Fix Cookbook download runner for Python 3.11

Avoid backslashes inside f-string expressions when generating Hugging Face install fallback commands. GitHub Actions compileall runs on Python 3.11, which rejects that syntax.

Verified with Python 3.11 feature-version AST parse, CI-scoped compileall, focused Cookbook import tests, and full container pytest: 4515 passed, 4 skipped.

.env.example

@@ -169,6 +169,26 @@ SEARXNG_INSTANCE=http://localhost:8080
 # ODYSSEUS_STT_MAX_AUDIO_BYTES=26214400                  # speech-to-text audio (25 MB)
 # ODYSSEUS_ICS_MAX_BYTES=10485760                        # calendar .ics import (10 MB)
 
+# ============================================================
+# Host Docker access (explicit opt-in)
+# ============================================================
+# Default Docker Compose does not mount /var/run/docker.sock. Existing
+# Ollama, vLLM, and other OpenAI-compatible endpoints remain usable without it.
+#
+# Enable this only for intentional Cookbook/local Docker-daemon management.
+# Raw socket access is high-trust and can grant broad control over the host
+# Docker daemon. Set DOCKER_GID to the host docker group's numeric GID.
+# Put these values in .env, or export them before running docker compose.
+# COMPOSE_FILE=docker-compose.yml:docker/host-docker.yml
+# DOCKER_GID=963
+# docker/host-docker.yml sets this inside the container. Keep it paired
+# with the socket overlay; setting it alone is not sufficient.
+# ODYSSEUS_ENABLE_HOST_DOCKER=true
+#
+# Host Docker access can be combined with one GPU overlay:
+# COMPOSE_FILE=docker-compose.yml:docker/gpu.nvidia.yml:docker/host-docker.yml
+# COMPOSE_FILE=docker-compose.yml:docker/gpu.amd.yml:docker/host-docker.yml
+
 # ============================================================
 # GPU support (Docker Compose)
 # ============================================================

.github/scripts/focused_test_guidance.py

@@ -0,0 +1,124 @@
+#!/usr/bin/env python3
+"""Report focused pytest guidance for changed paths under tests/."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import shlex
+import subprocess
+import sys
+from collections.abc import Iterable
+from pathlib import PurePosixPath
+
+
+def parse_paths(raw_paths: bytes) -> list[str]:
+    """Decode the NUL-delimited output of ``git diff --name-only -z``."""
+    return [os.fsdecode(path) for path in raw_paths.split(b"\0") if path]
+
+
+def changed_paths_from_merge_base(base_sha: str, head_sha: str) -> list[str]:
+    """Return changed ``tests/`` paths using GitHub PR three-dot semantics.
+
+    GitHub PR changed files are based on the merge base and the PR head, not a
+    direct endpoint diff between the current base branch tip and the PR head.
+    Using the direct endpoint diff can include files changed only on the base
+    branch when the PR branch is stale.
+    """
+    merge_base = subprocess.check_output(
+        ["git", "merge-base", base_sha, head_sha],
+        stderr=subprocess.DEVNULL,
+    ).strip()
+    raw_paths = subprocess.check_output(
+        [
+            "git",
+            "diff",
+            "--name-only",
+            "--diff-filter=ACMRT",
+            "-z",
+            os.fsdecode(merge_base),
+            head_sha,
+            "--",
+            "tests/",
+        ],
+    )
+    return parse_paths(raw_paths)
+
+
+def select_test_paths(paths: Iterable[str]) -> list[str]:
+    """Return unique, repository-relative paths contained by tests/."""
+    selected: set[str] = set()
+    for raw_path in paths:
+        path = PurePosixPath(raw_path)
+        if path.is_absolute() or ".." in path.parts:
+            continue
+        parts = tuple(part for part in path.parts if part != ".")
+        if len(parts) >= 2 and parts[0] == "tests":
+            selected.add(PurePosixPath(*parts).as_posix())
+    return sorted(selected)
+
+
+def is_pytest_file(path: str) -> bool:
+    """Return whether a changed path follows this repository's pytest naming."""
+    name = PurePosixPath(path).name
+    return name.endswith(".py") and (
+        name.startswith("test_") or name.endswith("_test.py")
+    )
+
+
+def pytest_command(paths: Iterable[str]) -> str:
+    """Build a copyable pytest command for changed runnable test files."""
+    command = ["python3", "-m", "pytest", "-q", *paths]
+    return shlex.join(command)
+
+
+def format_report(paths: Iterable[str]) -> str:
+    """Format focused guidance for CI logs and the workflow summary."""
+    changed_paths = select_test_paths(paths)
+    runnable_paths = [path for path in changed_paths if is_pytest_file(path)]
+    lines = ["## Focused test guidance (report-only)", ""]
+    if not changed_paths:
+        lines.append("No changed paths under `tests/`.")
+    else:
+        lines.extend(["Changed paths under `tests/`:", ""])
+        lines.extend(f"- `{path}`" for path in changed_paths)
+    lines.extend(["", "Suggested focused validation:", ""])
+    if runnable_paths:
+        lines.append(f"```sh\n{pytest_command(runnable_paths)}\n```")
+    else:
+        lines.append("No directly runnable pytest files changed.")
+    lines.extend(
+        [
+            "",
+            "This guidance does not infer tests from source changes. "
+            "Existing blocking CI remains the source of truth.",
+        ]
+    )
+    return "\n".join(lines)
+
+
+def _parse_args(argv: list[str]) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Report focused pytest guidance for changed tests/ paths.",
+    )
+    parser.add_argument("--base-sha", help="Pull request base commit SHA.")
+    parser.add_argument("--head-sha", help="Pull request head commit SHA.")
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = _parse_args(sys.argv[1:] if argv is None else argv)
+    if bool(args.base_sha) != bool(args.head_sha):
+        raise SystemExit("--base-sha and --head-sha must be provided together")
+
+    if args.base_sha and args.head_sha:
+        paths = changed_paths_from_merge_base(args.base_sha, args.head_sha)
+    else:
+        paths = parse_paths(sys.stdin.buffer.read())
+
+    print(format_report(paths))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

.github/workflows/ci.yml

@@ -15,6 +15,60 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  focused-test-guidance:
+    name: Focused test guidance (report-only)
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    continue-on-error: true
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0  # v7.0.0
+        with:
+          fetch-depth: 0
+          persist-credentials: false
+      - name: Report changed test paths
+        env:
+          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          report_file="$RUNNER_TEMP/focused-test-guidance.md"
+          publish_report() {
+            cat "$report_file"
+            if [ -n "${GITHUB_STEP_SUMMARY:-}" ]; then
+              cat "$report_file" >> "$GITHUB_STEP_SUMMARY" || true
+            fi
+            return 0
+          }
+
+          report_unavailable() {
+            {
+              printf '%s\n\n' '## Focused test guidance unavailable (report-only)'
+              printf '%s\n\n' "$1"
+              printf '%s\n' 'Existing blocking CI remains the source of truth.'
+            } > "$report_file"
+            publish_report
+            exit 0
+          }
+
+          if [ -z "$BASE_SHA" ] || [ -z "$HEAD_SHA" ]; then
+            report_unavailable "Pull request base/head metadata is missing."
+          fi
+
+          if ! git cat-file -e "${BASE_SHA}^{commit}" 2>/dev/null; then
+            report_unavailable "The pull request base commit is unavailable locally."
+          fi
+
+          if ! git cat-file -e "${HEAD_SHA}^{commit}" 2>/dev/null; then
+            report_unavailable "The pull request head commit is unavailable locally."
+          fi
+
+          if ! python3 .github/scripts/focused_test_guidance.py \
+            --base-sha "$BASE_SHA" \
+            --head-sha "$HEAD_SHA" > "$report_file"; then
+            report_unavailable "The focused test guidance helper could not produce a report."
+          fi
+
+          publish_report
+
   python-syntax:
     name: Python syntax (compileall)
     runs-on: ubuntu-latest

app.py

@@ -3,6 +3,7 @@ import mimetypes
 import os
 import sys
 import asyncio
+import time
 
 # On Windows, asyncio.create_subprocess_exec/shell require the ProactorEventLoop.
 # When started via `python -m uvicorn` from a terminal, uvicorn sets this
@@ -197,7 +198,50 @@ class _RequestTimeoutMiddleware(_BaseHTTPMiddleware):
             )
 
 
+class _InteractiveActivityMiddleware(_BaseHTTPMiddleware):
+    async def dispatch(self, request, call_next):
+        from src.interactive_gate import should_track_interactive_request, track_interactive_request
+
+        path = request.url.path or ""
+        if not should_track_interactive_request(path, request.method):
+            return await call_next(request)
+        async def _stop_background():
+            try:
+                await task_scheduler.stop_background_tasks_for_foreground(reason=f"foreground request {request.method} {path}")
+            except Exception:
+                logging.getLogger("app.foreground_gate").debug("foreground task stop failed", exc_info=True)
+        asyncio.create_task(_stop_background())
+        async with track_interactive_request(path, request.method):
+            return await call_next(request)
+
+
+class _SlowRequestLogMiddleware(_BaseHTTPMiddleware):
+    async def dispatch(self, request, call_next):
+        start = time.perf_counter()
+        status = 500
+        try:
+            response = await call_next(request)
+            status = getattr(response, "status_code", 0) or 0
+            return response
+        finally:
+            elapsed = time.perf_counter() - start
+            try:
+                threshold = float(os.getenv("ODYSSEUS_SLOW_REQUEST_LOG_SECONDS", "0.75") or "0.75")
+            except Exception:
+                threshold = 0.75
+            if elapsed >= threshold:
+                logging.getLogger("app.slow_request").warning(
+                    "slow_request method=%s path=%s status=%s elapsed=%.3fs",
+                    request.method,
+                    request.url.path,
+                    status,
+                    elapsed,
+                )
+
+
 app.add_middleware(_RequestTimeoutMiddleware)
+app.add_middleware(_InteractiveActivityMiddleware)
+app.add_middleware(_SlowRequestLogMiddleware)
 
 # ========= AUTH =========
 from routes.auth_routes import setup_auth_routes, SESSION_COOKIE
@@ -583,6 +627,20 @@ webhook_manager = WebhookManager(api_key_manager=api_key_manager)
 auth_router = setup_auth_routes(auth_manager)
 app.include_router(auth_router)
 
+
+@app.post("/api/activity/heartbeat")
+async def activity_heartbeat():
+    from src.interactive_gate import mark_browser_activity
+    await mark_browser_activity()
+    async def _stop_background():
+        try:
+            await task_scheduler.stop_background_tasks_for_foreground(reason="browser heartbeat")
+        except Exception:
+            logging.getLogger("app.foreground_gate").debug("heartbeat task stop failed", exc_info=True)
+    asyncio.create_task(_stop_background())
+    return {"ok": True}
+
+
 # Uploads
 from routes.upload_routes import setup_upload_routes
 upload_router, upload_cleanup_func = setup_upload_routes(upload_handler)
@@ -604,7 +662,7 @@ from routes.admin_wipe_routes import setup_admin_wipe_routes
 app.include_router(setup_admin_wipe_routes(session_manager))
 
 # Memory
-from routes.memory_routes import setup_memory_routes
+from routes.memory.memory_routes import setup_memory_routes
 memory_router = setup_memory_routes(memory_manager, session_manager, memory_vector=memory_vector)
 app.include_router(memory_router)
 from routes.skills_routes import setup_skills_routes
@@ -621,11 +679,11 @@ app.include_router(setup_chat_routes(
 ))
 
 # Research (background deep-research tasks)
-from routes.research_routes import setup_research_routes
+from routes.research.research_routes import setup_research_routes
 app.include_router(setup_research_routes(research_handler, session_manager=session_manager))
 
 # History
-from routes.history_routes import setup_history_routes
+from routes.history.history_routes import setup_history_routes
 app.include_router(setup_history_routes(session_manager))
 
 # Search
@@ -793,7 +851,7 @@ from routes.vault_routes import setup_vault_routes
 app.include_router(setup_vault_routes())
 
 # Contacts (CardDAV)
-from routes.contacts_routes import setup_contacts_routes
+from routes.contacts.contacts_routes import setup_contacts_routes
 app.include_router(setup_contacts_routes())
 
 from companion import setup_companion_routes
@@ -869,6 +927,34 @@ async def get_version():
 async def health_check() -> Dict[str, str]:
     return {"status": "healthy", "timestamp": datetime.now(timezone.utc).isoformat()}
 
+@app.post("/api/client-perf")
+async def client_perf(request: Request):
+    """Low-volume frontend timing reports for stalls that happen before SSE logs."""
+    try:
+        data = await request.json()
+    except Exception:
+        data = {}
+    try:
+        kind = str(data.get("type") or "client").replace("\n", " ")[:80]
+        total_ms = float(data.get("total_ms") or 0)
+        stages = data.get("stages") if isinstance(data.get("stages"), list) else []
+        stage_txt = " ".join(
+            f"{str(s.get('name') or '')[:40]}={float(s.get('delta_ms') or 0):.0f}ms"
+            for s in stages[:20]
+            if isinstance(s, dict)
+        )
+        extra = str(data.get("extra") or "").replace("\n", " ")[:200]
+        logging.getLogger("app.client_perf").warning(
+            "client_perf type=%s total=%.0fms %s%s",
+            kind,
+            total_ms,
+            stage_txt,
+            f" extra={extra}" if extra else "",
+        )
+    except Exception:
+        logging.getLogger("app.client_perf").debug("client_perf log failed", exc_info=True)
+    return {"ok": True}
+
 @app.get("/api/ready")
 async def readiness_check() -> JSONResponse:
     """Readiness / integrity self-check — DB, data dir, local-first storage.
@@ -965,57 +1051,59 @@ async def _startup_event():
 
     _startup_tasks.append(asyncio.create_task(_startup_mcp_connections()))
 
-    # Pre-warm the RAG tool index off the request path. Loading the local
-    # embedding model + opening ChromaDB + indexing the built-in tools is a
-    # one-time ~1-3s cost that otherwise lands on the user's FIRST message
-    # (showing up as a big `tool_selection` time). Doing it here makes the
-    # first turn as fast as subsequent ones (warm embed ≈ a few ms).
-    async def _warmup_tool_index():
-        try:
-            from src.tool_index import get_tool_index
-            idx = await asyncio.to_thread(get_tool_index)
-            if idx:
-                await asyncio.to_thread(idx.get_tools_for_query, "warmup", 8)
-                logger.info("[startup] Tool index pre-warmed")
-        except Exception as e:
-            logger.warning(f"Tool index warmup failed (non-critical): {type(e).__name__}: {e}")
-
-    _startup_tasks.append(asyncio.create_task(_warmup_tool_index()))
-    # Warmup: ping all known LLM endpoints to prime connections
-    async def _warmup_endpoints():
-        try:
-            import httpx
-            # model_discovery has no get_endpoints(); that call raised
-            # AttributeError every run and silently disabled warmup/keepalive.
-            # Resolve the /models probe URLs via the real discovery API, off the
-            # event loop since discovery does a blocking port scan.
-            urls = (
-                await asyncio.to_thread(model_discovery.warmup_ping_urls)
-                if model_discovery else []
-            )
-            for url in urls:
-                try:
-                    async with httpx.AsyncClient(timeout=5.0) as client:
-                        await client.get(url)
-                    logger.info(f"Warmup ping OK: {url}")
-                except Exception as e:
-                    logger.debug(f"Warmup ping failed for endpoint: {e}")
-        except Exception as e:
-            logger.debug(f"Warmup ping skipped: {e}")
-
-    _startup_tasks.append(asyncio.create_task(_warmup_endpoints()))
-
-    # Keep-alive: ping endpoints every 60 seconds to prevent cold starts
-    async def _keepalive_loop():
-        while True:
+    # Startup warmups are opt-in. They make later requests a little warmer, but
+    # they also compete with the first seconds of real UI use on slow or busy
+    # machines. Default to clear/idle startup and let requests warm what they use.
+    _startup_warmups_enabled = str(os.getenv("ODYSSEUS_STARTUP_WARMUPS", "")).lower() in {"1", "true", "yes", "on"}
+    if _startup_warmups_enabled:
+        async def _warmup_tool_index():
             try:
-                await asyncio.sleep(60)
-                await _warmup_endpoints()
+                from src.tool_index import get_tool_index
+                idx = await asyncio.to_thread(get_tool_index)
+                if idx:
+                    await asyncio.to_thread(idx.get_tools_for_query, "warmup", 8)
+                    logger.info("[startup] Tool index pre-warmed")
             except Exception as e:
-                logger.warning(f"Keepalive loop error: {e}")
-                await asyncio.sleep(300)  # Back off on error
+                logger.warning(f"Tool index warmup failed (non-critical): {type(e).__name__}: {e}")
 
-    _startup_tasks.append(asyncio.create_task(_keepalive_loop()))
+        _startup_tasks.append(asyncio.create_task(_warmup_tool_index()))
+
+        async def _warmup_endpoints():
+            try:
+                import httpx
+                urls = (
+                    await asyncio.to_thread(model_discovery.warmup_ping_urls)
+                    if model_discovery else []
+                )
+                for url in urls:
+                    try:
+                        async with httpx.AsyncClient(timeout=5.0) as client:
+                            await client.get(url)
+                        logger.info(f"Warmup ping OK: {url}")
+                    except Exception as e:
+                        logger.debug(f"Warmup ping failed for endpoint: {e}")
+            except Exception as e:
+                logger.debug(f"Warmup ping skipped: {e}")
+
+        _startup_tasks.append(asyncio.create_task(_warmup_endpoints()))
+    else:
+        logger.info("Startup warmups disabled (set ODYSSEUS_STARTUP_WARMUPS=1 to enable)")
+
+    # Keep-alive is opt-in. The ping path performs model discovery, and when
+    # stale LAN endpoints are configured it can add periodic backend pressure
+    # that delays unrelated UI requests such as Notes/Documents.
+    _keepalive_enabled = str(os.getenv("ODYSSEUS_MODEL_KEEPALIVE", "")).lower() in {"1", "true", "yes", "on"}
+    if _keepalive_enabled:
+        async def _keepalive_loop():
+            while True:
+                try:
+                    await asyncio.sleep(60)
+                    await _warmup_endpoints()
+                except Exception as e:
+                    logger.warning(f"Keepalive loop error: {e}")
+                    await asyncio.sleep(300)  # Back off on error
+
+        _startup_tasks.append(asyncio.create_task(_keepalive_loop()))
 
     async def _ensure_default_tasks():
         # Create/reconcile default automation tasks + personal assistant for every user.

core/atomic_io.py

@@ -34,6 +34,8 @@ def atomic_write_json(path: str, data: Any, *, indent: Optional[int] = None) ->
 
 
 def atomic_write_text(path: str, text: str) -> None:
+    if not isinstance(text, str):
+        raise TypeError("atomic_write_text expects a string")
     os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
     tmp = f"{path}.tmp.{os.getpid()}"
     with open(tmp, "w", encoding="utf-8") as f:

core/database.py

@@ -276,6 +276,7 @@ class GalleryImage(TimestampMixin, Base):
     id         = Column(String, primary_key=True, index=True)
     filename   = Column(String, nullable=False, unique=True)
     prompt     = Column(Text, nullable=False, default="")
+    caption    = Column(Text, nullable=True, default="")
     model      = Column(String, nullable=True)
     size       = Column(String, nullable=True)
     quality    = Column(String, nullable=True)
@@ -1182,6 +1183,29 @@ def _migrate_add_multiuser_owner_columns():
     _migrate_add_owner_to_table("documents", "ix_documents_owner")
 
 
+def _migrate_add_gallery_caption_column():
+    """Add OCR/vision caption storage for gallery images."""
+    import sqlite3
+    db_path = DATABASE_URL.replace("sqlite:///", "")
+    if not os.path.exists(db_path):
+        return
+    conn = None
+    try:
+        conn = sqlite3.connect(db_path)
+        columns = [row[1] for row in conn.execute("PRAGMA table_info(gallery_images)").fetchall()]
+        if columns and "caption" not in columns:
+            conn.execute("ALTER TABLE gallery_images ADD COLUMN caption TEXT DEFAULT ''")
+            conn.commit()
+            logging.getLogger(__name__).info("Migrated: added caption column to gallery_images")
+    except Exception as e:
+        logging.getLogger(__name__).warning(f"Migration gallery caption column failed: {e}")
+    finally:
+        try:
+            conn.close()
+        except Exception:
+            pass
+
+
 def _migrate_add_api_token_scopes_column():
     """Add API token scopes for existing installs.
 
@@ -1670,6 +1694,7 @@ class CalendarEvent(TimestampMixin, Base):
     # `Z`-suffix on serialization so the frontend interprets correctly.
     is_utc      = Column(Boolean, default=False, nullable=False)
     rrule       = Column(String, default="")
+    recurrence_exdates = Column(Text, default="")  # JSON list of skipped occurrence starts
     color       = Column(String, nullable=True)  # per-event color override
     status      = Column(String, default="confirmed")  # confirmed, cancelled
     importance  = Column(String, default="normal")    # low | normal | high | critical
@@ -1811,6 +1836,7 @@ def init_db():
     _migrate_add_token_columns()
     _migrate_add_mode_column()
     _migrate_add_multiuser_owner_columns()
+    _migrate_add_gallery_caption_column()
     _migrate_add_api_token_scopes_column()
     _migrate_backfill_document_owner_from_session()
     _migrate_assign_legacy_owner()
@@ -1833,6 +1859,7 @@ def init_db():
     _migrate_add_calendar_origin()
     _migrate_add_calendar_account_id()
     _migrate_add_caldav_sync_columns()
+    _migrate_add_calendar_recurrence_exdates()
     _migrate_chat_messages_fts()
     _migrate_encrypt_email_passwords()
     _migrate_encrypt_signatures()
@@ -2184,6 +2211,28 @@ def _migrate_add_calendar_metadata():
         except Exception:
             pass
 
+
+def _migrate_add_calendar_recurrence_exdates():
+    """Add skipped recurrence occurrences for deleting one instance of a series."""
+    import sqlite3
+    db_path = DATABASE_URL.replace("sqlite:///", "")
+    if not os.path.exists(db_path):
+        return
+    conn = None
+    try:
+        conn = sqlite3.connect(db_path)
+        columns = [row[1] for row in conn.execute("PRAGMA table_info(calendar_events)").fetchall()]
+        if columns and "recurrence_exdates" not in columns:
+            conn.execute("ALTER TABLE calendar_events ADD COLUMN recurrence_exdates TEXT DEFAULT ''")
+        conn.commit()
+    except Exception as e:
+        logging.getLogger(__name__).warning(f"calendar_events recurrence_exdates migration failed: {e}")
+    finally:
+        try:
+            conn.close()
+        except Exception:
+            pass
+
 def get_db():
     """
     Dependency to get a database session.

core/middleware.py

@@ -117,7 +117,7 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
                 f"script-src 'self' 'nonce-{nonce}' https://cdn.jsdelivr.net; "
                 "style-src 'self' 'unsafe-inline' https://cdn.jsdelivr.net; "
                 "font-src 'self' https://cdn.jsdelivr.net; "
-                "img-src 'self' data: blob:; "
+                "img-src 'self' data: blob: https:; "
                 "media-src 'self' blob:; "
                 "connect-src 'self'; "
                 "frame-src 'self'; "

docker-compose.gpu-amd.yml

@@ -28,14 +28,6 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
-      # Docker socket — lets Cookbook launch commands like
-      # `docker exec ollama-rocm ollama show <tag>` reach the host's
-      # Docker daemon (and sibling containers like ollama-rocm /
-      # ollama-test). The in-container user needs to be in the
-      # socket's owning group — see `group_add` below; the GID
-      # there must match the host's `docker` group (defaults to 963
-      # on Debian, 999 on Ubuntu — override via env if yours differs).
-      - /var/run/docker.sock:/var/run/docker.sock
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.
@@ -101,7 +93,6 @@ services:
       - /dev/kfd
       - /dev/dri
     group_add:
-      - "${DOCKER_GID:-963}"
       - video
       - ${RENDER_GID:-render}
 

docker-compose.gpu-nvidia.yml

@@ -27,16 +27,6 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
-      # Docker socket — lets Cookbook launch commands like
-      # `docker exec ollama-rocm ollama show <tag>` reach the host's
-      # Docker daemon (and sibling containers like ollama-rocm /
-      # ollama-test). The in-container user needs to be in the
-      # socket's owning group — see `group_add` below; the GID
-      # there must match the host's `docker` group (defaults to 963
-      # on Debian, 999 on Ubuntu — override via env if yours differs).
-      - /var/run/docker.sock:/var/run/docker.sock
-    group_add:
-      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.

docker-compose.yml

@@ -16,16 +16,6 @@ services:
       # land under /app/.local for the odysseus user. Persist them so a
       # container recreate does not silently remove installed serve engines.
       - ${APP_DATA_DIR:-./data}/local:/app/.local:z
-      # Docker socket — lets Cookbook launch commands like
-      # `docker exec ollama-rocm ollama show <tag>` reach the host's
-      # Docker daemon (and sibling containers like ollama-rocm /
-      # ollama-test). The in-container user needs to be in the
-      # socket's owning group — see `group_add` below; the GID
-      # there must match the host's `docker` group (defaults to 963
-      # on Debian, 999 on Ubuntu — override via env if yours differs).
-      - /var/run/docker.sock:/var/run/docker.sock
-    group_add:
-      - "${DOCKER_GID:-963}"
     extra_hosts:
       # Lets the container reach local services on the Docker host, including
       # Ollama at http://host.docker.internal:11434.

docker/entrypoint.sh

@@ -29,12 +29,12 @@ fi
 ODY_USER="$(getent passwd "$PUID" | cut -d: -f1)"
 [ -z "$ODY_USER" ] && ODY_USER=odysseus
 
-# Docker-socket group plumbing. When /var/run/docker.sock is bind-mounted
-# (Cookbook uses docker exec to reach sibling containers), the socket is
-# owned by root:<host docker gid>. Add the app user to that group and later
-# call gosu by username so supplementary groups are retained.
+# Docker-socket group plumbing for the explicit host-Docker overlay. When
+# opted in, the socket is owned by root:<host docker gid>. Add the app user
+# to that group and later call gosu by username so supplementary groups are
+# retained.
 DOCKER_SOCK="${DOCKER_SOCK:-/var/run/docker.sock}"
-if [ -S "$DOCKER_SOCK" ]; then
+if [ "${ODYSSEUS_ENABLE_HOST_DOCKER:-}" = "true" ] && [ -S "$DOCKER_SOCK" ]; then
     SOCK_GID="$(stat -c '%g' "$DOCKER_SOCK" 2>/dev/null || echo '')"
     if [ -n "$SOCK_GID" ] && [ "$SOCK_GID" != "0" ]; then
         if ! getent group "$SOCK_GID" >/dev/null 2>&1; then

docker/host-docker.yml

@@ -0,0 +1,12 @@
+# High-trust host Docker access. Enable only when local Docker-daemon
+# management from Cookbook is required and you accept that raw socket access
+# grants broad control over the host Docker daemon.
+#   COMPOSE_FILE=docker-compose.yml:docker/host-docker.yml
+#   DOCKER_GID=<numeric host Docker group id>
+services:
+  odysseus:
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock
+    group_add: ["${DOCKER_GID:-963}"]
+    environment:
+      - ODYSSEUS_ENABLE_HOST_DOCKER=true

docs/setup.md

@@ -99,6 +99,33 @@ Odysseus SSH key and add the public key to the remote server's
 ssh-copy-id -i data/ssh/id_ed25519.pub user@server
 ```
 
+**Host Docker access (explicit opt-in).** Default Docker Compose intentionally
+does not mount `/var/run/docker.sock`. You can still connect Odysseus to
+existing Ollama, vLLM, and other OpenAI-compatible endpoints without Docker
+socket access.
+
+Cookbook/local Docker-daemon management requires the opt-in overlay below. Raw
+Docker socket access is high-trust because it can effectively grant broad
+control over the host Docker daemon. Remote server Docker workflows over SSH
+remain preferred.
+
+Place these values in `.env`, or export them in the shell before running
+`docker compose`:
+
+```bash
+COMPOSE_FILE=docker-compose.yml:docker/host-docker.yml
+DOCKER_GID=<host docker group gid>
+```
+
+Combine host Docker access with a GPU overlay when both are intentionally
+required:
+
+```bash
+COMPOSE_FILE=docker-compose.yml:docker/gpu.nvidia.yml:docker/host-docker.yml
+# or
+COMPOSE_FILE=docker-compose.yml:docker/gpu.amd.yml:docker/host-docker.yml
+```
+
 **Docker GPU overlays.** CPU-only users can skip this section. Cookbook can
 only detect GPUs that Docker exposes to the container — if the host runtime or
 device passthrough is not configured, Cookbook sees the iGPU, another card, or
@@ -445,4 +472,4 @@ All user data lives in `data/` (gitignored): `app.db` (sessions, messages, docum
 `memory.json`, `presets.json`, `uploads/`, `personal_docs/`, `chroma/`, `settings.json`.
 
 To back up or restore everything in `data/`, see the
-[Backup & Restore guide](docs/backup-restore.md).
+[Backup & Restore guide](backup-restore.md).

mcp_servers/email_server.py

@@ -58,6 +58,11 @@ def _uid_fetch_rows(data) -> list:
 _ACCOUNT_CACHE: dict = {}  # key = normalized account selector -> config dict
 _MCP_OWNER_ARG = "_odysseus_owner"
 _CURRENT_OWNER: ContextVar[str | None] = ContextVar("email_mcp_owner", default=None)
+_OWNER_ENV_KEYS = ("ODYSSEUS_MCP_EMAIL_OWNER", "ODYSSEUS_EMAIL_OWNER")
+_OWNER_SCOPE_ERROR = (
+    "Error: email MCP requires an authenticated owner or ODYSSEUS_MCP_EMAIL_OWNER "
+    "when owner-scoped email accounts are configured."
+)
 
 
 def _clean_header_value(value) -> str:
@@ -71,13 +76,29 @@ def _db_path() -> Path:
     return Path(APP_DB)
 
 
+def _configured_owner() -> str | None:
+    for key in _OWNER_ENV_KEYS:
+        owner = os.environ.get(key, "").strip()
+        if owner:
+            return owner
+    return None
+
+
 def _current_owner() -> str:
     owner = _CURRENT_OWNER.get()
-    return str(owner or "").strip()
+    return str(owner or _configured_owner() or "").strip()
+
+
+def _account_owner(row: dict) -> str:
+    return str(row.get("owner") or "").strip()
+
+
+def _has_owner_scoped_accounts(rows: list[dict]) -> bool:
+    return any(_account_owner(r) for r in rows)
 
 
 def _account_visible_to_owner(row: dict, owner: str) -> bool:
-    row_owner = str(row.get("owner") or "").strip()
+    row_owner = _account_owner(row)
     if row_owner == owner:
         return True
     if row_owner:
@@ -96,8 +117,7 @@ def _filter_accounts_for_owner(rows: list[dict]) -> list[dict]:
     if owner:
         return [r for r in rows if _account_visible_to_owner(r, owner)]
 
-    owners = {str(r.get("owner") or "").strip() for r in rows if str(r.get("owner") or "").strip()}
-    if len(owners) > 1:
+    if _has_owner_scoped_accounts(rows):
         return []
     return rows
 
@@ -106,8 +126,7 @@ def _mcp_owner_required(rows: list[dict] | None = None) -> bool:
     if _current_owner():
         return False
     rows = rows if rows is not None else _read_accounts_from_db()
-    owners = {str(r.get("owner") or "").strip() for r in rows if str(r.get("owner") or "").strip()}
-    return len(owners) > 1
+    return _has_owner_scoped_accounts(rows)
 
 
 def _load_email_writing_style() -> str:
@@ -274,6 +293,8 @@ def _load_config(account: str | None = None) -> dict:
     }
 
     raw_rows = _read_accounts_from_db()
+    if _mcp_owner_required(raw_rows):
+        raise ValueError(_OWNER_SCOPE_ERROR)
     rows = _filter_accounts_for_owner(raw_rows)
     row = _resolve_account_from_rows(rows, account)
     if _current_owner() and raw_rows and not rows:
@@ -538,6 +559,148 @@ def _get_cached_summaries():
         return {}
 
 
+def _fixture_email_file() -> Path:
+    return DATA_DIR / "fixture_email_messages.json"
+
+
+def _fixture_email_enabled() -> bool:
+    return _fixture_email_file().exists()
+
+
+def _parse_fixture_date(raw_date: str) -> tuple[str, float]:
+    if not raw_date:
+        return "", 0.0
+    parsed = None
+    try:
+        parsed = datetime.fromisoformat(str(raw_date).replace("Z", "+00:00"))
+    except Exception:
+        try:
+            parsed = email.utils.parsedate_to_datetime(str(raw_date))
+        except Exception:
+            parsed = None
+    if parsed:
+        return parsed.isoformat(), parsed.timestamp()
+    return str(raw_date), 0.0
+
+
+def _fixture_email_record(row: dict, uid_num: int, owner: str) -> dict:
+    sender = str(row.get("from") or "Fixture Sender <fixture@example.invalid>")
+    sender_name, sender_addr = email.utils.parseaddr(sender)
+    date_str, date_epoch = _parse_fixture_date(str(row.get("date") or ""))
+    subject = str(row.get("subject") or "(no subject)")
+    body = str(row.get("body") or "")
+    owner_key = re.sub(r"[^A-Za-z0-9_.-]", "-", owner or "default")
+    uid = str(uid_num)
+    return {
+        "uid": uid,
+        "message_id": f"<fixture-email-{uid}-{owner_key}@fixtures.odysseus.local>",
+        "subject": subject,
+        "from": sender_name or sender_addr or sender,
+        "from_address": sender_addr,
+        "date": date_str,
+        "date_epoch": date_epoch,
+        "summary": body[:240],
+        "body": body,
+        "account": "Fixture Inbox",
+        "account_email": owner or str(row.get("owner") or ""),
+        "account_id": "fixture-email",
+        "attachments": [],
+    }
+
+
+def _fixture_email_rows(owner: str | None = None) -> list[dict]:
+    path = _fixture_email_file()
+    if not path.exists():
+        return []
+    try:
+        raw = json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return []
+    rows = raw.get("messages") if isinstance(raw, dict) else raw
+    out = []
+    owner = str(owner or "").strip()
+    for i, row in enumerate(rows if isinstance(rows, list) else [], start=1):
+        if not isinstance(row, dict):
+            continue
+        row_owner = str(row.get("owner") or "").strip()
+        if owner and row_owner and row_owner != owner:
+            continue
+        out.append(_fixture_email_record(row, i, owner or row_owner))
+    out.sort(key=lambda item: item.get("date_epoch") or 0, reverse=True)
+    return out
+
+
+def _fixture_account_rows() -> list[dict]:
+    if not _fixture_email_enabled():
+        return []
+    owner = _current_owner()
+    owners = []
+    for row in _fixture_email_rows(owner or None):
+        email_addr = row.get("account_email") or owner or "fixture@fixtures.odysseus.local"
+        if email_addr not in owners:
+            owners.append(email_addr)
+    if not owners:
+        owners = [owner or "fixture@fixtures.odysseus.local"]
+    return [
+        {
+            "id": "fixture-email",
+            "owner": owner or owners[0],
+            "name": "Fixture Inbox",
+            "is_default": True,
+            "imap_user": owners[0],
+            "from_address": owners[0],
+        }
+    ]
+
+
+def _fixture_email_matches(item: dict, query: str) -> bool:
+    if not query:
+        return True
+    terms = [term for term in re.split(r"\W+", str(query).lower()) if term]
+    haystack = "\n".join(
+        str(item.get(key) or "")
+        for key in ("subject", "from", "from_address", "body", "summary")
+    ).lower()
+    return all(term in haystack for term in terms)
+
+
+def _fixture_list_emails(folder="INBOX", max_results=20, unresponded_only=False,
+                         unread_only=False, account=None) -> list[dict] | None:
+    if not _fixture_email_enabled():
+        return None
+    if account and str(account).strip().lower() not in {
+        "fixture-email",
+        "fixture inbox",
+        "fixture",
+        str(_current_owner()).lower(),
+    }:
+        return []
+    if (folder or "INBOX").upper() not in {"INBOX", "ALL", "ALL MAIL"}:
+        return []
+    return _fixture_email_rows(_current_owner())[: int(max_results or 20)]
+
+
+def _fixture_search_emails(query, folders=None, max_results=20, account=None) -> list[dict] | None:
+    if not _fixture_email_enabled():
+        return None
+    rows = _fixture_list_emails("INBOX", max_results=1000, account=account) or []
+    out = [dict(row, _folder="INBOX") for row in rows if _fixture_email_matches(row, str(query or ""))]
+    return out[: int(max_results or 20)]
+
+
+def _fixture_read_email(uid=None, message_id=None, folder="INBOX", account=None) -> dict | None:
+    if not _fixture_email_enabled():
+        return None
+    if (folder or "INBOX").upper() not in {"INBOX", "ALL", "ALL MAIL"}:
+        return {"error": f"Email UID {uid or message_id} not found"}
+    for item in _fixture_email_rows(_current_owner()):
+        if uid and str(item.get("uid")) == str(uid):
+            return item
+        if message_id and str(item.get("message_id")) == str(message_id):
+            return item
+    return {"error": f"Email not found with UID/Message-ID: {uid or message_id}"}
+
+
 # ── Tool implementations ──
 
 
@@ -548,6 +711,9 @@ def _list_emails(folder="INBOX", max_results=20, unresponded_only=False,
     Pass unread_only=True and/or unresponded_only=True for attention scans.
     account selects mailbox (None = default).
     """
+    fixture = _fixture_list_emails(folder, max_results, unresponded_only, unread_only, account)
+    if fixture is not None:
+        return fixture
     conn = None
     try:
         conn = _imap_connect(account)
@@ -629,6 +795,9 @@ def _result_sort_time(result: dict) -> datetime:
 
 def _list_emails_across_accounts(folder="INBOX", max_results=20,
                                  unresponded_only=False, unread_only=False):
+    fixture = _fixture_list_emails(folder, max_results, unresponded_only, unread_only, None)
+    if fixture is not None:
+        return fixture, []
     rows = _list_accounts_raw()
     combined = []
     errors = []
@@ -662,6 +831,9 @@ def _search_emails(query, folders=None, max_results=20, account=None):
     _list_emails plus an `_folder` tag."""
     if not query or not str(query).strip():
         return []
+    fixture = _fixture_search_emails(query, folders=folders, max_results=max_results, account=account)
+    if fixture is not None:
+        return fixture
     q = str(query).replace("\\", "\\\\").replace('"', '\\"')
     # Mail clients commonly use OR FROM/SUBJECT/TEXT to match either field.
     # IMAP SEARCH OR is binary, so we nest it.
@@ -784,6 +956,9 @@ def _extract_attachment_to_disk(msg, index, target_dir):
 
 def _read_email(uid=None, message_id=None, folder="INBOX", account=None):
     """Read full email content by UID or message-ID. account = mailbox selector."""
+    fixture = _fixture_read_email(uid=uid, message_id=message_id, folder=folder, account=account)
+    if fixture is not None:
+        return fixture
     cfg = _load_config(account)
     conn = None
     try:
@@ -837,6 +1012,9 @@ def _read_email(uid=None, message_id=None, folder="INBOX", account=None):
 
 
 def _read_email_across_accounts(uid=None, message_id=None, folder="INBOX"):
+    fixture = _fixture_read_email(uid=uid, message_id=message_id, folder=folder, account=None)
+    if fixture is not None:
+        return fixture
     rows = _list_accounts_raw()
     matches = []
     errors = []
@@ -1036,10 +1214,14 @@ def _send_email(to, subject, body, in_reply_to=None, references=None, cc=None, b
     UI. This closes the auto-send hole that let earlier models invent
     signatures and ship them to real recipients without confirmation."""
     if _read_agent_email_confirm_setting():
+        # Even confirmation-first sends must resolve the selected account now.
+        # Otherwise a caller could stage a pending draft against another
+        # owner's account selector before browser approval handles it.
+        cfg = _load_config(account)
         return _stash_agent_draft(
             to=to, subject=subject, body=body,
             in_reply_to=in_reply_to, references=references,
-            cc=cc, bcc=bcc, account=account,
+            cc=cc, bcc=bcc, account=cfg.get("account_id") or account,
         )
     send_account, cfg = _resolve_send_config(account)
     msg = EmailMessage()
@@ -1775,9 +1957,10 @@ async def list_tools() -> list[Tool]:
         Tool(
             name="reply_to_email",
             description=(
-                "Reply to an existing email by UID. This sends immediately; for normal "
-                "assistant-written replies, prefer draft_email_reply so the user can "
-                "review and send from Odysseus. Automatically threads the reply with "
+                "Reply to an existing email by UID. This sends immediately. Do NOT use "
+                "for normal 'write/draft a reply saying X' requests; use "
+                "draft_email_reply so the user can review and send from Odysseus. "
+                "Only use this when the user explicitly says to send now. Automatically threads the reply with "
                 "In-Reply-To and References headers, prefixes 'Re:' on the subject, and "
                 "uses the original sender as the recipient. Set reply_all=true to also CC "
                 "the original To/Cc recipients. For follow-up 'reply ...' requests, use "
@@ -1984,13 +2167,12 @@ async def call_tool(name: str, arguments: dict) -> list[TextContent]:
     try:
         all_db_accounts = _read_accounts_from_db()
         if _mcp_owner_required(all_db_accounts):
-            return [TextContent(
-                type="text",
-                text="Error: email MCP requires an authenticated owner when multiple email account owners are configured.",
-            )]
+            return [TextContent(type="text", text=_OWNER_SCOPE_ERROR)]
 
         if name == "list_email_accounts":
             rows = _filter_accounts_for_owner(all_db_accounts)
+            if not rows:
+                rows = _fixture_account_rows()
             if not rows:
                 if all_db_accounts and owner:
                     return [TextContent(type="text", text="No email accounts configured for this owner.")]

requirements.txt

@@ -3,6 +3,7 @@ uvicorn
 python-multipart
 python-dotenv
 httpx
+httpcore>=1.0,<2.0
 pydantic>=2.13.4
 pydantic-settings>=2.14.1
 SQLAlchemy

routes/calendar_routes.py

@@ -1,6 +1,7 @@
 """Calendar routes — local SQLite-backed calendar CRUD."""
 
 import logging
+import json
 import re
 import uuid
 from datetime import datetime, date, timedelta
@@ -541,6 +542,7 @@ def _event_to_dict(ev: CalendarEvent) -> dict:
         "description": ev.description or "",
         "location": ev.location or "",
         "rrule": ev.rrule or "",
+        "recurrence_exdates": _recurrence_exdates(ev),
         "calendar": ev.calendar.name if ev.calendar else "",
         "calendar_href": ev.calendar_id,
         "color": ev.color or (ev.calendar.color if ev.calendar else ""),
@@ -554,6 +556,28 @@ def _event_to_dict(ev: CalendarEvent) -> dict:
 _RRULE_EXPANSION_LIMIT = 1000
 
 
+def _recurrence_exdates(ev: CalendarEvent) -> list[str]:
+    raw = getattr(ev, "recurrence_exdates", "") or ""
+    if not raw:
+        return []
+    try:
+        values = json.loads(raw)
+    except Exception:
+        return []
+    if not isinstance(values, list):
+        return []
+    return [str(v) for v in values if isinstance(v, str) and v.strip()]
+
+
+def _occurrence_exdate_key(uid: str, ev: CalendarEvent) -> str:
+    if "::" not in uid:
+        return ""
+    suffix = uid.split("::", 1)[1]
+    if ev.all_day:
+        return suffix[:10]
+    return suffix[:16]
+
+
 def _expand_rrule(
     ev: CalendarEvent, start: datetime, end: datetime
 ) -> List[dict]:
@@ -618,6 +642,7 @@ def _expand_rrule(
     results = []
     truncated = False
     base = _event_to_dict(ev)
+    exdates = set(_recurrence_exdates(ev))
 
     for occ_start in rule.xafter(expand_start, inc=True):
         if occ_start >= end:
@@ -638,8 +663,13 @@ def _expand_rrule(
         # Build the compound uid: {base_uid}::{date} or ::{datetime}
         if ev.all_day:
             occ_uid = f"{ev.uid}::{occ_start.strftime('%Y-%m-%d')}"
+            exdate_key = occ_start.strftime("%Y-%m-%d")
         else:
             occ_uid = f"{ev.uid}::{occ_start.strftime('%Y-%m-%dT%H:%M')}"
+            exdate_key = occ_start.strftime("%Y-%m-%dT%H:%M")
+
+        if exdate_key in exdates:
+            continue
 
         d = dict(base)
         d["uid"] = occ_uid
@@ -1150,7 +1180,7 @@ def setup_calendar_routes() -> APIRouter:
             db.close()
 
     @router.delete("/events/{uid}")
-    async def delete_event(request: Request, uid: str):
+    async def delete_event(request: Request, uid: str, scope: str = "series"):
         owner = _require_user(request)
         try:
             base_uid = _resolve_base_uid(uid)
@@ -1159,7 +1189,22 @@ def setup_calendar_routes() -> APIRouter:
         db = SessionLocal()
         try:
             ev = _get_or_404_event(db, base_uid, owner)
+            is_occurrence_delete = scope in {"occurrence", "instance"} and "::" in uid and bool(ev.rrule)
             is_caldav = ev.calendar and ev.calendar.source == "caldav"
+            if is_occurrence_delete:
+                key = _occurrence_exdate_key(uid, ev)
+                if not key:
+                    raise HTTPException(400, "Invalid recurring occurrence uid")
+                exdates = _recurrence_exdates(ev)
+                if key not in exdates:
+                    exdates.append(key)
+                ev.recurrence_exdates = json.dumps(sorted(exdates))
+                if is_caldav:
+                    ev.caldav_sync_pending = "update"
+                db.commit()
+                if is_caldav:
+                    await _push_caldav_event_after_commit(owner, base_uid, "update")
+                return {"ok": True, "scope": "occurrence", "exdate": key}
             if is_caldav:
                 _record_caldav_delete_tombstone(db, ev, owner)
             db.delete(ev)

routes/chat_helpers.py

@@ -14,6 +14,7 @@ from core.database import Session as DBSession, ModelEndpoint
 from src.llm_core import normalize_model_id
 from src.endpoint_resolver import normalize_base
 from src.context_compactor import maybe_compact, trim_for_context
+from src.model_context import estimate_tokens
 from src.auth_helpers import effective_user
 from src.prompt_security import untrusted_context_message
 from routes.prefs_routes import _load_for_user as load_prefs_for_user
@@ -99,6 +100,11 @@ class ChatContext:
     uprefs: dict
     preset: PresetInfo
     preprocessed: PreprocessedMessage
+    context_trimmed: bool = False
+    context_messages_before_trim: int = 0
+    context_messages_after_trim: int = 0
+    context_tokens_before_trim: int = 0
+    context_tokens_after_trim: int = 0
     # Documents auto-created server-side during preprocess (e.g. when an
     # attached fillable PDF gets rendered into a markdown editor doc).
     # The chat route emits a doc_update SSE event for each before streaming
@@ -777,7 +783,12 @@ async def build_chat_context(
     messages, context_length, was_compacted = await maybe_compact(
         sess, sess.endpoint_url, sess.model, messages, sess.headers, owner=user,
     )
+    _before_trim_messages = len(messages)
+    _before_trim_tokens = estimate_tokens(messages)
     messages = trim_for_context(messages, context_length)
+    _after_trim_messages = len(messages)
+    _after_trim_tokens = estimate_tokens(messages)
+    _context_trimmed = _after_trim_messages < _before_trim_messages or _after_trim_tokens < _before_trim_tokens
 
     return ChatContext(
         preface=preface,
@@ -791,6 +802,11 @@ async def build_chat_context(
         uprefs=uprefs,
         preset=preset,
         preprocessed=preprocessed,
+        context_trimmed=_context_trimmed,
+        context_messages_before_trim=_before_trim_messages,
+        context_messages_after_trim=_after_trim_messages,
+        context_tokens_before_trim=_before_trim_tokens,
+        context_tokens_after_trim=_after_trim_tokens,
         auto_opened_docs=auto_opened_docs,
         uploaded_files=uploaded_files,
     )

routes/chat_routes.py

@@ -3,6 +3,7 @@
 import asyncio
 import json
 import os
+import re
 import time
 import logging
 from datetime import datetime
@@ -40,7 +41,7 @@ from routes.chat_helpers import (
     clean_thinking_for_save,
     _enforce_chat_privileges,
 )
-from src.action_intents import classify_tool_intent as _classify_tool_intent
+from src.action_intents import ToolIntent, classify_tool_intent as _classify_tool_intent
 from src.tool_policy import build_effective_tool_policy
 
 logger = logging.getLogger(__name__)
@@ -63,6 +64,78 @@ def _stream_set(session_id: str, **fields) -> None:
     rec.update(fields)
 
 
+def _message_plain_text(content: Any) -> str:
+    if isinstance(content, list):
+        parts: List[str] = []
+        for block in content:
+            if isinstance(block, dict):
+                text = block.get("text")
+                if isinstance(text, str):
+                    parts.append(text)
+            elif isinstance(block, str):
+                parts.append(block)
+        return " ".join(parts)
+    return str(content or "")
+
+
+def _last_user_plain_text(messages: List[Dict[str, Any]]) -> str:
+    for msg in reversed(messages or []):
+        if msg.get("role") == "user":
+            return _message_plain_text(msg.get("content"))
+    return ""
+
+
+def _ensure_current_request_is_latest_user(messages: List[Dict[str, Any]], current_message: str) -> List[Dict[str, Any]]:
+    """Defensively keep detached streams grounded on the request that created them."""
+    current = str(current_message or "").strip()
+    if not current:
+        return messages
+    latest = _last_user_plain_text(messages).strip()
+    if latest == current or current in latest or latest in current:
+        return messages
+    logger.warning(
+        "[chat_stream] latest user context mismatch; appending current request for model call. latest=%r current=%r",
+        latest[:120],
+        current[:120],
+    )
+    repaired = list(messages or [])
+    repaired.append({"role": "user", "content": current})
+    return repaired
+
+
+_WEB_FOLLOWUP_RE = re.compile(
+    r"^\s*(?:(?:can|could|would|will)\s+you\s+)?"
+    r"(?:check|try\s+again|look(?:\s+now|\s+it\s+up)?|search(?:\s+now|\s+online|\s+it)?|"
+    r"do\s+it|again)\??\s*$",
+    re.I,
+)
+_RECENT_WEB_CONTEXT_RE = re.compile(
+    r"\b(?:weather|forecast|rain|raining|hourly|news|headlines|rate|exchange|currency|"
+    r"price|current|latest|search|look\s+up|online)\b",
+    re.I,
+)
+
+
+def _recent_session_text(sess, limit: int = 8, max_chars: int = 2000) -> str:
+    history = getattr(sess, "history", None) or getattr(sess, "_history", None) or []
+    chunks: List[str] = []
+    for msg in history[-limit:]:
+        content = getattr(msg, "content", None)
+        if content is None and isinstance(msg, dict):
+            content = msg.get("content")
+        text = _message_plain_text(content).strip()
+        if text:
+            chunks.append(text)
+    return " ".join(chunks)[-max_chars:]
+
+
+def _is_contextual_web_followup(message: str, sess) -> bool:
+    """Treat short retry/check replies as web lookups when recent context was web."""
+    if not message or not _WEB_FOLLOWUP_RE.search(message):
+        return False
+    return bool(_RECENT_WEB_CONTEXT_RE.search(_recent_session_text(sess)))
+
+
 def _resolve_request_workspace(request, raw_value) -> tuple:
     """Resolve the posted workspace for this request: (workspace, rejected).
 
@@ -510,6 +583,10 @@ def setup_chat_routes(
         # below). Skill extraction should only learn from real agent sessions,
         # not chats we quietly promoted for a notes/calendar intent.
         user_requested_agent = (chat_mode == "agent")
+        _search_enabled = (
+            str(allow_web_search).lower() == "true"
+            or str(use_web).lower() == "true"
+        )
         # Intent auto-escalation: if the user is clearly asking the assistant
         # to create a todo, reminder, or calendar event, promote chat → agent
         # for this turn so the LLM has access to manage_notes / manage_calendar.
@@ -527,6 +604,10 @@ def setup_chat_routes(
                 _tool_intent.category,
                 _tool_intent.reason,
             )
+        elif chat_mode == "chat" and _search_enabled:
+            chat_mode = "agent"
+            auto_escalated = True
+            logger.info("chat→agent auto-escalation: search enabled")
         active_doc_id = form_data.get("active_doc_id", "").strip()
         logger.info(f"[doc-inject] chat_mode={chat_mode}, active_doc_id={active_doc_id!r}")
 
@@ -619,6 +700,20 @@ def setup_chat_routes(
                     400,
                     "No model selected for this chat. Open the model picker and choose one before sending.",
                 )
+            if (
+                chat_mode == "chat"
+                and isinstance(message, str)
+                and (not _tool_intent or not _tool_intent.needs_tools)
+                and _is_contextual_web_followup(message, sess)
+            ):
+                _tool_intent = ToolIntent(True, "web", "contextual web lookup follow-up")
+                chat_mode = "agent"
+                auto_escalated = True
+                logger.info(
+                    "chat→agent auto-escalation: category=%s reason=%s",
+                    _tool_intent.category,
+                    _tool_intent.reason,
+                )
         except SessionNotFoundError as e:
             raise HTTPException(404, str(e))
         except (ValueError, ValidationError):
@@ -729,6 +824,15 @@ def setup_chat_routes(
                         logger.info(f"[doc-inject] found by ID: title={active_doc.title!r}, lang={active_doc.language!r}, is_active={active_doc.is_active}, content_len={len(active_doc.current_content or '')}")
                 else:
                     logger.warning(f"[doc-inject] NOT FOUND by ID {active_doc_id}")
+            if not active_doc:
+                _email_doc_q = _doc_db.query(DBDocument).filter(
+                    DBDocument.session_id == session,
+                    DBDocument.is_active == True,
+                    DBDocument.language == "email",
+                )
+                active_doc = _owner_session_filter(_email_doc_q, ctx.user).order_by(DBDocument.updated_at.desc()).first()
+                if active_doc:
+                    logger.info(f"[doc-inject] found email draft by session fallback: title={active_doc.title!r}")
             if not active_doc:
                 _session_doc_q = _doc_db.query(DBDocument).filter(
                     DBDocument.session_id == session,
@@ -772,14 +876,31 @@ def setup_chat_routes(
         # by default without having to send allow_bash in every request.
         if allow_bash is not None and str(allow_bash).lower() != "true":
             disabled_tools.add("bash")
-        _explicit_web_intent = bool(_tool_intent and _tool_intent.category == "web")
         if (
             allow_web_search is not None
             and str(allow_web_search).lower() != "true"
-            and not _explicit_web_intent
         ):
             disabled_tools.add("web_search")
             disabled_tools.add("web_fetch")
+        if _explicit_web_intent:
+            # A direct lookup/search request should not drift into personal
+            # tools or shell fallbacks. We still keep web_search/web_fetch
+            # available even when the frontend toggle is stale/falsy because
+            # the user's words are the stronger signal.
+            disabled_tools.update({
+                "bash", "python",
+                "search_chats", "manage_skills", "manage_memory",
+                "read_file", "write_file", "edit_file",
+                "create_document", "edit_document", "update_document",
+                "send_email", "reply_to_email",
+                "manage_notes", "manage_calendar", "manage_tasks",
+                "api_call", "builtin_browser",
+            })
+            disabled_tools.discard("web_search")
+            disabled_tools.discard("web_fetch")
+        elif _search_enabled:
+            disabled_tools.discard("web_search")
+            disabled_tools.discard("web_fetch")
 
         # Nobody/incognito mode: deny tools that would expose the user's
         # persistent memory, past chats, or other identity-linked data.
@@ -790,19 +911,19 @@ def setup_chat_routes(
                 "manage_skills",      # skill presets tied to user
             })
 
-        # Active email reader open → strip the tools that let the agent
-        # "drift" to a new compose: create_document (writes a fake email-
-        # shaped .md file) and send_email (sends fresh to a recipient the
-        # agent invented). With those gone, the only paths left for "write
-        # email saying X" are ui_control open_email_reply (draft) and
-        # reply_to_email (immediate send) — both of which use the open
-        # email's UID. Code-level enforcement instead of relying on a
-        # prompt rule the model can ignore.
+        # Active email reader open → strip the tools that let the agent drift
+        # away from the visible email or skip review. The only allowed compose
+        # path is ui_control open_email_reply, which opens the same draft editor
+        # as the Reply button with the generated body pre-filled. This prevents
+        # the model from falling back to direct SMTP when it botches a draft
+        # call, and prevents fake email-shaped documents.
         if active_email_ctx and active_email_ctx.get("uid"):
             disabled_tools.update({
                 "create_document",
                 "send_email",
+                "reply_to_email",
                 "mcp__email__send_email",
+                "mcp__email__reply_to_email",
             })
 
         # Enforce per-user privileges
@@ -830,7 +951,10 @@ def setup_chat_routes(
         from src.settings import get_setting
         _global_disabled = get_setting("disabled_tools", [])
         if _global_disabled and isinstance(_global_disabled, list):
-            explicit_web_allowed = allow_web_search is not None and str(allow_web_search).lower() == "true"
+            explicit_web_allowed = (
+                _explicit_web_intent
+                or (allow_web_search is not None and str(allow_web_search).lower() == "true")
+            )
             if explicit_web_allowed:
                 disabled_tools.update(t for t in _global_disabled if t not in {"web_search", "web_fetch"})
             else:
@@ -1048,13 +1172,16 @@ def setup_chat_routes(
                     _active_streams.pop(session, None)
                     return
 
-            messages = ctx.messages
+            messages = _ensure_current_request_is_latest_user(ctx.messages, message)
 
             # Auto-compact notification
             if ctx.was_compacted:
                 yield f"data: {json.dumps({'type': 'compacted', 'context_length': ctx.context_length})}\n\n"
+            if ctx.context_trimmed and not ctx.was_compacted:
+                yield f"data: {json.dumps({'type': 'context_trimmed', 'data': {'context_length': ctx.context_length, 'messages_before': ctx.context_messages_before_trim, 'messages_after': ctx.context_messages_after_trim, 'tokens_before': ctx.context_tokens_before_trim, 'tokens_after': ctx.context_tokens_after_trim}})}\n\n"
 
             full_response = ""
+            thinking_response = ""
             last_metrics = None
 
             # Configured fallback chain for the default chat model. Tried in
@@ -1144,7 +1271,9 @@ def setup_chat_routes(
                                     # Forward them so the client can show a thinking
                                     # indicator, but don't fold them into the saved
                                     # reply (mirrors the rewrite path below).
-                                    if not data.get("thinking"):
+                                    if data.get("thinking"):
+                                        thinking_response += data["delta"]
+                                    else:
                                         full_response += data["delta"]
                                         _stream_set(session, partial=full_response)
                                     yield chunk
@@ -1164,6 +1293,12 @@ def setup_chat_routes(
                                     _reported_model = last_metrics.get("model")
                                     last_metrics["requested_model"] = _requested_model
                                     last_metrics["model"] = _reported_model or _actual_model or _answered_by or _requested_model
+                                    if ctx.context_trimmed:
+                                        last_metrics["context_trimmed"] = True
+                                        last_metrics["context_messages_before_trim"] = ctx.context_messages_before_trim
+                                        last_metrics["context_messages_after_trim"] = ctx.context_messages_after_trim
+                                        last_metrics["context_tokens_before_trim"] = ctx.context_tokens_before_trim
+                                        last_metrics["context_tokens_after_trim"] = ctx.context_tokens_after_trim
                                     if ctx.context_length and last_metrics.get("input_tokens"):
                                         pct = min(round((last_metrics["input_tokens"] / ctx.context_length) * 100, 1), 100.0)
                                         last_metrics["context_percent"] = pct
@@ -1206,8 +1341,11 @@ def setup_chat_routes(
                                 }
                                 yield f'data: {json.dumps({"type": "metrics", "data": last_metrics})}\n\n'
                             if full_response:
+                                _metrics_to_save = dict(last_metrics or {})
+                                if thinking_response.strip() and not _metrics_to_save.get("thinking"):
+                                    _metrics_to_save["thinking"] = thinking_response.strip()
                                 _saved_id = save_assistant_response(
-                                    sess, session_manager, session, full_response, last_metrics,
+                                    sess, session_manager, session, full_response, _metrics_to_save,
                                     character_name=ctx.preset.character_name,
                                     web_sources=web_sources,
                                     rag_sources=ctx.rag_sources,
@@ -1220,7 +1358,7 @@ def setup_chat_routes(
                                     yield f'data: {json.dumps({"type": "message_saved", "id": _saved_id})}\n\n'
                                 run_post_response_tasks(
                                     sess, session_manager, session, message, full_response,
-                                    last_metrics, ctx.uprefs, memory_manager, memory_vector, webhook_manager,
+                                    _metrics_to_save, ctx.uprefs, memory_manager, memory_vector, webhook_manager,
                                     incognito=incognito, compare_mode=compare_mode,
                                     character_name=ctx.preset.character_name,
                                     owner=_user,
@@ -1272,7 +1410,9 @@ def setup_chat_routes(
                     _max_rounds = max(1, min(_max_rounds, 200))
 
                     _forced_tools = None
-                    if allow_web_search is not None and str(allow_web_search).lower() == "true":
+                    if _explicit_web_intent:
+                        _forced_tools = {"web_search", "web_fetch"}
+                    elif _search_enabled:
                         _forced_tools = {"web_search", "web_fetch"}
 
                     async for chunk in stream_agent_loop(
@@ -1306,7 +1446,9 @@ def setup_chat_routes(
                                     # Reasoning tokens arrive flagged thinking:true.
                                     # Forward them for the live indicator, but keep
                                     # them out of the saved reply (same as chat mode).
-                                    if not data.get("thinking"):
+                                    if data.get("thinking"):
+                                        thinking_response += data["delta"]
+                                    else:
                                         full_response += data["delta"]
                                         _stream_set(session, partial=full_response)
                                     yield chunk
@@ -1344,15 +1486,26 @@ def setup_chat_routes(
                                     _reported_model = last_metrics.get("model")
                                     last_metrics["requested_model"] = last_metrics.get("requested_model") or _requested_model
                                     last_metrics["model"] = _reported_model or _actual_model or _answered_by or _requested_model
+                                    if ctx.context_trimmed:
+                                        last_metrics["context_trimmed"] = True
+                                        last_metrics["context_messages_before_trim"] = ctx.context_messages_before_trim
+                                        last_metrics["context_messages_after_trim"] = ctx.context_messages_after_trim
+                                        last_metrics["context_tokens_before_trim"] = ctx.context_tokens_before_trim
+                                        last_metrics["context_tokens_after_trim"] = ctx.context_tokens_after_trim
                                     yield f'data: {json.dumps({"type": "metrics", "data": last_metrics})}\n\n'
                             except json.JSONDecodeError:
                                 yield chunk
                         elif chunk.startswith("event: "):
                             yield chunk
                         elif chunk == "data: [DONE]\n\n":
-                            if full_response:
+                            _has_tool_events = bool((last_metrics or {}).get("tool_events"))
+                            if full_response or _has_tool_events:
+                                _response_to_save = full_response or "Done."
+                                _metrics_to_save = dict(last_metrics or {})
+                                if thinking_response.strip() and not _metrics_to_save.get("thinking"):
+                                    _metrics_to_save["thinking"] = thinking_response.strip()
                                 _saved_id = save_assistant_response(
-                                    sess, session_manager, session, full_response, last_metrics,
+                                    sess, session_manager, session, _response_to_save, _metrics_to_save,
                                     character_name=ctx.preset.character_name,
                                     web_sources=web_sources,
                                     rag_sources=ctx.rag_sources,
@@ -1362,8 +1515,8 @@ def setup_chat_routes(
                                 if _saved_id:
                                     yield f'data: {json.dumps({"type": "message_saved", "id": _saved_id})}\n\n'
                                 run_post_response_tasks(
-                                    sess, session_manager, session, message, full_response,
-                                    last_metrics, ctx.uprefs, memory_manager, memory_vector, webhook_manager,
+                                    sess, session_manager, session, message, _response_to_save,
+                                    _metrics_to_save, ctx.uprefs, memory_manager, memory_vector, webhook_manager,
                                     incognito=incognito, compare_mode=compare_mode,
                                     character_name=ctx.preset.character_name,
                                                             agent_rounds=_agent_rounds,

routes/contacts/__init__.py

@@ -0,0 +1,5 @@
+"""Contacts route domain package (slice 2e, #4082/#4071).
+
+Contains contacts_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/contacts_routes.py re-exports from here.
+"""

routes/contacts/contacts_routes.py

@@ -0,0 +1,916 @@
+"""
+contacts_routes.py
+
+CardDAV contacts integration. Reads from local Radicale, supports
+search and adding new contacts.
+"""
+
+import re
+import logging
+import uuid
+import json
+import csv
+import io
+import os
+import inspect
+import httpx
+from pathlib import Path
+from datetime import datetime
+from urllib.parse import urljoin, urlparse, urlunparse
+
+from core.log_safety import redact_url
+from fastapi import APIRouter, Query, Depends, Response, HTTPException
+from typing import List, Dict, Optional
+
+from core.middleware import require_admin
+from src.url_safety import check_outbound_url
+
+logger = logging.getLogger(__name__)
+
+from src.constants import DATA_DIR as _DATA_DIR, SETTINGS_FILE as _SETTINGS_FILE, CONTACTS_FILE as _CONTACTS_FILE
+DATA_DIR = Path(_DATA_DIR)
+SETTINGS_FILE = Path(_SETTINGS_FILE)
+LOCAL_CONTACTS_FILE = Path(_CONTACTS_FILE)
+
+
+def _load_settings():
+    if SETTINGS_FILE.exists():
+        return json.loads(SETTINGS_FILE.read_text(encoding="utf-8"))
+    return {}
+
+
+def _save_settings(settings):
+    from core.atomic_io import atomic_write_json
+    atomic_write_json(str(SETTINGS_FILE), settings, indent=2)
+
+
+def _get_carddav_config():
+    import os
+    settings = _load_settings()
+    password = settings.get("carddav_password", os.environ.get("CARDDAV_PASSWORD", ""))
+    if password and "carddav_password" in settings:
+        from src.secret_storage import decrypt
+        password = decrypt(password)
+    return {
+        "url": settings.get("carddav_url", os.environ.get("CARDDAV_URL", "")),
+        "username": settings.get("carddav_username", os.environ.get("CARDDAV_USERNAME", "")),
+        "password": password,
+    }
+
+
+def _carddav_configured(cfg: Optional[Dict] = None) -> bool:
+    cfg = cfg or _get_carddav_config()
+    return bool((cfg.get("url") or "").strip())
+
+
+def _validate_carddav_url(url: str) -> str:
+    cleaned = (url if isinstance(url, str) else "").strip().rstrip("/")
+    ok, reason = check_outbound_url(
+        cleaned,
+        block_private=os.getenv("CARDDAV_BLOCK_PRIVATE_IPS", "false").lower() == "true",
+    )
+    if not ok:
+        raise ValueError(f"Rejected CardDAV URL: {reason}")
+    return cleaned
+
+
+def _carddav_base_url(cfg: Dict) -> str:
+    return _validate_carddav_url(cfg.get("url") or "")
+
+
+def _normalize_contact(contact: Dict) -> Dict:
+    emails = []
+    for e in contact.get("emails") or ([] if not contact.get("email") else [contact.get("email")]):
+        e = str(e or "").strip()
+        if e and e not in emails:
+            emails.append(e)
+    phones = []
+    for p in contact.get("phones") or ([] if not contact.get("phone") else [contact.get("phone")]):
+        p = str(p or "").strip()
+        if p and p not in phones:
+            phones.append(p)
+    name = str(contact.get("name") or "").strip()
+    if not name and emails:
+        name = emails[0].split("@")[0]
+    address = str(contact.get("address") or "").strip()
+    return {
+        "uid": str(contact.get("uid") or uuid.uuid4()),
+        "name": name,
+        "emails": emails,
+        "phones": phones,
+        "address": address,
+    }
+
+
+def _load_local_contacts() -> List[Dict]:
+    try:
+        if not LOCAL_CONTACTS_FILE.exists():
+            return []
+        data = json.loads(LOCAL_CONTACTS_FILE.read_text(encoding="utf-8"))
+        rows = data.get("contacts", data) if isinstance(data, dict) else data
+        return [_normalize_contact(c) for c in (rows or []) if isinstance(c, dict)]
+    except Exception as e:
+        logger.error(f"Failed to load local contacts: {e}")
+        return []
+
+
+def _save_local_contacts(contacts: List[Dict]) -> None:
+    from core.atomic_io import atomic_write_json
+    DATA_DIR.mkdir(parents=True, exist_ok=True)
+    atomic_write_json(str(LOCAL_CONTACTS_FILE), {"contacts": [_normalize_contact(c) for c in contacts]}, indent=2)
+    _contact_cache["contacts"] = [_normalize_contact(c) for c in contacts]
+    _contact_cache["fetched_at"] = datetime.utcnow()
+
+
+# ── vCard parsing ──
+
+def _vunesc(value: str) -> str:
+    """Reverse _vesc() — turn escaped vCard text back into the raw value.
+    Order matters: handle \\n/\\, /\\; first, backslash-unescape last."""
+    if not value:
+        return value
+    out = []
+    i = 0
+    while i < len(value):
+        ch = value[i]
+        if ch == "\\" and i + 1 < len(value):
+            nxt = value[i + 1]
+            if nxt in ("n", "N"):
+                out.append("\n")
+            elif nxt in (",", ";", "\\"):
+                out.append(nxt)
+            else:
+                out.append(nxt)
+            i += 2
+        else:
+            out.append(ch)
+            i += 1
+    return "".join(out)
+
+
+def _parse_vcards(text: str) -> List[Dict]:
+    """Parse a stream of vCards into dicts with name, email, phone."""
+    # Unfold RFC 6350 3.2 line folding first: a CRLF/LF followed by a single
+    # space or tab is a continuation of the previous logical line. Real
+    # CardDAV servers (Radicale, iCloud, Apple/Google) fold long EMAIL / FN /
+    # PHOTO lines, and splitting on raw newlines without unfolding dropped the
+    # continuation (e.g. "...@example\n .com" lost the ".com"), truncating the
+    # email/name.
+    text = re.sub(r"\r\n[ \t]", "", text or "")
+    text = re.sub(r"\n[ \t]", "", text)
+    contacts = []
+    for block in re.split(r"BEGIN:VCARD", text):
+        if not block.strip():
+            continue
+        contact = {"name": "", "emails": [], "phones": [], "uid": "", "address": ""}
+        for line in block.split("\n"):
+            line = line.strip()
+            # Strip an optional RFC 6350 group prefix (e.g. "item1.EMAIL;...")
+            # that Apple Contacts / iCloud / many CardDAV servers emit by
+            # default — without this the property-name checks below miss those
+            # lines and silently drop the email / phone. The group token only
+            # precedes the property name, so it is safe to strip for matching
+            # and value extraction, and a no-op for non-grouped lines.
+            name_part = re.sub(r"^[A-Za-z0-9-]+\.", "", line, count=1)
+            if name_part.startswith("FN:") or name_part.startswith("FN;"):
+                contact["name"] = _vunesc(name_part.split(":", 1)[1]) if ":" in name_part else ""
+            elif name_part.startswith("EMAIL"):
+                # Handle EMAIL:foo@bar OR EMAIL;TYPE=...:foo@bar OR EMAIL;PREF=1:foo@bar
+                if ":" in name_part:
+                    email_addr = _vunesc(name_part.split(":", 1)[1])
+                    if email_addr and email_addr not in contact["emails"]:
+                        contact["emails"].append(email_addr)
+            elif name_part.startswith("TEL"):
+                if ":" in name_part:
+                    phone = _vunesc(name_part.split(":", 1)[1])
+                    if phone and phone not in contact["phones"]:
+                        contact["phones"].append(phone)
+            elif name_part.startswith("ADR"):
+                # vCard ADR is 7 semicolon-separated components:
+                # post-office-box;extended-address;street;locality;region;postal-code;country.
+                # Recover a human-readable string by joining non-empty
+                # components with ", ".
+                if ":" in name_part:
+                    raw = name_part.split(":", 1)[1]
+                    parts = [_vunesc(p).strip() for p in raw.split(";")]
+                    contact["address"] = ", ".join(p for p in parts if p)
+            elif name_part.startswith("UID:"):
+                contact["uid"] = _vunesc(name_part[4:])
+        if contact["name"] or contact["emails"]:
+            contacts.append(contact)
+    return contacts
+
+
+def _vesc(value: str) -> str:
+    """Escape a vCard property VALUE per RFC 6350 §3.4: backslash, comma,
+    semicolon, and newlines. Without this, a name like 'Sekisui House,Ltd'
+    or any value containing a newline produces a malformed vCard (broken
+    N/FN fields) or could inject arbitrary properties."""
+    return (
+        (value or "")
+        .replace("\\", "\\\\")
+        .replace("\n", "\\n")
+        .replace("\r", "")
+        .replace(",", "\\,")
+        .replace(";", "\\;")
+    )
+
+
+def _build_vcard(name: str, email: str, uid: Optional[str] = None,
+                 emails: Optional[List[str]] = None,
+                 phones: Optional[List[str]] = None,
+                 address: Optional[str] = None) -> str:
+    """Build a vCard. Accepts either a single `email` (legacy callers) or
+    full `emails`/`phones` lists (edit path). The first email is marked
+    PREF=1. All values are RFC-6350-escaped."""
+    if not uid:
+        uid = str(uuid.uuid4())
+    # Normalize email lists — `email` arg is a convenience for single-email
+    # creation; `emails` (if given) is authoritative.
+    email_list = [e.strip() for e in (emails if emails is not None else ([email] if email else [])) if e and e.strip()]
+    phone_list = [p.strip() for p in (phones or []) if p and p.strip()]
+    # Try to split name into first/last
+    parts = name.strip().split()
+    if len(parts) >= 2:
+        first = parts[0]
+        last = " ".join(parts[1:])
+    else:
+        first = name
+        last = ""
+    # N field is structured (5 components separated by ';') — escape each
+    # component individually so a comma in the name doesn't split it.
+    n_field = f"{_vesc(last)};{_vesc(first)};;;"
+    lines = [
+        "BEGIN:VCARD",
+        "VERSION:4.0",
+        f"UID:{_vesc(uid)}",
+        f"FN:{_vesc(name)}",
+        f"N:{n_field}",
+    ]
+    for i, em in enumerate(email_list):
+        # First email is the preferred one.
+        lines.append(f"EMAIL;PREF=1:{_vesc(em)}" if i == 0 else f"EMAIL:{_vesc(em)}")
+    for ph in phone_list:
+        lines.append(f"TEL:{_vesc(ph)}")
+    # Address: stuff the whole human-readable string into the street
+    # component of ADR. vCard ADR has 7 semicolon-separated components:
+    # post-office-box;extended-address;street;locality;region;postal-code;country.
+    addr = (address or "").strip()
+    if addr:
+        lines.append(f"ADR:;;{_vesc(addr)};;;;")
+    lines.append("END:VCARD")
+    return "\r\n".join(lines) + "\r\n"
+
+
+# ── In-memory cache ──
+
+_contact_cache = {"contacts": [], "fetched_at": None}
+
+
+def _abs_url(href: str) -> str:
+    """Combine a multistatus <href> (an absolute path like
+    /user/contacts/x.vcf) with the configured CardDAV server origin so we
+    get a fully-qualified URL to PUT/DELETE. Absolute hrefs are accepted only
+    for the configured origin; a cross-origin href is treated as a path on the
+    configured server so a malicious CardDAV response cannot redirect later
+    writes/deletes to cloud metadata or another host."""
+    cfg = _get_carddav_config()
+    base = _carddav_base_url(cfg)
+    base_p = urlparse(base)
+    joined = urljoin(base.rstrip("/") + "/", href or "")
+    joined_p = urlparse(joined)
+    if (joined_p.scheme, joined_p.netloc) != (base_p.scheme, base_p.netloc):
+        joined = urlunparse((base_p.scheme, base_p.netloc, joined_p.path or "/", "", joined_p.query, ""))
+    return _validate_carddav_url(joined)
+
+
+# CardDAV REPORT body — pull every card's etag + raw vCard in ONE request,
+# alongside the resource href. Lets us map each contact's UID to the real
+# server resource path (which is NOT always <uid>.vcf for contacts created
+# by other clients).
+_ADDRESSBOOK_QUERY = (
+    '<?xml version="1.0" encoding="utf-8"?>'
+    '<C:addressbook-query xmlns:D="DAV:" xmlns:C="urn:ietf:params:xml:ns:carddav">'
+    '<D:prop><D:getetag/><C:address-data/></D:prop>'
+    '<C:filter/>'
+    '</C:addressbook-query>'
+)
+
+
+def _fetch_via_report(cfg, auth):
+    """Try a CardDAV REPORT addressbook-query — returns contacts WITH an
+    `href` field, or None if the server doesn't support it / errors."""
+    from defusedxml import ElementTree as ET
+    try:
+        r = httpx.request(
+            "REPORT", cfg["url"],
+            content=_ADDRESSBOOK_QUERY.encode("utf-8"),
+            headers={"Content-Type": "application/xml; charset=utf-8", "Depth": "1"},
+            auth=auth, timeout=10,
+        )
+        if r.status_code not in (207, 200):
+            return None
+        root = ET.fromstring(r.text)
+        ns = {"D": "DAV:", "C": "urn:ietf:params:xml:ns:carddav"}
+        out = []
+        for resp in root.findall("D:response", ns):
+            href_el = resp.find("D:href", ns)
+            data_el = resp.find(".//C:address-data", ns)
+            if href_el is None or data_el is None or not (data_el.text or "").strip():
+                continue
+            parsed = _parse_vcards(data_el.text)
+            if not parsed:
+                continue
+            c = parsed[0]
+            c["href"] = href_el.text.strip()
+            out.append(c)
+        # If the REPORT parsed to ZERO contacts, don't trust it — some
+        # CardDAV servers treat an empty <filter/> as "match nothing" and
+        # return a valid-but-empty 207. Return None so the caller falls
+        # back to the plain GET (which lists everything). A genuinely empty
+        # address book just costs one extra GET that also returns nothing.
+        if not out:
+            return None
+        return out
+    except Exception as e:
+        logger.warning(f"CardDAV REPORT failed, falling back to GET: {e}")
+        return None
+
+
+def _fetch_contacts(force=False):
+    """Fetch all contacts. Uses CardDAV when configured, otherwise local JSON."""
+    if not force and _contact_cache["fetched_at"]:
+        age = (datetime.utcnow() - _contact_cache["fetched_at"]).total_seconds()
+        if age < 60:
+            return _contact_cache["contacts"]
+
+    cfg = _get_carddav_config()
+    if not _carddav_configured(cfg):
+        contacts = _load_local_contacts()
+        _contact_cache["contacts"] = contacts
+        _contact_cache["fetched_at"] = datetime.utcnow()
+        return contacts
+
+    try:
+        cfg["url"] = _carddav_base_url(cfg)
+        auth = None
+        if cfg["username"]:
+            auth = (cfg["username"], cfg["password"])
+        # Preferred path: REPORT gives us hrefs for reliable edit/delete.
+        contacts = _fetch_via_report(cfg, auth)
+        if contacts is None:
+            # Fallback: plain GET, concatenated vCards, no hrefs.
+            r = httpx.get(cfg["url"], auth=auth, timeout=10)
+            if r.status_code != 200:
+                logger.warning(f"CardDAV returned {r.status_code}")
+                return _contact_cache["contacts"]
+            contacts = _parse_vcards(r.text)
+        _contact_cache["contacts"] = contacts
+        _contact_cache["fetched_at"] = datetime.utcnow()
+        return contacts
+    except Exception as e:
+        logger.error(f"Failed to fetch contacts: {e}")
+        return _contact_cache["contacts"]
+
+
+def _resolve_resource_url(uid: str) -> str:
+    """Map a contact UID to its real CardDAV resource URL. Uses the href
+    captured during fetch when available (handles contacts whose filename
+    != UID); falls back to the <uid>.vcf guess for app-created contacts or
+    when no href is known."""
+    def _lookup():
+        for c in _contact_cache.get("contacts", []):
+            if c.get("uid") == uid and c.get("href"):
+                return _abs_url(c["href"])
+        return None
+    found = _lookup()
+    if found:
+        return found
+    # Not in cache (or no href) — refresh once and retry before guessing.
+    try:
+        _fetch_contacts(force=True)
+    except Exception:
+        pass
+    return _lookup() or _vcard_url(uid)
+
+
+def _create_contact(name: str, email: str = "", address: str = "", phones: Optional[List[str]] = None) -> bool:
+    """Add a new contact via CardDAV or local contacts."""
+    email = (email or "").strip()
+    phone_list = [str(p or "").strip() for p in (phones or []) if str(p or "").strip()]
+    cfg = _get_carddav_config()
+    if not _carddav_configured(cfg):
+        contacts = _load_local_contacts()
+        email_l = email.lower()
+        for c in contacts:
+            if email_l and email_l in [e.lower() for e in c.get("emails", [])]:
+                return True
+            if phone_list and any(p in (c.get("phones") or []) for p in phone_list):
+                return True
+        contacts.append(_normalize_contact({
+            "name": name,
+            "emails": [email] if email else [],
+            "phones": phone_list,
+            "address": address,
+        }))
+        _save_local_contacts(contacts)
+        return True
+
+    contact_uid = str(uuid.uuid4())
+    vcard = _build_vcard(name, email, contact_uid, address=address, phones=phone_list)
+    try:
+        url = _carddav_base_url(cfg) + "/" + contact_uid + ".vcf"
+        auth = None
+        if cfg["username"]:
+            auth = (cfg["username"], cfg["password"])
+        r = httpx.put(
+            url,
+            data=vcard.encode("utf-8"),
+            headers={"Content-Type": "text/vcard; charset=utf-8"},
+            auth=auth,
+            timeout=10,
+        )
+        if r.status_code in (200, 201, 204):
+            # Invalidate cache
+            _contact_cache["fetched_at"] = None
+            return True
+        logger.warning(f"CardDAV PUT returned {r.status_code}: {r.text[:200]}")
+        return False
+    except Exception as e:
+        logger.error(f"Failed to create contact: {e}")
+        return False
+
+
+def _vcard_url(uid: str) -> str:
+    """The CardDAV resource URL for a given contact UID. The uid is URL-
+    encoded so a value containing '/', '..' or other path chars can't
+    escape the collection and target an arbitrary CardDAV resource."""
+    from urllib.parse import quote
+    cfg = _get_carddav_config()
+    return _carddav_base_url(cfg) + "/" + quote(uid, safe="") + ".vcf"
+
+
+def _import_vcards(text: str) -> Dict:
+    """Import a (possibly multi-card) .vcf blob. Each card is PUT to the
+    CardDAV server PRESERVING its full original content (ADR/ORG/photo/
+    etc.) — we don't rebuild it, just ensure it has VERSION + UID and
+    normalize line endings. Returns {imported, failed, total}."""
+    from urllib.parse import quote
+    cfg = _get_carddav_config()
+    if not cfg.get("url"):
+        parsed = _parse_vcards(text)
+        contacts = _load_local_contacts()
+        existing = {
+            e.lower()
+            for c in contacts
+            for e in (c.get("emails") or [])
+            if e
+        }
+        imported = 0
+        for c in parsed:
+            emails = [e for e in (c.get("emails") or []) if e]
+            if emails and any(e.lower() in existing for e in emails):
+                continue
+            contacts.append(_normalize_contact(c))
+            for e in emails:
+                existing.add(e.lower())
+            imported += 1
+        if imported:
+            _save_local_contacts(contacts)
+        return {"imported": imported, "failed": 0, "total": len(parsed)}
+    try:
+        base_url = _carddav_base_url(cfg)
+    except ValueError as e:
+        logger.warning("CardDAV import URL rejected: %s", e)
+        return {"imported": 0, "failed": 0, "total": 0, "error": str(e)}
+    auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
+    # Split into individual cards. re.split drops the BEGIN line, so we
+    # re-add it. Normalize CRLF.
+    raw = (text or "").replace("\r\n", "\n").replace("\r", "\n")
+    blocks = []
+    for chunk in raw.split("BEGIN:VCARD"):
+        chunk = chunk.strip()
+        if not chunk:
+            continue
+        # Trim anything after END:VCARD (defensive).
+        end = chunk.upper().find("END:VCARD")
+        body = chunk[: end + len("END:VCARD")] if end != -1 else chunk
+        blocks.append("BEGIN:VCARD\n" + body)
+    imported = 0
+    failed = 0
+    for block in blocks:
+        # Extract or assign a UID.
+        m = re.search(r"^UID:(.+)$", block, re.MULTILINE)
+        uid = (m.group(1).strip() if m else "") or str(uuid.uuid4())
+        if not m:
+            # Inject a UID right after the VERSION line (or after BEGIN).
+            if re.search(r"^VERSION:", block, re.MULTILINE):
+                block = re.sub(r"(^VERSION:.*$)", r"\1\nUID:" + uid, block, count=1, flags=re.MULTILINE)
+            else:
+                block = block.replace("BEGIN:VCARD", f"BEGIN:VCARD\nVERSION:4.0\nUID:{uid}", 1)
+        elif not re.search(r"^VERSION:", block, re.MULTILINE):
+            block = block.replace("BEGIN:VCARD", "BEGIN:VCARD\nVERSION:4.0", 1)
+        vcard = block.replace("\n", "\r\n") + "\r\n"
+        url = base_url + "/" + quote(uid, safe="") + ".vcf"
+        try:
+            r = httpx.put(
+                url, data=vcard.encode("utf-8"),
+                headers={"Content-Type": "text/vcard; charset=utf-8"},
+                auth=auth, timeout=15,
+            )
+            if r.status_code in (200, 201, 204):
+                imported += 1
+            else:
+                failed += 1
+                logger.warning(f"Import PUT {uid} returned {r.status_code}: {r.text[:120]}")
+        except Exception as e:
+            failed += 1
+            logger.error(f"Import PUT {uid} failed: {e}")
+    if imported:
+        _contact_cache["fetched_at"] = None
+    return {"imported": imported, "failed": failed, "total": len(blocks)}
+
+
+def _import_csv_contacts(text: str) -> Dict:
+    """Import contacts from CSV. Supports common headers:
+    name/full_name/display_name, email/email_address/e-mail, phone/tel.
+    Falls back to first columns as name,email,phone when no headers exist."""
+    raw = (text or "").strip()
+    if not raw:
+        return {"imported": 0, "failed": 0, "total": 0, "error": "No CSV data found"}
+
+    try:
+        sample = raw[:2048]
+        dialect = csv.Sniffer().sniff(sample)
+    except Exception:
+        dialect = csv.excel
+
+    stream = io.StringIO(raw)
+    try:
+        has_header = csv.Sniffer().has_header(raw[:2048])
+    except Exception:
+        has_header = True
+
+    rows = []
+    if has_header:
+        reader = csv.DictReader(stream, dialect=dialect)
+        for row in reader:
+            lowered = {str(k or "").strip().lower(): (v or "").strip() for k, v in row.items()}
+            name = (
+                lowered.get("name") or lowered.get("full name") or lowered.get("full_name")
+                or lowered.get("display name") or lowered.get("display_name")
+                or lowered.get("fn") or ""
+            )
+            email = (
+                lowered.get("email") or lowered.get("email address")
+                or lowered.get("email_address") or lowered.get("e-mail")
+                or lowered.get("mail") or ""
+            )
+            phone = lowered.get("phone") or lowered.get("telephone") or lowered.get("tel") or ""
+            rows.append((name, email, phone))
+    else:
+        stream.seek(0)
+        reader = csv.reader(stream, dialect=dialect)
+        for row in reader:
+            cols = [(c or "").strip() for c in row]
+            if not any(cols):
+                continue
+            rows.append((
+                cols[0] if len(cols) > 0 else "",
+                cols[1] if len(cols) > 1 else "",
+                cols[2] if len(cols) > 2 else "",
+            ))
+
+    imported = 0
+    failed = 0
+    total = 0
+    existing_emails = {
+        e.lower()
+        for c in _fetch_contacts()
+        for e in (c.get("emails") or [])
+        if e
+    }
+    for name, email, phone in rows:
+        email = (email or "").strip()
+        name = (name or "").strip() or (email.split("@")[0] if email else "")
+        if not email:
+            continue
+        total += 1
+        if email.lower() in existing_emails:
+            continue
+        ok = _create_contact(name, email)
+        if ok:
+            imported += 1
+            existing_emails.add(email.lower())
+            # If the CSV had a phone number, rewrite the just-created row
+            # through the richer update path so phone lands in CardDAV too.
+            if phone:
+                try:
+                    contacts = _fetch_contacts(force=True)
+                    created = next((c for c in contacts if email.lower() in [e.lower() for e in c.get("emails", [])]), None)
+                    if created and created.get("uid"):
+                        _update_contact(created["uid"], name, [email], [phone])
+                except Exception:
+                    pass
+        else:
+            failed += 1
+
+    if imported:
+        _contact_cache["fetched_at"] = None
+    return {"imported": imported, "failed": failed, "total": total}
+
+
+def _contacts_to_vcf(contacts: List[Dict]) -> str:
+    return "".join(
+        _build_vcard(
+            c.get("name") or ((c.get("emails") or [""])[0].split("@")[0] if c.get("emails") else "Contact"),
+            "",
+            uid=c.get("uid") or str(uuid.uuid4()),
+            emails=c.get("emails") or [],
+            phones=c.get("phones") or [],
+        )
+        for c in contacts
+    )
+
+
+def _contacts_to_csv(contacts: List[Dict]) -> str:
+    out = io.StringIO()
+    writer = csv.writer(out)
+    writer.writerow(["name", "email", "phone"])
+    for c in contacts:
+        emails = c.get("emails") or [""]
+        phones = c.get("phones") or [""]
+        max_len = max(len(emails), len(phones), 1)
+        for i in range(max_len):
+            writer.writerow([
+                c.get("name") or "",
+                emails[i] if i < len(emails) else "",
+                phones[i] if i < len(phones) else "",
+            ])
+    return out.getvalue()
+
+
+def _update_contact(uid: str, name: str, emails: List[str], phones: List[str], address: str = "") -> bool:
+    """Rewrite an existing contact via CardDAV or local contacts."""
+    cfg = _get_carddav_config()
+    if not _carddav_configured(cfg):
+        contacts = _load_local_contacts()
+        found = False
+        out = []
+        for c in contacts:
+            if c.get("uid") == uid:
+                # Preserve existing address when caller passes "" (only
+                # updating name/emails/phones, not touching address).
+                addr = address if address else c.get("address", "")
+                out.append(_normalize_contact({"uid": uid, "name": name, "emails": emails, "phones": phones, "address": addr}))
+                found = True
+            else:
+                out.append(c)
+        if not found:
+            out.append(_normalize_contact({"uid": uid, "name": name, "emails": emails, "phones": phones, "address": address}))
+        _save_local_contacts(out)
+        return True
+
+    vcard = _build_vcard(name, "", uid=uid, emails=emails, phones=phones, address=address)
+    # Use the real resource href (handles externally-created contacts whose
+    # filename != UID); falls back to the <uid>.vcf guess.
+    try:
+        url = _resolve_resource_url(uid)
+        auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
+        r = httpx.put(
+            url,
+            data=vcard.encode("utf-8"),
+            headers={"Content-Type": "text/vcard; charset=utf-8"},
+            auth=auth,
+            timeout=10,
+        )
+        if r.status_code in (200, 201, 204):
+            _contact_cache["fetched_at"] = None
+            return True
+        logger.warning(f"CardDAV update PUT returned {r.status_code}: {r.text[:200]}")
+        return False
+    except Exception as e:
+        logger.error(f"Failed to update contact: {e}")
+        return False
+
+
+def _delete_contact(uid: str) -> bool:
+    """Delete a contact via CardDAV or local contacts."""
+    cfg = _get_carddav_config()
+    if not _carddav_configured(cfg):
+        contacts = _load_local_contacts()
+        remaining = [c for c in contacts if c.get("uid") != uid]
+        _save_local_contacts(remaining)
+        return True
+
+    try:
+        url = _resolve_resource_url(uid)
+        auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
+        r = httpx.delete(url, auth=auth, timeout=10)
+        if r.status_code in (200, 204, 404):
+            # Invalidate cache so the next fetch sees the server truth.
+            _contact_cache["fetched_at"] = None
+            # Verify: force a fresh fetch and check the UID is actually gone.
+            # A 404 on the guessed URL ({uid}.vcf) can mean the contact
+            # lives at a different resource URL — the DELETE missed it but
+            # we'd silently report success. This check catches that.
+            fresh = _fetch_contacts(force=True)
+            still_there = any(c.get("uid") == uid for c in fresh)
+            if still_there:
+                logger.warning(
+                    f"CardDAV DELETE reported success for {uid} "
+                    f"but UID still present after re-fetch — "
+                    f"resource URL may differ from {redact_url(url)}"
+                )
+                return False
+            if r.status_code == 404:
+                logger.info(f"CardDAV DELETE 404 for {uid} — already gone")
+            return True
+        logger.warning(f"CardDAV DELETE returned {r.status_code}: {r.text[:200]}")
+        return False
+    except Exception as e:
+        logger.error(f"Failed to delete contact: {e}")
+        return False
+
+
+# ── Routes ──
+
+def setup_contacts_routes():
+    router = APIRouter(prefix="/api/contacts", tags=["contacts"])
+
+    @router.get("/list")
+    async def list_contacts(_admin: str = Depends(require_admin)):
+        """List all contacts."""
+        contacts = _fetch_contacts()
+        return {"contacts": contacts, "count": len(contacts)}
+
+    @router.get("/search")
+    async def search_contacts(q: str = Query(""), _admin: str = Depends(require_admin)):
+        """Search contacts by name or email. Returns up to 10 matches."""
+        contacts = _fetch_contacts()
+        if not q:
+            return {"results": []}
+        q_lower = q.lower()
+        results = []
+        for c in contacts:
+            if q_lower in c["name"].lower():
+                results.append(c)
+                continue
+            for em in c["emails"]:
+                if q_lower in em.lower():
+                    results.append(c)
+                    break
+        return {"results": results[:10]}
+
+    @router.post("/add")
+    async def add_contact(data: dict, _admin: str = Depends(require_admin)):
+        """Add a new contact."""
+        name = (data.get("name") or "").strip()
+        email = (data.get("email") or "").strip()
+        phone = (data.get("phone") or "").strip()
+        phones = [str(p or "").strip() for p in (data.get("phones") or []) if str(p or "").strip()]
+        if phone and phone not in phones:
+            phones.insert(0, phone)
+        address = (data.get("address") or "").strip()
+        if not name and email:
+            name = email.split("@")[0]
+        if not name and not email and not phones and not address:
+            return {"success": False, "error": "Name, email, phone, or address required"}
+        if not name:
+            name = email.split("@")[0] if email else (phones[0] if phones else "Contact")
+        contacts = _fetch_contacts()
+        for c in contacts:
+            if email and email.lower() in [e.lower() for e in c.get("emails", [])]:
+                return {"success": True, "message": "Already exists", "contact": c}
+            if phones and any(p in (c.get("phones") or []) for p in phones):
+                return {"success": True, "message": "Already exists", "contact": c}
+        create_params = inspect.signature(_create_contact).parameters
+        if "phones" in create_params:
+            ok = _create_contact(name, email, address, phones=phones)
+        elif len(create_params) >= 3:
+            ok = _create_contact(name, email, address)
+        else:
+            ok = _create_contact(name, email)
+        # If a phone was provided, do an immediate update to thread it
+        # through (the simple _create_contact signature only takes name +
+        # email + address; phones happen via update).
+        if ok and phones and "phones" not in create_params:
+            try:
+                fresh = _fetch_contacts(force=True)
+                created = next((c for c in fresh if name == c.get("name") and (not email or email in c.get("emails", []))), None)
+                if created:
+                    _update_contact(
+                        created["uid"], name,
+                        created.get("emails", []),
+                        phones,
+                        address,
+                    )
+            except Exception:
+                pass
+        return {"success": ok}
+
+    @router.post("/import")
+    async def import_vcf(data: dict, _admin: str = Depends(require_admin)):
+        """Import contacts from .vcf or CSV. Body: {"vcf": "..."} or {"csv": "..."}."""
+        # Coerce defensively: a non-string vcf/text/csv (e.g. a number or list
+        # in the JSON body) would otherwise reach .strip() and 500 with an
+        # AttributeError instead of degrading to a clean "no data" response.
+        text = str(data.get("vcf") or data.get("text") or "")
+        csv_text = str(data.get("csv") or "")
+        if text.strip():
+            if "BEGIN:VCARD" not in text.upper():
+                return {"success": False, "error": "No vCard data found"}
+            result = _import_vcards(text)
+        elif csv_text.strip():
+            result = _import_csv_contacts(csv_text)
+        else:
+            return {"success": False, "error": "No contact data found"}
+        result["success"] = result.get("imported", 0) > 0
+        return result
+
+    @router.get("/export")
+    async def export_contacts(
+        format: str = Query("vcf", pattern="^(vcf|csv)$"),
+        _admin: str = Depends(require_admin),
+    ):
+        """Export all contacts as vCard or CSV."""
+        contacts = _fetch_contacts(force=True)
+        if format == "csv":
+            content = _contacts_to_csv(contacts)
+            media_type = "text/csv; charset=utf-8"
+            filename = "odysseus-contacts.csv"
+        else:
+            content = _contacts_to_vcf(contacts)
+            media_type = "text/vcard; charset=utf-8"
+            filename = "odysseus-contacts.vcf"
+        return Response(
+            content=content,
+            media_type=media_type,
+            headers={"Content-Disposition": f'attachment; filename="{filename}"'},
+        )
+
+    @router.get("/config")
+    async def get_config(_admin: str = Depends(require_admin)):
+        cfg = _get_carddav_config()
+        # Mask password
+        if cfg["password"]:
+            cfg["password"] = "***"
+        return cfg
+
+    @router.put("/config")
+    async def update_config(data: dict, _admin: str = Depends(require_admin)):
+        settings = _load_settings()
+        for key in ("carddav_url", "carddav_username", "carddav_password"):
+            if key in data:
+                if key == "carddav_url" and str(data[key] or "").strip():
+                    try:
+                        settings[key] = _validate_carddav_url(data[key])
+                    except ValueError as e:
+                        raise HTTPException(400, str(e))
+                else:
+                    value = data[key]
+                    if key == "carddav_password" and value:
+                        from src.secret_storage import encrypt
+                        value = encrypt(value)
+                    settings[key] = value
+        _save_settings(settings)
+        # Force re-fetch
+        _contact_cache["fetched_at"] = None
+        return {"success": True}
+
+    @router.delete("/clear")
+    async def clear_contacts(_admin: str = Depends(require_admin)):
+        """Clear all local contacts. If CardDAV is configured, only clears the local fallback cache."""
+        _save_local_contacts([])
+        return {"success": True}
+
+    # NOTE: the /{uid} routes are declared LAST so the literal paths above
+    # (/list, /search, /add, /config) win — otherwise PUT /config would
+    # match PUT /{uid} with uid="config".
+    @router.put("/{uid}")
+    async def edit_contact(uid: str, data: dict, _admin: str = Depends(require_admin)):
+        """Edit an existing contact — name / emails / phones / address."""
+        name = (data.get("name") or "").strip()
+        emails = data.get("emails")
+        phones = data.get("phones")
+        if emails is None and data.get("email"):
+            emails = [data["email"]]
+        emails = [e.strip() for e in (emails or []) if e and e.strip()]
+        phones = [p.strip() for p in (phones or []) if p and p.strip()]
+        address = (data.get("address") or "").strip()
+        if not name and not emails and not address:
+            return {"success": False, "error": "Name, email, or address required"}
+        if not name and emails:
+            name = emails[0].split("@")[0]
+        ok = _update_contact(uid, name, emails, phones, address)
+        return {"success": ok}
+
+    @router.delete("/{uid}")
+    async def delete_contact(uid: str, _admin: str = Depends(require_admin)):
+        """Delete a contact by UID."""
+        if not uid:
+            return {"success": False, "error": "UID required"}
+        ok = _delete_contact(uid)
+        return {"success": ok}
+
+    return router

routes/contacts_routes.py

@@ -1,900 +1,13 @@
-"""
-contacts_routes.py
+"""Backward-compat shim — canonical location is routes/contacts/contacts_routes.py.
 
-CardDAV contacts integration. Reads from local Radicale, supports
-search and adding new contacts.
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.contacts_routes``, ``from routes.contacts_routes import X``,
+``importlib.import_module("routes.contacts_routes")``, and string-targeted
+monkeypatches all operate on the same object the application actually uses.
 """
 
-import re
-import logging
-import uuid
-import json
-import csv
-import io
-import os
-import inspect
-import httpx
-from pathlib import Path
-from datetime import datetime
-from urllib.parse import urljoin, urlparse, urlunparse
+import sys as _sys
 
-from core.log_safety import redact_url
-from fastapi import APIRouter, Query, Depends, Response, HTTPException
-from typing import List, Dict, Optional
+from routes.contacts import contacts_routes as _canonical  # noqa: F401
 
-from core.middleware import require_admin
-from src.url_safety import check_outbound_url
-
-logger = logging.getLogger(__name__)
-
-from src.constants import DATA_DIR as _DATA_DIR, SETTINGS_FILE as _SETTINGS_FILE, CONTACTS_FILE as _CONTACTS_FILE
-DATA_DIR = Path(_DATA_DIR)
-SETTINGS_FILE = Path(_SETTINGS_FILE)
-LOCAL_CONTACTS_FILE = Path(_CONTACTS_FILE)
-
-
-def _load_settings():
-    if SETTINGS_FILE.exists():
-        return json.loads(SETTINGS_FILE.read_text(encoding="utf-8"))
-    return {}
-
-
-def _save_settings(settings):
-    from core.atomic_io import atomic_write_json
-    atomic_write_json(str(SETTINGS_FILE), settings, indent=2)
-
-
-def _get_carddav_config():
-    import os
-    settings = _load_settings()
-    password = settings.get("carddav_password", os.environ.get("CARDDAV_PASSWORD", ""))
-    if password and "carddav_password" in settings:
-        from src.secret_storage import decrypt
-        password = decrypt(password)
-    return {
-        "url": settings.get("carddav_url", os.environ.get("CARDDAV_URL", "")),
-        "username": settings.get("carddav_username", os.environ.get("CARDDAV_USERNAME", "")),
-        "password": password,
-    }
-
-
-def _carddav_configured(cfg: Optional[Dict] = None) -> bool:
-    cfg = cfg or _get_carddav_config()
-    return bool((cfg.get("url") or "").strip())
-
-
-def _validate_carddav_url(url: str) -> str:
-    cleaned = (url if isinstance(url, str) else "").strip().rstrip("/")
-    ok, reason = check_outbound_url(
-        cleaned,
-        block_private=os.getenv("CARDDAV_BLOCK_PRIVATE_IPS", "false").lower() == "true",
-    )
-    if not ok:
-        raise ValueError(f"Rejected CardDAV URL: {reason}")
-    return cleaned
-
-
-def _carddav_base_url(cfg: Dict) -> str:
-    return _validate_carddav_url(cfg.get("url") or "")
-
-
-def _normalize_contact(contact: Dict) -> Dict:
-    emails = []
-    for e in contact.get("emails") or ([] if not contact.get("email") else [contact.get("email")]):
-        e = str(e or "").strip()
-        if e and e not in emails:
-            emails.append(e)
-    phones = []
-    for p in contact.get("phones") or ([] if not contact.get("phone") else [contact.get("phone")]):
-        p = str(p or "").strip()
-        if p and p not in phones:
-            phones.append(p)
-    name = str(contact.get("name") or "").strip()
-    if not name and emails:
-        name = emails[0].split("@")[0]
-    address = str(contact.get("address") or "").strip()
-    return {
-        "uid": str(contact.get("uid") or uuid.uuid4()),
-        "name": name,
-        "emails": emails,
-        "phones": phones,
-        "address": address,
-    }
-
-
-def _load_local_contacts() -> List[Dict]:
-    try:
-        if not LOCAL_CONTACTS_FILE.exists():
-            return []
-        data = json.loads(LOCAL_CONTACTS_FILE.read_text(encoding="utf-8"))
-        rows = data.get("contacts", data) if isinstance(data, dict) else data
-        return [_normalize_contact(c) for c in (rows or []) if isinstance(c, dict)]
-    except Exception as e:
-        logger.error(f"Failed to load local contacts: {e}")
-        return []
-
-
-def _save_local_contacts(contacts: List[Dict]) -> None:
-    from core.atomic_io import atomic_write_json
-    DATA_DIR.mkdir(parents=True, exist_ok=True)
-    atomic_write_json(str(LOCAL_CONTACTS_FILE), {"contacts": [_normalize_contact(c) for c in contacts]}, indent=2)
-    _contact_cache["contacts"] = [_normalize_contact(c) for c in contacts]
-    _contact_cache["fetched_at"] = datetime.utcnow()
-
-
-# ── vCard parsing ──
-
-def _vunesc(value: str) -> str:
-    """Reverse _vesc() — turn escaped vCard text back into the raw value.
-    Order matters: handle \\n/\\, /\\; first, backslash-unescape last."""
-    if not value:
-        return value
-    out = []
-    i = 0
-    while i < len(value):
-        ch = value[i]
-        if ch == "\\" and i + 1 < len(value):
-            nxt = value[i + 1]
-            if nxt in ("n", "N"):
-                out.append("\n")
-            elif nxt in (",", ";", "\\"):
-                out.append(nxt)
-            else:
-                out.append(nxt)
-            i += 2
-        else:
-            out.append(ch)
-            i += 1
-    return "".join(out)
-
-
-def _parse_vcards(text: str) -> List[Dict]:
-    """Parse a stream of vCards into dicts with name, email, phone."""
-    # Unfold RFC 6350 3.2 line folding first: a CRLF/LF followed by a single
-    # space or tab is a continuation of the previous logical line. Real
-    # CardDAV servers (Radicale, iCloud, Apple/Google) fold long EMAIL / FN /
-    # PHOTO lines, and splitting on raw newlines without unfolding dropped the
-    # continuation (e.g. "...@example\n .com" lost the ".com"), truncating the
-    # email/name.
-    text = re.sub(r"\r\n[ \t]", "", text or "")
-    text = re.sub(r"\n[ \t]", "", text)
-    contacts = []
-    for block in re.split(r"BEGIN:VCARD", text):
-        if not block.strip():
-            continue
-        contact = {"name": "", "emails": [], "phones": [], "uid": "", "address": ""}
-        for line in block.split("\n"):
-            line = line.strip()
-            # Strip an optional RFC 6350 group prefix (e.g. "item1.EMAIL;...")
-            # that Apple Contacts / iCloud / many CardDAV servers emit by
-            # default — without this the property-name checks below miss those
-            # lines and silently drop the email / phone. The group token only
-            # precedes the property name, so it is safe to strip for matching
-            # and value extraction, and a no-op for non-grouped lines.
-            name_part = re.sub(r"^[A-Za-z0-9-]+\.", "", line, count=1)
-            if name_part.startswith("FN:") or name_part.startswith("FN;"):
-                contact["name"] = _vunesc(name_part.split(":", 1)[1]) if ":" in name_part else ""
-            elif name_part.startswith("EMAIL"):
-                # Handle EMAIL:foo@bar OR EMAIL;TYPE=...:foo@bar OR EMAIL;PREF=1:foo@bar
-                if ":" in name_part:
-                    email_addr = _vunesc(name_part.split(":", 1)[1])
-                    if email_addr and email_addr not in contact["emails"]:
-                        contact["emails"].append(email_addr)
-            elif name_part.startswith("TEL"):
-                if ":" in name_part:
-                    phone = _vunesc(name_part.split(":", 1)[1])
-                    if phone and phone not in contact["phones"]:
-                        contact["phones"].append(phone)
-            elif name_part.startswith("ADR"):
-                # vCard ADR is 7 semicolon-separated components:
-                # post-office-box;extended-address;street;locality;region;postal-code;country.
-                # Recover a human-readable string by joining non-empty
-                # components with ", ".
-                if ":" in name_part:
-                    raw = name_part.split(":", 1)[1]
-                    parts = [_vunesc(p).strip() for p in raw.split(";")]
-                    contact["address"] = ", ".join(p for p in parts if p)
-            elif name_part.startswith("UID:"):
-                contact["uid"] = _vunesc(name_part[4:])
-        if contact["name"] or contact["emails"]:
-            contacts.append(contact)
-    return contacts
-
-
-def _vesc(value: str) -> str:
-    """Escape a vCard property VALUE per RFC 6350 §3.4: backslash, comma,
-    semicolon, and newlines. Without this, a name like 'Sekisui House,Ltd'
-    or any value containing a newline produces a malformed vCard (broken
-    N/FN fields) or could inject arbitrary properties."""
-    return (
-        (value or "")
-        .replace("\\", "\\\\")
-        .replace("\n", "\\n")
-        .replace("\r", "")
-        .replace(",", "\\,")
-        .replace(";", "\\;")
-    )
-
-
-def _build_vcard(name: str, email: str, uid: Optional[str] = None,
-                 emails: Optional[List[str]] = None,
-                 phones: Optional[List[str]] = None,
-                 address: Optional[str] = None) -> str:
-    """Build a vCard. Accepts either a single `email` (legacy callers) or
-    full `emails`/`phones` lists (edit path). The first email is marked
-    PREF=1. All values are RFC-6350-escaped."""
-    if not uid:
-        uid = str(uuid.uuid4())
-    # Normalize email lists — `email` arg is a convenience for single-email
-    # creation; `emails` (if given) is authoritative.
-    email_list = [e.strip() for e in (emails if emails is not None else ([email] if email else [])) if e and e.strip()]
-    phone_list = [p.strip() for p in (phones or []) if p and p.strip()]
-    # Try to split name into first/last
-    parts = name.strip().split()
-    if len(parts) >= 2:
-        first = parts[0]
-        last = " ".join(parts[1:])
-    else:
-        first = name
-        last = ""
-    # N field is structured (5 components separated by ';') — escape each
-    # component individually so a comma in the name doesn't split it.
-    n_field = f"{_vesc(last)};{_vesc(first)};;;"
-    lines = [
-        "BEGIN:VCARD",
-        "VERSION:4.0",
-        f"UID:{_vesc(uid)}",
-        f"FN:{_vesc(name)}",
-        f"N:{n_field}",
-    ]
-    for i, em in enumerate(email_list):
-        # First email is the preferred one.
-        lines.append(f"EMAIL;PREF=1:{_vesc(em)}" if i == 0 else f"EMAIL:{_vesc(em)}")
-    for ph in phone_list:
-        lines.append(f"TEL:{_vesc(ph)}")
-    # Address: stuff the whole human-readable string into the street
-    # component of ADR. vCard ADR has 7 semicolon-separated components:
-    # post-office-box;extended-address;street;locality;region;postal-code;country.
-    addr = (address or "").strip()
-    if addr:
-        lines.append(f"ADR:;;{_vesc(addr)};;;;")
-    lines.append("END:VCARD")
-    return "\r\n".join(lines) + "\r\n"
-
-
-# ── In-memory cache ──
-
-_contact_cache = {"contacts": [], "fetched_at": None}
-
-
-def _abs_url(href: str) -> str:
-    """Combine a multistatus <href> (an absolute path like
-    /user/contacts/x.vcf) with the configured CardDAV server origin so we
-    get a fully-qualified URL to PUT/DELETE. Absolute hrefs are accepted only
-    for the configured origin; a cross-origin href is treated as a path on the
-    configured server so a malicious CardDAV response cannot redirect later
-    writes/deletes to cloud metadata or another host."""
-    cfg = _get_carddav_config()
-    base = _carddav_base_url(cfg)
-    base_p = urlparse(base)
-    joined = urljoin(base.rstrip("/") + "/", href or "")
-    joined_p = urlparse(joined)
-    if (joined_p.scheme, joined_p.netloc) != (base_p.scheme, base_p.netloc):
-        joined = urlunparse((base_p.scheme, base_p.netloc, joined_p.path or "/", "", joined_p.query, ""))
-    return _validate_carddav_url(joined)
-
-
-# CardDAV REPORT body — pull every card's etag + raw vCard in ONE request,
-# alongside the resource href. Lets us map each contact's UID to the real
-# server resource path (which is NOT always <uid>.vcf for contacts created
-# by other clients).
-_ADDRESSBOOK_QUERY = (
-    '<?xml version="1.0" encoding="utf-8"?>'
-    '<C:addressbook-query xmlns:D="DAV:" xmlns:C="urn:ietf:params:xml:ns:carddav">'
-    '<D:prop><D:getetag/><C:address-data/></D:prop>'
-    '<C:filter/>'
-    '</C:addressbook-query>'
-)
-
-
-def _fetch_via_report(cfg, auth):
-    """Try a CardDAV REPORT addressbook-query — returns contacts WITH an
-    `href` field, or None if the server doesn't support it / errors."""
-    from defusedxml import ElementTree as ET
-    try:
-        r = httpx.request(
-            "REPORT", cfg["url"],
-            content=_ADDRESSBOOK_QUERY.encode("utf-8"),
-            headers={"Content-Type": "application/xml; charset=utf-8", "Depth": "1"},
-            auth=auth, timeout=10,
-        )
-        if r.status_code not in (207, 200):
-            return None
-        root = ET.fromstring(r.text)
-        ns = {"D": "DAV:", "C": "urn:ietf:params:xml:ns:carddav"}
-        out = []
-        for resp in root.findall("D:response", ns):
-            href_el = resp.find("D:href", ns)
-            data_el = resp.find(".//C:address-data", ns)
-            if href_el is None or data_el is None or not (data_el.text or "").strip():
-                continue
-            parsed = _parse_vcards(data_el.text)
-            if not parsed:
-                continue
-            c = parsed[0]
-            c["href"] = href_el.text.strip()
-            out.append(c)
-        # If the REPORT parsed to ZERO contacts, don't trust it — some
-        # CardDAV servers treat an empty <filter/> as "match nothing" and
-        # return a valid-but-empty 207. Return None so the caller falls
-        # back to the plain GET (which lists everything). A genuinely empty
-        # address book just costs one extra GET that also returns nothing.
-        if not out:
-            return None
-        return out
-    except Exception as e:
-        logger.warning(f"CardDAV REPORT failed, falling back to GET: {e}")
-        return None
-
-
-def _fetch_contacts(force=False):
-    """Fetch all contacts. Uses CardDAV when configured, otherwise local JSON."""
-    if not force and _contact_cache["fetched_at"]:
-        age = (datetime.utcnow() - _contact_cache["fetched_at"]).total_seconds()
-        if age < 60:
-            return _contact_cache["contacts"]
-
-    cfg = _get_carddav_config()
-    if not _carddav_configured(cfg):
-        contacts = _load_local_contacts()
-        _contact_cache["contacts"] = contacts
-        _contact_cache["fetched_at"] = datetime.utcnow()
-        return contacts
-
-    try:
-        cfg["url"] = _carddav_base_url(cfg)
-        auth = None
-        if cfg["username"]:
-            auth = (cfg["username"], cfg["password"])
-        # Preferred path: REPORT gives us hrefs for reliable edit/delete.
-        contacts = _fetch_via_report(cfg, auth)
-        if contacts is None:
-            # Fallback: plain GET, concatenated vCards, no hrefs.
-            r = httpx.get(cfg["url"], auth=auth, timeout=10)
-            if r.status_code != 200:
-                logger.warning(f"CardDAV returned {r.status_code}")
-                return _contact_cache["contacts"]
-            contacts = _parse_vcards(r.text)
-        _contact_cache["contacts"] = contacts
-        _contact_cache["fetched_at"] = datetime.utcnow()
-        return contacts
-    except Exception as e:
-        logger.error(f"Failed to fetch contacts: {e}")
-        return _contact_cache["contacts"]
-
-
-def _resolve_resource_url(uid: str) -> str:
-    """Map a contact UID to its real CardDAV resource URL. Uses the href
-    captured during fetch when available (handles contacts whose filename
-    != UID); falls back to the <uid>.vcf guess for app-created contacts or
-    when no href is known."""
-    def _lookup():
-        for c in _contact_cache.get("contacts", []):
-            if c.get("uid") == uid and c.get("href"):
-                return _abs_url(c["href"])
-        return None
-    found = _lookup()
-    if found:
-        return found
-    # Not in cache (or no href) — refresh once and retry before guessing.
-    try:
-        _fetch_contacts(force=True)
-    except Exception:
-        pass
-    return _lookup() or _vcard_url(uid)
-
-
-def _create_contact(name: str, email: str, address: str = "") -> bool:
-    """Add a new contact via CardDAV or local contacts."""
-    cfg = _get_carddav_config()
-    if not _carddav_configured(cfg):
-        contacts = _load_local_contacts()
-        email_l = (email or "").strip().lower()
-        for c in contacts:
-            if email_l and email_l in [e.lower() for e in c.get("emails", [])]:
-                return True
-        contacts.append(_normalize_contact({"name": name, "emails": [email], "address": address}))
-        _save_local_contacts(contacts)
-        return True
-
-    contact_uid = str(uuid.uuid4())
-    vcard = _build_vcard(name, email, contact_uid, address=address)
-    try:
-        url = _carddav_base_url(cfg) + "/" + contact_uid + ".vcf"
-        auth = None
-        if cfg["username"]:
-            auth = (cfg["username"], cfg["password"])
-        r = httpx.put(
-            url,
-            data=vcard.encode("utf-8"),
-            headers={"Content-Type": "text/vcard; charset=utf-8"},
-            auth=auth,
-            timeout=10,
-        )
-        if r.status_code in (200, 201, 204):
-            # Invalidate cache
-            _contact_cache["fetched_at"] = None
-            return True
-        logger.warning(f"CardDAV PUT returned {r.status_code}: {r.text[:200]}")
-        return False
-    except Exception as e:
-        logger.error(f"Failed to create contact: {e}")
-        return False
-
-
-def _vcard_url(uid: str) -> str:
-    """The CardDAV resource URL for a given contact UID. The uid is URL-
-    encoded so a value containing '/', '..' or other path chars can't
-    escape the collection and target an arbitrary CardDAV resource."""
-    from urllib.parse import quote
-    cfg = _get_carddav_config()
-    return _carddav_base_url(cfg) + "/" + quote(uid, safe="") + ".vcf"
-
-
-def _import_vcards(text: str) -> Dict:
-    """Import a (possibly multi-card) .vcf blob. Each card is PUT to the
-    CardDAV server PRESERVING its full original content (ADR/ORG/photo/
-    etc.) — we don't rebuild it, just ensure it has VERSION + UID and
-    normalize line endings. Returns {imported, failed, total}."""
-    from urllib.parse import quote
-    cfg = _get_carddav_config()
-    if not cfg.get("url"):
-        parsed = _parse_vcards(text)
-        contacts = _load_local_contacts()
-        existing = {
-            e.lower()
-            for c in contacts
-            for e in (c.get("emails") or [])
-            if e
-        }
-        imported = 0
-        for c in parsed:
-            emails = [e for e in (c.get("emails") or []) if e]
-            if emails and any(e.lower() in existing for e in emails):
-                continue
-            contacts.append(_normalize_contact(c))
-            for e in emails:
-                existing.add(e.lower())
-            imported += 1
-        if imported:
-            _save_local_contacts(contacts)
-        return {"imported": imported, "failed": 0, "total": len(parsed)}
-    try:
-        base_url = _carddav_base_url(cfg)
-    except ValueError as e:
-        logger.warning("CardDAV import URL rejected: %s", e)
-        return {"imported": 0, "failed": 0, "total": 0, "error": str(e)}
-    auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
-    # Split into individual cards. re.split drops the BEGIN line, so we
-    # re-add it. Normalize CRLF.
-    raw = (text or "").replace("\r\n", "\n").replace("\r", "\n")
-    blocks = []
-    for chunk in raw.split("BEGIN:VCARD"):
-        chunk = chunk.strip()
-        if not chunk:
-            continue
-        # Trim anything after END:VCARD (defensive).
-        end = chunk.upper().find("END:VCARD")
-        body = chunk[: end + len("END:VCARD")] if end != -1 else chunk
-        blocks.append("BEGIN:VCARD\n" + body)
-    imported = 0
-    failed = 0
-    for block in blocks:
-        # Extract or assign a UID.
-        m = re.search(r"^UID:(.+)$", block, re.MULTILINE)
-        uid = (m.group(1).strip() if m else "") or str(uuid.uuid4())
-        if not m:
-            # Inject a UID right after the VERSION line (or after BEGIN).
-            if re.search(r"^VERSION:", block, re.MULTILINE):
-                block = re.sub(r"(^VERSION:.*$)", r"\1\nUID:" + uid, block, count=1, flags=re.MULTILINE)
-            else:
-                block = block.replace("BEGIN:VCARD", f"BEGIN:VCARD\nVERSION:4.0\nUID:{uid}", 1)
-        elif not re.search(r"^VERSION:", block, re.MULTILINE):
-            block = block.replace("BEGIN:VCARD", "BEGIN:VCARD\nVERSION:4.0", 1)
-        vcard = block.replace("\n", "\r\n") + "\r\n"
-        url = base_url + "/" + quote(uid, safe="") + ".vcf"
-        try:
-            r = httpx.put(
-                url, data=vcard.encode("utf-8"),
-                headers={"Content-Type": "text/vcard; charset=utf-8"},
-                auth=auth, timeout=15,
-            )
-            if r.status_code in (200, 201, 204):
-                imported += 1
-            else:
-                failed += 1
-                logger.warning(f"Import PUT {uid} returned {r.status_code}: {r.text[:120]}")
-        except Exception as e:
-            failed += 1
-            logger.error(f"Import PUT {uid} failed: {e}")
-    if imported:
-        _contact_cache["fetched_at"] = None
-    return {"imported": imported, "failed": failed, "total": len(blocks)}
-
-
-def _import_csv_contacts(text: str) -> Dict:
-    """Import contacts from CSV. Supports common headers:
-    name/full_name/display_name, email/email_address/e-mail, phone/tel.
-    Falls back to first columns as name,email,phone when no headers exist."""
-    raw = (text or "").strip()
-    if not raw:
-        return {"imported": 0, "failed": 0, "total": 0, "error": "No CSV data found"}
-
-    try:
-        sample = raw[:2048]
-        dialect = csv.Sniffer().sniff(sample)
-    except Exception:
-        dialect = csv.excel
-
-    stream = io.StringIO(raw)
-    try:
-        has_header = csv.Sniffer().has_header(raw[:2048])
-    except Exception:
-        has_header = True
-
-    rows = []
-    if has_header:
-        reader = csv.DictReader(stream, dialect=dialect)
-        for row in reader:
-            lowered = {str(k or "").strip().lower(): (v or "").strip() for k, v in row.items()}
-            name = (
-                lowered.get("name") or lowered.get("full name") or lowered.get("full_name")
-                or lowered.get("display name") or lowered.get("display_name")
-                or lowered.get("fn") or ""
-            )
-            email = (
-                lowered.get("email") or lowered.get("email address")
-                or lowered.get("email_address") or lowered.get("e-mail")
-                or lowered.get("mail") or ""
-            )
-            phone = lowered.get("phone") or lowered.get("telephone") or lowered.get("tel") or ""
-            rows.append((name, email, phone))
-    else:
-        stream.seek(0)
-        reader = csv.reader(stream, dialect=dialect)
-        for row in reader:
-            cols = [(c or "").strip() for c in row]
-            if not any(cols):
-                continue
-            rows.append((
-                cols[0] if len(cols) > 0 else "",
-                cols[1] if len(cols) > 1 else "",
-                cols[2] if len(cols) > 2 else "",
-            ))
-
-    imported = 0
-    failed = 0
-    total = 0
-    existing_emails = {
-        e.lower()
-        for c in _fetch_contacts()
-        for e in (c.get("emails") or [])
-        if e
-    }
-    for name, email, phone in rows:
-        email = (email or "").strip()
-        name = (name or "").strip() or (email.split("@")[0] if email else "")
-        if not email:
-            continue
-        total += 1
-        if email.lower() in existing_emails:
-            continue
-        ok = _create_contact(name, email)
-        if ok:
-            imported += 1
-            existing_emails.add(email.lower())
-            # If the CSV had a phone number, rewrite the just-created row
-            # through the richer update path so phone lands in CardDAV too.
-            if phone:
-                try:
-                    contacts = _fetch_contacts(force=True)
-                    created = next((c for c in contacts if email.lower() in [e.lower() for e in c.get("emails", [])]), None)
-                    if created and created.get("uid"):
-                        _update_contact(created["uid"], name, [email], [phone])
-                except Exception:
-                    pass
-        else:
-            failed += 1
-
-    if imported:
-        _contact_cache["fetched_at"] = None
-    return {"imported": imported, "failed": failed, "total": total}
-
-
-def _contacts_to_vcf(contacts: List[Dict]) -> str:
-    return "".join(
-        _build_vcard(
-            c.get("name") or ((c.get("emails") or [""])[0].split("@")[0] if c.get("emails") else "Contact"),
-            "",
-            uid=c.get("uid") or str(uuid.uuid4()),
-            emails=c.get("emails") or [],
-            phones=c.get("phones") or [],
-        )
-        for c in contacts
-    )
-
-
-def _contacts_to_csv(contacts: List[Dict]) -> str:
-    out = io.StringIO()
-    writer = csv.writer(out)
-    writer.writerow(["name", "email", "phone"])
-    for c in contacts:
-        emails = c.get("emails") or [""]
-        phones = c.get("phones") or [""]
-        max_len = max(len(emails), len(phones), 1)
-        for i in range(max_len):
-            writer.writerow([
-                c.get("name") or "",
-                emails[i] if i < len(emails) else "",
-                phones[i] if i < len(phones) else "",
-            ])
-    return out.getvalue()
-
-
-def _update_contact(uid: str, name: str, emails: List[str], phones: List[str], address: str = "") -> bool:
-    """Rewrite an existing contact via CardDAV or local contacts."""
-    cfg = _get_carddav_config()
-    if not _carddav_configured(cfg):
-        contacts = _load_local_contacts()
-        found = False
-        out = []
-        for c in contacts:
-            if c.get("uid") == uid:
-                # Preserve existing address when caller passes "" (only
-                # updating name/emails/phones, not touching address).
-                addr = address if address else c.get("address", "")
-                out.append(_normalize_contact({"uid": uid, "name": name, "emails": emails, "phones": phones, "address": addr}))
-                found = True
-            else:
-                out.append(c)
-        if not found:
-            out.append(_normalize_contact({"uid": uid, "name": name, "emails": emails, "phones": phones, "address": address}))
-        _save_local_contacts(out)
-        return True
-
-    vcard = _build_vcard(name, "", uid=uid, emails=emails, phones=phones, address=address)
-    # Use the real resource href (handles externally-created contacts whose
-    # filename != UID); falls back to the <uid>.vcf guess.
-    try:
-        url = _resolve_resource_url(uid)
-        auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
-        r = httpx.put(
-            url,
-            data=vcard.encode("utf-8"),
-            headers={"Content-Type": "text/vcard; charset=utf-8"},
-            auth=auth,
-            timeout=10,
-        )
-        if r.status_code in (200, 201, 204):
-            _contact_cache["fetched_at"] = None
-            return True
-        logger.warning(f"CardDAV update PUT returned {r.status_code}: {r.text[:200]}")
-        return False
-    except Exception as e:
-        logger.error(f"Failed to update contact: {e}")
-        return False
-
-
-def _delete_contact(uid: str) -> bool:
-    """Delete a contact via CardDAV or local contacts."""
-    cfg = _get_carddav_config()
-    if not _carddav_configured(cfg):
-        contacts = _load_local_contacts()
-        remaining = [c for c in contacts if c.get("uid") != uid]
-        _save_local_contacts(remaining)
-        return True
-
-    try:
-        url = _resolve_resource_url(uid)
-        auth = (cfg["username"], cfg["password"]) if cfg["username"] else None
-        r = httpx.delete(url, auth=auth, timeout=10)
-        if r.status_code in (200, 204, 404):
-            # Invalidate cache so the next fetch sees the server truth.
-            _contact_cache["fetched_at"] = None
-            # Verify: force a fresh fetch and check the UID is actually gone.
-            # A 404 on the guessed URL ({uid}.vcf) can mean the contact
-            # lives at a different resource URL — the DELETE missed it but
-            # we'd silently report success. This check catches that.
-            fresh = _fetch_contacts(force=True)
-            still_there = any(c.get("uid") == uid for c in fresh)
-            if still_there:
-                logger.warning(
-                    f"CardDAV DELETE reported success for {uid} "
-                    f"but UID still present after re-fetch — "
-                    f"resource URL may differ from {redact_url(url)}"
-                )
-                return False
-            if r.status_code == 404:
-                logger.info(f"CardDAV DELETE 404 for {uid} — already gone")
-            return True
-        logger.warning(f"CardDAV DELETE returned {r.status_code}: {r.text[:200]}")
-        return False
-    except Exception as e:
-        logger.error(f"Failed to delete contact: {e}")
-        return False
-
-
-# ── Routes ──
-
-def setup_contacts_routes():
-    router = APIRouter(prefix="/api/contacts", tags=["contacts"])
-
-    @router.get("/list")
-    async def list_contacts(_admin: str = Depends(require_admin)):
-        """List all contacts."""
-        contacts = _fetch_contacts()
-        return {"contacts": contacts, "count": len(contacts)}
-
-    @router.get("/search")
-    async def search_contacts(q: str = Query(""), _admin: str = Depends(require_admin)):
-        """Search contacts by name or email. Returns up to 10 matches."""
-        contacts = _fetch_contacts()
-        if not q:
-            return {"results": []}
-        q_lower = q.lower()
-        results = []
-        for c in contacts:
-            if q_lower in c["name"].lower():
-                results.append(c)
-                continue
-            for em in c["emails"]:
-                if q_lower in em.lower():
-                    results.append(c)
-                    break
-        return {"results": results[:10]}
-
-    @router.post("/add")
-    async def add_contact(data: dict, _admin: str = Depends(require_admin)):
-        """Add a new contact."""
-        name = (data.get("name") or "").strip()
-        email = (data.get("email") or "").strip()
-        phone = (data.get("phone") or "").strip()
-        address = (data.get("address") or "").strip()
-        if not email:
-            return {"success": False, "error": "Email required"}
-        # Check if already exists by email
-        if email:
-            contacts = _fetch_contacts()
-            for c in contacts:
-                if email.lower() in [e.lower() for e in c["emails"]]:
-                    return {"success": True, "message": "Already exists", "contact": c}
-        if not name:
-            name = email.split("@")[0]
-        create_params = inspect.signature(_create_contact).parameters
-        if len(create_params) >= 3:
-            ok = _create_contact(name, email, address)
-        else:
-            ok = _create_contact(name, email)
-        # If a phone was provided, do an immediate update to thread it
-        # through (the simple _create_contact signature only takes name +
-        # email + address; phones happen via update).
-        if ok and phone:
-            try:
-                fresh = _fetch_contacts(force=True)
-                created = next((c for c in fresh if name == c.get("name") and (not email or email in c.get("emails", []))), None)
-                if created:
-                    _update_contact(
-                        created["uid"], name,
-                        created.get("emails", []),
-                        [phone],
-                        address,
-                    )
-            except Exception:
-                pass
-        return {"success": ok}
-
-    @router.post("/import")
-    async def import_vcf(data: dict, _admin: str = Depends(require_admin)):
-        """Import contacts from .vcf or CSV. Body: {"vcf": "..."} or {"csv": "..."}."""
-        # Coerce defensively: a non-string vcf/text/csv (e.g. a number or list
-        # in the JSON body) would otherwise reach .strip() and 500 with an
-        # AttributeError instead of degrading to a clean "no data" response.
-        text = str(data.get("vcf") or data.get("text") or "")
-        csv_text = str(data.get("csv") or "")
-        if text.strip():
-            if "BEGIN:VCARD" not in text.upper():
-                return {"success": False, "error": "No vCard data found"}
-            result = _import_vcards(text)
-        elif csv_text.strip():
-            result = _import_csv_contacts(csv_text)
-        else:
-            return {"success": False, "error": "No contact data found"}
-        result["success"] = result.get("imported", 0) > 0
-        return result
-
-    @router.get("/export")
-    async def export_contacts(
-        format: str = Query("vcf", pattern="^(vcf|csv)$"),
-        _admin: str = Depends(require_admin),
-    ):
-        """Export all contacts as vCard or CSV."""
-        contacts = _fetch_contacts(force=True)
-        if format == "csv":
-            content = _contacts_to_csv(contacts)
-            media_type = "text/csv; charset=utf-8"
-            filename = "odysseus-contacts.csv"
-        else:
-            content = _contacts_to_vcf(contacts)
-            media_type = "text/vcard; charset=utf-8"
-            filename = "odysseus-contacts.vcf"
-        return Response(
-            content=content,
-            media_type=media_type,
-            headers={"Content-Disposition": f'attachment; filename="{filename}"'},
-        )
-
-    @router.get("/config")
-    async def get_config(_admin: str = Depends(require_admin)):
-        cfg = _get_carddav_config()
-        # Mask password
-        if cfg["password"]:
-            cfg["password"] = "***"
-        return cfg
-
-    @router.put("/config")
-    async def update_config(data: dict, _admin: str = Depends(require_admin)):
-        settings = _load_settings()
-        for key in ("carddav_url", "carddav_username", "carddav_password"):
-            if key in data:
-                if key == "carddav_url" and str(data[key] or "").strip():
-                    try:
-                        settings[key] = _validate_carddav_url(data[key])
-                    except ValueError as e:
-                        raise HTTPException(400, str(e))
-                else:
-                    value = data[key]
-                    if key == "carddav_password" and value:
-                        from src.secret_storage import encrypt
-                        value = encrypt(value)
-                    settings[key] = value
-        _save_settings(settings)
-        # Force re-fetch
-        _contact_cache["fetched_at"] = None
-        return {"success": True}
-
-    @router.delete("/clear")
-    async def clear_contacts(_admin: str = Depends(require_admin)):
-        """Clear all local contacts. If CardDAV is configured, only clears the local fallback cache."""
-        _save_local_contacts([])
-        return {"success": True}
-
-    # NOTE: the /{uid} routes are declared LAST so the literal paths above
-    # (/list, /search, /add, /config) win — otherwise PUT /config would
-    # match PUT /{uid} with uid="config".
-    @router.put("/{uid}")
-    async def edit_contact(uid: str, data: dict, _admin: str = Depends(require_admin)):
-        """Edit an existing contact — name / emails / phones / address."""
-        name = (data.get("name") or "").strip()
-        emails = data.get("emails")
-        phones = data.get("phones")
-        if emails is None and data.get("email"):
-            emails = [data["email"]]
-        emails = [e.strip() for e in (emails or []) if e and e.strip()]
-        phones = [p.strip() for p in (phones or []) if p and p.strip()]
-        address = (data.get("address") or "").strip()
-        if not name and not emails and not address:
-            return {"success": False, "error": "Name, email, or address required"}
-        if not name and emails:
-            name = emails[0].split("@")[0]
-        ok = _update_contact(uid, name, emails, phones, address)
-        return {"success": ok}
-
-    @router.delete("/{uid}")
-    async def delete_contact(uid: str, _admin: str = Depends(require_admin)):
-        """Delete a contact by UID."""
-        if not uid:
-            return {"success": False, "error": "UID required"}
-        ok = _delete_contact(uid)
-        return {"success": ok}
-
-    return router
+_sys.modules[__name__] = _canonical

routes/cookbook_helpers.py

@@ -439,15 +439,30 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "                if f.is_file(): nf += 1; sz += f.stat().st_size",
         "                if f.name.endswith('.incomplete'): ic = True",
         "        snap = os.path.join(cache, d, 'snapshots')",
-        "        # Windows HF cache stores files directly in snapshots/; blobs/ may be empty.",
-        "        # Fallback: scan snapshots for real files when blobs yielded nothing.",
-        "        if sz == 0 and os.path.isdir(snap):",
+        "        def snapshot_size():",
+        "            total, count, incomplete = 0, 0, False",
+        "            seen_real = set()",
         "            for sd in os.listdir(snap):",
         "                sf = os.path.join(snap, sd)",
         "                if not os.path.isdir(sf): continue",
-        "                for f in os.scandir(sf):",
-        "                    if f.is_file(): nf += 1; sz += f.stat().st_size",
-        "                    if f.name.endswith('.incomplete'): ic = True",
+        "                for root, dirs, fns in safe_walk(sf):",
+        "                    for fn in fns:",
+        "                        fp = os.path.join(root, fn)",
+        "                        if fn.endswith('.incomplete'): incomplete = True",
+        "                        try:",
+        "                            real = os.path.realpath(fp)",
+        "                            if real in seen_real: continue",
+        "                            seen_real.add(real)",
+        "                            total += os.path.getsize(real)",
+        "                            count += 1",
+        "                        except Exception:",
+        "                            pass",
+        "            return total, count, incomplete",
+        "        # Some HF caches (macOS/MLX/Xet-style) keep blobs elsewhere or expose",
+        "        # snapshot symlinks only. Size snapshots too when blob accounting is empty.",
+        "        if sz == 0 and os.path.isdir(snap):",
+        "            sz2, nf2, ic2 = snapshot_size()",
+        "            sz, nf, ic = sz2, nf2, ic or ic2",
         "        is_diffusion = False; gguf_files = []",
         "        if os.path.isdir(snap):",
         "            for sd in os.listdir(snap):",
@@ -471,7 +486,18 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "    add('/app/.cache/huggingface/hub')",
         f"    add({add_hf_cache!r})" if add_hf_cache else "",
         "    return candidates",
+        "def normalize_model_dir(p):",
+        "    p = os.path.expanduser((p or '').strip())",
+        "    if not p: return p",
+        "    if os.path.isdir(p) or os.path.isabs(p): return p",
+        "    # Users often paste Linux absolute paths without the leading slash.",
+        "    # Treat home/<user>/... as /home/<user>/... so remote scans work.",
+        "    if p.startswith(('home/', 'mnt/', 'media/', 'data/', 'opt/', 'srv/', 'var/')):",
+        "        prefixed = '/' + p",
+        "        if os.path.isdir(prefixed): return prefixed",
+        "    return p",
         "def scan_dir(p):",
+        "    p = normalize_model_dir(p)",
         "    if not os.path.isdir(p) or not safe_path(p): return",
         "    for d in sorted(os.listdir(p)):",
         "        if d.startswith('.'): continue",
@@ -541,7 +567,7 @@ def _cached_model_scan_script(model_dirs: list[str] | None = None, add_hf_cache:
         "scan_ollama()",
     ]
     for model_dir in model_dirs or []:
-        lines.append(f"scan_dir(os.path.expanduser({model_dir!r}))")
+        lines.append(f"scan_dir({model_dir!r})")
     lines.append("print(json.dumps(models))")
     return "\n".join(lines) + "\n"
 
@@ -558,6 +584,18 @@ def _bash_squote(v: str) -> str:
     return v.replace("'", "'\\''")
 
 
+# Shown by generated runner scripts when the ollama binary is missing on the
+# target host. Must stay free of backticks/$( ) and be emitted single-quoted:
+# an earlier version wrapped the install one-liner in backticks inside a
+# double-quoted echo, which bash executed as command substitution and ran the
+# system-wide installer (including on remote SSH hosts) instead of printing
+# the hint.
+OLLAMA_MISSING_HINT = (
+    "ERROR: Ollama not found on this server. Install it from "
+    "https://ollama.com/download or run: curl -fsSL https://ollama.com/install.sh | sh"
+)
+
+
 # Allow-list of binaries permitted as the leading token of `req.cmd` for /api/model/serve.
 # Anything else is rejected before the cmd is interpolated into a tmux/PowerShell wrapper.
 _SERVE_CMD_ALLOWLIST = {
@@ -577,6 +615,16 @@ _SERVE_CMD_ALLOWLIST = {
 _GGUF_PRELUDE_RE = re.compile(
     r'^MODEL_FILE=\$\([^\n]*?\)\s*&&\s*\{[^{}]*\}\s*\|\|\s*\{[^{}]*\}\s*&&\s*'
 )
+_SAFE_SUBSHELL_TEXT = r"[^'\n;&|`$()<>]+"
+_SAFE_SUBSHELL_DQ_HOME_PATH = r'"\$HOME/[^"\n;&|`()<>]*"'
+_SAFE_PRINTF_SUBSHELL_RE = re.compile(
+    rf"^\$\(printf[ \t]+%s[ \t]+(?:'{_SAFE_SUBSHELL_TEXT}'|\$\{{HOME\}}'/{_SAFE_SUBSHELL_TEXT}')\)$"
+)
+_SAFE_FIND_MMPROJ_SUBSHELL_RE = re.compile(
+    rf"^\$\(find[ \t]+(?:'{_SAFE_SUBSHELL_TEXT}'|{_SAFE_SUBSHELL_DQ_HOME_PATH}|{_SAFE_SUBSHELL_TEXT})"
+    r"[ \t]+-iname[ \t]+'mmproj\*\.gguf'"
+    r"(?:[ \t]+2>/dev/null)?[ \t]*\|[ \t]*sort[ \t]*\|[ \t]*head[ \t]+-1\)$"
+)
 _OLLAMA_HOST_ASSIGNMENT_RE = re.compile(r"(?:^|\s)OLLAMA_HOST=([^\s]+)")
 _OLLAMA_BIND_RE = re.compile(r"^\[([^\]]+)\]:(\d+)$|^([^:]+):(\d+)$")
 _OLLAMA_BIND_HOST_RE = re.compile(r"^[A-Za-z0-9._:-]+$")
@@ -677,6 +725,13 @@ def _check_serve_binary(seg: str) -> None:
         )
 
 
+def _is_safe_serve_subshell(subshell: str) -> bool:
+    return bool(
+        _SAFE_PRINTF_SUBSHELL_RE.fullmatch(subshell)
+        or _SAFE_FIND_MMPROJ_SUBSHELL_RE.fullmatch(subshell)
+    )
+
+
 def _validate_serve_cmd(v: str | None) -> str | None:
     """Reject serve commands that aren't in the allowlist or contain shell metachars.
 
@@ -708,15 +763,15 @@ def _validate_serve_cmd(v: str | None) -> str | None:
             _check_serve_binary(part.strip())
         return v
 
-    # Otherwise: a single invocation — no shell metacharacters allowed.
-    # Temporarily replace safe $(printf %s ...) expressions with a placeholder
-    # to avoid triggering the metacharacter/command-injection checks.
-    cleaned_v = v
-    printf_matches = list(re.finditer(r"\$\(\s*printf\s+%s\s+([^\n()]*?)\)", v))
-    for match in printf_matches:
-        inner = match.group(1)
-        if not any(c in inner for c in (";", "&&", "||", "$(", "`")):
-            cleaned_v = cleaned_v.replace(match.group(0), "/placeholder/safe/path.gguf")
+    # Otherwise: a single invocation — no shell metacharacters allowed. Replace
+    # only the exact command substitutions emitted by the Cookbook UI:
+    # $(printf %s 'safe-path') and the mmproj lookup
+    # $(find <path> -iname 'mmproj*.gguf' 2>/dev/null | sort | head -1).
+    def _replace_safe_subshell(match: re.Match[str]) -> str:
+        subshell = match.group(0)
+        return "/placeholder/safe/path" if _is_safe_serve_subshell(subshell) else subshell
+
+    cleaned_v = re.sub(r"\$\([^()]*\)", _replace_safe_subshell, v)
 
     # (`$(` was the original intent; bare `$` is fine for shell-safe paths.)
     if any(c in cleaned_v for c in (";", "&&", "||", "$(")):
@@ -1244,13 +1299,15 @@ def _diagnose_serve_output(text: str) -> dict | None:
             [{"label": "install vLLM in Cookbook Dependencies", "op": "dependency", "package": "vllm"}],
         ),
         (
-            r"sgl_kernel[\s\S]*(Python\.h|libnuma\.so\.1|common_ops)|"
-            r"(Python\.h|libnuma\.so\.1|common_ops)[\s\S]*sgl_kernel|"
+            r"sgl_kernel[\s\S]*(Python\.h|libnuma\.so\.1|common_ops|libnvrtc\.so)|"
+            r"(Python\.h|libnuma\.so\.1|common_ops|libnvrtc\.so)[\s\S]*sgl_kernel|"
+            r"Could not load any common_ops library|"
             r"Please ensure sgl_kernel is properly installed",
-            "SGLang native dependencies are missing on this server.",
+            "SGLang native kernel/runtime is missing or mismatched on this server.",
             [
+                {"label": "repair sglang-kernel in this Python environment", "op": "dependency", "package": "sglang-kernel"},
                 {"label": "install OS packages: libnuma-dev python3.12-dev build-essential", "op": "manual"},
-                {"label": "upgrade sglang-kernel after OS packages are installed", "op": "manual"},
+                {"label": "if libnvrtc is still missing, install the matching CUDA/NVRTC runtime on this host", "op": "manual"},
             ],
         ),
         (
@@ -1258,6 +1315,19 @@ def _diagnose_serve_output(text: str) -> dict | None:
             "SGLang is not installed or not in PATH on this server.",
             [{"label": "install SGLang in Cookbook Dependencies", "op": "dependency", "package": "sglang[all]"}],
         ),
+        (
+            r"No module named ['\"]?mlx_lm|mlx_lm.*command not found|MLX is not installed|MLX LM is not installed",
+            "MLX LM is not installed on this server.",
+            [{"label": "install mlx-lm in Cookbook Dependencies", "op": "dependency", "package": "mlx-lm"}],
+        ),
+        (
+            r"Unable to quantize model of type <class ['\"]mlx_lm\.models\.switch_layers\.QuantizedSwitchLinear['\"]>|QuantizedSwitchLinear",
+            "MLX-LM tried to quantize an already-quantized DeepSeek switch layer.",
+            [
+                {"label": "relaunch from the cached local Hugging Face snapshot path on this Mac", "op": "manual"},
+                {"label": "Odysseus now rewrites MLX repo-id launches to a cached snapshot when one exists", "op": "manual"},
+            ],
+        ),
         # System build deps come BEFORE the generic llama.cpp catch-all so
         # cmake / build-essential / git missing → a specific OS-package
         # remediation instead of "install llama-cpp-python[server]" (which

routes/document_helpers.py

@@ -29,6 +29,7 @@ class DocumentCreate(BaseModel):
 class DocumentUpdate(BaseModel):
     content: str
     summary: Optional[str] = None
+    force_version: bool = False
 
 class DocumentPatch(BaseModel):
     title: Optional[str] = None

routes/document_routes.py

@@ -54,6 +54,18 @@ def _library_language_for_document(doc: Document) -> str:
     return doc.language or "text"
 
 
+def _email_source_key(content: str) -> tuple[str, str]:
+    """Return the source email identity embedded in an email draft document."""
+    import re
+
+    text = content or ""
+    uid_m = re.search(r"(?im)^X-Source-UID:\s*(.+?)\s*$", text)
+    folder_m = re.search(r"(?im)^X-Source-Folder:\s*(.+?)\s*$", text)
+    uid = (uid_m.group(1).strip() if uid_m else "")
+    folder = (folder_m.group(1).strip() if folder_m else "INBOX")
+    return uid, folder
+
+
 from routes.document_helpers import (
     DocumentCreate, DocumentUpdate, DocumentPatch,
     _doc_to_dict, _version_to_dict,
@@ -99,24 +111,62 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                 # the existing lenient path.
                 session = _get_session_or_404(db, req.session_id, user)
 
-            doc_id = str(uuid.uuid4())
-            ver_id = str(uuid.uuid4())
-
             # If no language was supplied (e.g. cloning a doc whose language
             # was never set), detect it from the content rather than storing
             # NULL — which made the editor fall back to plain text. Defaults
             # to markdown for prose.
             language = req.language
             if not language:
-                from src.agent_tools.document_tools import _looks_like_email_document, _sniff_doc_language
+                from src.agent_tools.document_tools import _looks_like_email_document, _sniff_doc_language, _coerce_email_document_content
                 language = _sniff_doc_language(req.content)
             else:
-                from src.agent_tools.document_tools import _looks_like_email_document
+                from src.agent_tools.document_tools import _looks_like_email_document, _coerce_email_document_content
             if _looks_like_email_document(req.content, req.title):
                 language = "email"
 
             _assert_pdf_marker_upload_owned(request, req.content, user, upload_handler)
 
+            # Reply drafts are keyed to the source email. If a UI/tool path tries
+            # to create a second draft for the same email in the same chat,
+            # update the existing draft instead so quoted thread history stays
+            # attached to the visible document.
+            if language == "email" and req.session_id:
+                source_uid, source_folder = _email_source_key(req.content)
+                if source_uid:
+                    candidates = (
+                        db.query(Document)
+                        .filter(Document.session_id == req.session_id)
+                        .filter(Document.is_active == True)
+                        .filter(Document.language == "email")
+                        .order_by(Document.updated_at.desc())
+                        .limit(25)
+                        .all()
+                    )
+                    for existing in candidates:
+                        old_uid, old_folder = _email_source_key(existing.current_content or "")
+                        if old_uid != source_uid or old_folder != source_folder:
+                            continue
+                        merged = _coerce_email_document_content(existing.current_content or "", req.content)
+                        if existing.current_content != merged:
+                            new_ver = (existing.version_count or 1) + 1
+                            existing.current_content = merged
+                            existing.title = req.title or existing.title
+                            existing.version_count = new_ver
+                            db.add(DocumentVersion(
+                                id=str(uuid.uuid4()),
+                                document_id=existing.id,
+                                version_number=new_ver,
+                                content=merged,
+                                summary="Updated existing email draft",
+                                source="user",
+                            ))
+                            db.commit()
+                            db.refresh(existing)
+                        return _doc_to_dict(existing)
+
+            doc_id = str(uuid.uuid4())
+            ver_id = str(uuid.uuid4())
+
             doc = Document(
                 id=doc_id,
                 session_id=req.session_id,
@@ -570,11 +620,23 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                 raise HTTPException(404, "Document not found")
             _verify_doc_owner(db, doc, user)
 
-            # Skip if content is identical
-            if doc.current_content == req.content:
+            incoming_content = req.content
+            from src.agent_tools.document_tools import _coerce_email_document_content, _looks_like_email_document
+            is_email_doc = (
+                (doc.language or "").lower() == "email"
+                or _looks_like_email_document(doc.current_content or "", doc.title or "")
+                or _looks_like_email_document(req.content or "", doc.title or "")
+            )
+            if is_email_doc:
+                incoming_content = _coerce_email_document_content(doc.current_content or "", req.content)
+                doc.language = "email"
+
+            # Skip if content is identical unless the caller explicitly wants
+            # a checkpoint version from the current editor state.
+            if doc.current_content == incoming_content and not req.force_version:
                 return _doc_to_dict(doc)
 
-            _assert_pdf_marker_upload_owned(request, req.content, user, upload_handler)
+            _assert_pdf_marker_upload_owned(request, incoming_content, user, upload_handler)
 
             # Check if we can coalesce with the latest version
             latest_ver = db.query(DocumentVersion).filter(
@@ -583,14 +645,14 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
 
             now = datetime.now(timezone.utc)
             coalesced = False
-            if latest_ver and latest_ver.source == "user":
+            if latest_ver and latest_ver.source == "user" and not req.force_version:
                 ver_time = latest_ver.created_at
                 if ver_time.tzinfo is None:
                     ver_time = ver_time.replace(tzinfo=timezone.utc)
                 age = (now - ver_time).total_seconds()
                 if age < VERSION_COALESCE_SECONDS:
                     # Update the existing version in-place
-                    latest_ver.content = req.content
+                    latest_ver.content = incoming_content
                     latest_ver.created_at = now
                     if req.summary:
                         latest_ver.summary = req.summary
@@ -602,14 +664,14 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                     id=str(uuid.uuid4()),
                     document_id=doc_id,
                     version_number=new_ver,
-                    content=req.content,
+                    content=incoming_content,
                     summary=req.summary or "Manual edit",
                     source="user",
                 )
                 doc.version_count = new_ver
                 db.add(ver)
 
-            doc.current_content = req.content
+            doc.current_content = incoming_content
             db.commit()
             db.refresh(doc)
             return _doc_to_dict(doc)
@@ -799,10 +861,26 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
             from src.document_actions import _JUNK_TITLES
 
             to_delete = []
+            now = datetime.now(timezone.utc)
             for doc in docs:
+                created = doc.created_at
+                if created and created.tzinfo is None:
+                    created = created.replace(tzinfo=timezone.utc)
+
+                # Skip freshly created documents to avoid deleting them while the user is actively editing
+                if created and (now - created).total_seconds() < 900:  # 15 minutes
+                    continue
+
                 content = (doc.current_content or "").strip()
                 title_raw = (doc.title or "").strip()
                 title = title_raw.lower()
+                is_fresh_empty = (
+                    not content
+                    and created is not None
+                    and (now - created).total_seconds() < 1800
+                )
+                if is_fresh_empty:
+                    continue
 
                 # Strip markdown noise to get a "real" character count
                 stripped = _re.sub(r"^#{1,6}\s+", "", content, flags=_re.MULTILINE)
@@ -837,10 +915,6 @@ def setup_document_routes(session_manager, upload_handler=None) -> APIRouter:
                     to_delete.append(doc); deleted += 1; continue
                 if title in _JUNK_TITLES:
                     to_delete.append(doc); deleted += 1; continue
-                if real_len < 30:
-                    to_delete.append(doc); deleted += 1; continue
-                if "\n" not in content and real_len < 50:
-                    to_delete.append(doc); deleted += 1; continue
 
                 # Fix empty or placeholder titles on survivors
                 if not title_raw or title_raw == "Untitled":

routes/email_helpers.py

@@ -349,7 +349,7 @@ def _assert_owns_account(account_id: str, owner: str) -> None:
             row = db.query(_EA).filter(_EA.id == account_id).first()
             if row is None:
                 raise HTTPException(404, "Account not found")
-            if row.owner and row.owner != owner:
+            if not _account_visible_to_owner(row, owner):
                 # Treat as 404 (not 403) so we don't leak existence.
                 raise HTTPException(404, "Account not found")
         finally:
@@ -362,6 +362,26 @@ def _assert_owns_account(account_id: str, owner: str) -> None:
         logger.error(f"Account-owner check failed: {e}")
         raise HTTPException(503, "Account check failed")
 
+
+def _account_visible_to_owner(row, owner: str) -> bool:
+    """Whether an authenticated `owner` may act on this EmailAccount row.
+
+    Mirrors the SQL predicate in `_get_email_config`'s
+    `_owner_or_matching_legacy_account`: a caller sees an account they own, or a
+    legacy owner-less account (owner NULL/"") only when its own mailbox
+    (`imap_user` / `from_address`) is the caller's. `email_accounts` is the one
+    owner-scoped table deliberately left out of the legacy-owner migration
+    backfill, so ownerless rows persist on multi-user deploys — making this the
+    gate that keeps one tenant off another's imported mailbox and its decrypted
+    IMAP/SMTP credentials."""
+    row_owner = getattr(row, "owner", None) or ""
+    if row_owner:
+        return row_owner == owner
+    return owner in {
+        getattr(row, "imap_user", None) or "",
+        getattr(row, "from_address", None) or "",
+    }
+
 def _q(name: str) -> str:
     """Quote an IMAP mailbox name. Defensive: escapes `\\` and `"` and wraps
     in double quotes so user-supplied folder names with spaces or quotes can't
@@ -424,12 +444,19 @@ SCHEDULED_DB = Path(SCHEDULED_EMAILS_DB)
 OWNER_SCOPED_EMAIL_CACHE_TABLES = {
     "email_summaries",
     "email_ai_replies",
+    "email_translations",
     "email_calendar_extractions",
     "email_urgency_alerts",
     "sender_signatures",
 }
 
 
+def email_translation_body_hash(body: str) -> str:
+    import hashlib as _hashlib
+    normalized = (body or "").strip()
+    return _hashlib.sha256(normalized.encode("utf-8", errors="ignore")).hexdigest()
+
+
 def _email_cache_owner_clause(owner: str = "") -> tuple[str, tuple[str, ...]]:
     owner = (owner or "").strip()
     if owner:
@@ -437,14 +464,34 @@ def _email_cache_owner_clause(owner: str = "") -> tuple[str, tuple[str, ...]]:
     return "(owner = '' OR owner IS NULL)", ()
 
 
-def _ensure_owner_scoped_email_cache_table(conn, table: str, create_sql: str, columns: list[str]):
+def _ensure_owner_scoped_email_cache_table(
+    conn,
+    table: str,
+    create_sql: str,
+    columns: list[str],
+    pk_columns: list[str] | None = None,
+):
     """Rebuild legacy Message-ID-only cache tables with owner in the PK."""
+    desired_pk_cols = pk_columns or ["message_id", "owner"]
     conn.execute(create_sql)
     try:
         info = conn.execute(f"PRAGMA table_info({table})").fetchall()
         cols = [r[1] for r in info]
         pk_cols = [r[1] for r in sorted((r for r in info if r[5]), key=lambda r: r[5])]
-        if "owner" in cols and pk_cols == ["message_id", "owner"]:
+        for col in columns:
+            if col not in cols:
+                if col == "owner":
+                    conn.execute(f"ALTER TABLE {table} ADD COLUMN owner TEXT DEFAULT ''")
+                elif col in {"event_uids"}:
+                    conn.execute(f"ALTER TABLE {table} ADD COLUMN {col} TEXT DEFAULT '[]'")
+                elif col.startswith("has_") or col.endswith("_created") or col.endswith("_count"):
+                    conn.execute(f"ALTER TABLE {table} ADD COLUMN {col} INTEGER DEFAULT 0")
+                elif col == "created_at":
+                    conn.execute(f"ALTER TABLE {table} ADD COLUMN {col} TEXT DEFAULT ''")
+                else:
+                    conn.execute(f"ALTER TABLE {table} ADD COLUMN {col} TEXT")
+                cols.append(col)
+        if "owner" in cols and pk_cols == desired_pk_cols:
             return
 
         conn.execute(f"ALTER TABLE {table} RENAME TO {table}__old")
@@ -577,6 +624,25 @@ def _init_scheduled_db():
             PRIMARY KEY (message_id, owner)
         )
     """, ["message_id", "owner", "uid", "folder", "reply", "model_used", "created_at"])
+    _ensure_owner_scoped_email_cache_table(conn, "email_translations", """
+        CREATE TABLE IF NOT EXISTS email_translations (
+            body_hash TEXT,
+            owner TEXT DEFAULT '',
+            target_language TEXT DEFAULT 'English',
+            uid TEXT,
+            folder TEXT,
+            subject TEXT,
+            sender TEXT,
+            translation TEXT,
+            same_language INTEGER DEFAULT 0,
+            model_used TEXT,
+            created_at TEXT NOT NULL,
+            PRIMARY KEY (body_hash, owner, target_language)
+        )
+    """, [
+        "body_hash", "owner", "target_language", "uid", "folder", "subject", "sender",
+        "translation", "same_language", "model_used", "created_at",
+    ], ["body_hash", "owner", "target_language"])
     # Email tags / spam classification cache. SECURITY: keyed by
     # (message_id, owner) because Message-IDs are GLOBAL (a newsletter goes
     # to many users with the same Message-ID). Without owner-scoping, a
@@ -586,6 +652,7 @@ def _init_scheduled_db():
         CREATE TABLE IF NOT EXISTS email_tags (
             message_id TEXT,
             owner TEXT DEFAULT '',
+            account_id TEXT DEFAULT '',
             uid TEXT,
             folder TEXT,
             subject TEXT,
@@ -596,7 +663,7 @@ def _init_scheduled_db():
             moved_to TEXT,
             model_used TEXT,
             created_at TEXT NOT NULL,
-            PRIMARY KEY (message_id, owner)
+            PRIMARY KEY (message_id, owner, account_id)
         )
     """)
     # Backfill migration: older installs created the table with
@@ -604,28 +671,35 @@ def _init_scheduled_db():
     # promote it into the PK by rebuild-copy-swap (SQLite can't ALTER PK).
     try:
         _cols = [r[1] for r in conn.execute("PRAGMA table_info(email_tags)")]
+        _pk_cols = [r[1] for r in sorted(conn.execute("PRAGMA table_info(email_tags)").fetchall(), key=lambda row: row[5] or 99) if r[5]]
         if "owner" not in _cols:
-            # Add the column first so reads/writes don't break mid-migration.
             conn.execute("ALTER TABLE email_tags ADD COLUMN owner TEXT DEFAULT ''")
-            # Rebuild with composite PK. Existing rows get owner='' (legacy
-            # single-user); the urgency scanner will overwrite as it
-            # re-classifies. No data loss.
+            _cols.append("owner")
+        if "account_id" not in _cols:
+            conn.execute("ALTER TABLE email_tags ADD COLUMN account_id TEXT DEFAULT ''")
+            _cols.append("account_id")
+        if _pk_cols != ["message_id", "owner", "account_id"]:
+            # Rebuild with account-aware composite PK. Existing rows get
+            # account_id='' and are still readable as legacy fallback rows;
+            # fresh task runs write exact account ids and no longer block each
+            # other when two accounts share a Message-ID.
             conn.execute("""
                 CREATE TABLE IF NOT EXISTS email_tags__new (
                     message_id TEXT,
                     owner TEXT DEFAULT '',
+                    account_id TEXT DEFAULT '',
                     uid TEXT, folder TEXT, subject TEXT, sender TEXT,
                     tags TEXT, spam_verdict INTEGER DEFAULT 0,
                     spam_reason TEXT, moved_to TEXT, model_used TEXT,
                     created_at TEXT NOT NULL,
-                    PRIMARY KEY (message_id, owner)
+                    PRIMARY KEY (message_id, owner, account_id)
                 )
             """)
             conn.execute("""
                 INSERT OR IGNORE INTO email_tags__new
-                  (message_id, owner, uid, folder, subject, sender, tags,
+                  (message_id, owner, account_id, uid, folder, subject, sender, tags,
                    spam_verdict, spam_reason, moved_to, model_used, created_at)
-                SELECT message_id, COALESCE(owner, ''), uid, folder, subject,
+                SELECT message_id, COALESCE(owner, ''), COALESCE(account_id, ''), uid, folder, subject,
                        sender, tags, spam_verdict, spam_reason, moved_to,
                        model_used, created_at
                 FROM email_tags
@@ -641,11 +715,12 @@ def _init_scheduled_db():
             message_id TEXT,
             owner TEXT DEFAULT '',
             uid TEXT,
+            event_uids TEXT DEFAULT '[]',
             events_created INTEGER DEFAULT 0,
             created_at TEXT NOT NULL,
             PRIMARY KEY (message_id, owner)
         )
-    """, ["message_id", "owner", "uid", "events_created", "created_at"])
+    """, ["message_id", "owner", "uid", "event_uids", "events_created", "created_at"])
     _ensure_owner_scoped_email_cache_table(conn, "email_urgency_alerts", """
         CREATE TABLE IF NOT EXISTS email_urgency_alerts (
             message_id TEXT,
@@ -671,6 +746,64 @@ def _init_scheduled_db():
             PRIMARY KEY (owner, account_key, folder, message_key)
         )
     """)
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS email_message_index (
+            owner TEXT NOT NULL DEFAULT '',
+            account_key TEXT NOT NULL DEFAULT '',
+            folder TEXT NOT NULL,
+            uid TEXT NOT NULL,
+            message_id TEXT,
+            subject TEXT,
+            from_name TEXT,
+            from_address TEXT,
+            to_text TEXT,
+            cc_text TEXT,
+            date_iso TEXT,
+            date_display TEXT,
+            date_epoch REAL DEFAULT 0,
+            size INTEGER DEFAULT 0,
+            flags TEXT DEFAULT '',
+            has_attachments INTEGER DEFAULT 0,
+            updated_at TEXT NOT NULL,
+            PRIMARY KEY (owner, account_key, folder, uid)
+        )
+    """)
+    conn.execute("""
+        CREATE INDEX IF NOT EXISTS ix_email_message_index_folder_date
+        ON email_message_index(owner, account_key, folder, date_epoch DESC)
+    """)
+    conn.execute("""
+        CREATE INDEX IF NOT EXISTS ix_email_message_index_message_id
+        ON email_message_index(owner, account_key, message_id)
+    """)
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS email_body_preview_cache (
+            owner TEXT NOT NULL DEFAULT '',
+            account_key TEXT NOT NULL DEFAULT '',
+            folder TEXT NOT NULL,
+            uid TEXT NOT NULL,
+            message_id TEXT,
+            payload_json TEXT NOT NULL,
+            updated_at TEXT NOT NULL,
+            PRIMARY KEY (owner, account_key, folder, uid)
+        )
+    """)
+    conn.execute("""
+        CREATE INDEX IF NOT EXISTS ix_email_body_preview_message_id
+        ON email_body_preview_cache(owner, account_key, message_id)
+    """)
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS email_attachment_metadata_cache (
+            owner TEXT NOT NULL DEFAULT '',
+            account_key TEXT NOT NULL DEFAULT '',
+            folder TEXT NOT NULL,
+            uid TEXT NOT NULL,
+            message_id TEXT,
+            attachments_json TEXT NOT NULL,
+            updated_at TEXT NOT NULL,
+            PRIMARY KEY (owner, account_key, folder, uid)
+        )
+    """)
     # Boundary cache — LLM-detected sig/quote start positions in the body.
     # Stored as char offsets (-1 = no boundary found). Once cached, the
     # client uses these to fold without ever re-calling the LLM.
@@ -790,12 +923,13 @@ def _get_email_config(account_id: str | None = None, owner: str = "") -> dict:
         try:
             if account_id:
                 row = db.query(_EA).filter(_EA.id == account_id, _EA.enabled == True).first()  # noqa: E712
-                # If the resolved row belongs to a different owner, treat as
+                # If the resolved row isn't visible to this owner, treat as
                 # not-found rather than silently serving it. This is a defense
                 # in depth — `require_owner` already calls `_assert_owns_account`
                 # for query-param account_ids, but other callers (cookbook
-                # rules, scheduled poller) may not.
-                if row is not None and owner and row.owner and row.owner != owner:
+                # rules, scheduled poller) may not. Ownerless legacy rows are
+                # only visible on a mailbox match, same as the fallback below.
+                if row is not None and owner and not _account_visible_to_owner(row, owner):
                     row = None
             # Fallback path — restrict to this owner's accounts so we don't
             # leak another user's default mailbox to an unconfigured user.
@@ -1160,10 +1294,15 @@ def _imap_move(uid, dest, src="INBOX", account_id: str | None = None, owner: str
     try:
         c = _imap_connect(account_id, owner=owner)
         c.select(_q(src))
-        status, _ = c.copy(uid, _q(dest))
+        # Callers pass a real IMAP UID (from conn.uid("SEARCH", ...)). copy()
+        # and store() operate on message SEQUENCE NUMBERS, so addressing them
+        # with a UID moved/deleted the wrong message (or silently no-oped when
+        # the UID exceeded the message count). Use the UID commands, matching
+        # the move/delete path in email_routes.py.
+        status, _ = c.uid("COPY", uid, _q(dest))
         if status != "OK":
             return False
-        c.store(uid, "+FLAGS", "\\Deleted")
+        c.uid("STORE", uid, "+FLAGS", "\\Deleted")
         c.expunge()
         return True
     except Exception as e:
@@ -1276,12 +1415,14 @@ def _list_attachments_from_msg(msg):
             except Exception:
                 payload = b""
         size = len(payload) if payload is not None else 0
+        content_id = (part.get("Content-ID") or "").strip().strip("<>")
         attachments.append({
             "index": idx,
             "filename": filename,
             "content_type": ct,
             "size": size,
             "is_inline": "inline" in cd.lower(),
+            "content_id": content_id,
         })
         idx += 1
     return attachments
@@ -1720,6 +1861,10 @@ class SendEmailRequest(BaseModel):
     attachments: Optional[List[str]] = None
     # Which account to send from. None = default account.
     account_id: Optional[str] = None
+    # Source message for replies. When present, /send marks this exact message
+    # answered after successful delivery so it leaves undone/reply-soon views.
+    source_uid: Optional[str] = None
+    source_folder: Optional[str] = None
     # Internal marker for Odysseus-generated mail (e.g. reminder, scheduled).
     odysseus_kind: Optional[str] = None
     # If true, /send waits for SMTP + Sent append and returns the sent UID.

routes/email_pollers.py

@@ -29,7 +29,7 @@ from datetime import datetime
 from email.mime.text import MIMEText
 from email.mime.multipart import MIMEMultipart
 
-from src.llm_core import llm_call_async
+from src.task_endpoint import resolve_task_candidates, task_llm_call_async
 
 from routes.email_helpers import (
     _strip_think, _extract_reply, _apply_email_style_mechanics, _load_settings, _save_settings, _get_email_config,
@@ -56,6 +56,35 @@ _CAL_ACTION_ARRAY_RE = re.compile(
 )
 
 
+def _extract_json_array_from_text(text: str):
+    """Return the last valid JSON array embedded in model output, if any."""
+    if not text:
+        return None
+    cleaned = re.sub(r"^```(?:json)?\s*|\s*```$", "", text.strip(), flags=re.MULTILINE).strip()
+    decoder = json.JSONDecoder()
+    try:
+        parsed = decoder.decode(cleaned)
+        if isinstance(parsed, list):
+            return parsed
+    except Exception:
+        pass
+
+    # Models often explain themselves and finish with `[]` or `[{"action":...}]`.
+    # Scan every array opener and keep the last complete JSON array, rather than
+    # using a greedy regex that can swallow prose containing square brackets.
+    last = None
+    for idx, ch in enumerate(cleaned):
+        if ch != "[":
+            continue
+        try:
+            parsed, _end = decoder.raw_decode(cleaned[idx:])
+        except Exception:
+            continue
+        if isinstance(parsed, list):
+            last = parsed
+    return last
+
+
 def _owner_for_email_account(account_id: str | None) -> str:
     if not account_id:
         return ""
@@ -88,6 +117,8 @@ async def _run_auto_summarize_once(do_summary: bool = True, do_reply: bool = Tru
                                    do_tag: bool = False, do_spam: bool = False,
                                    do_calendar: bool = False,
                                    days_back: int = 1,
+                                   account_id: str | None = None,
+                                   max_process: int | None = None,
                                    progress_cb=None) -> str:
     """One iteration of the email scan. Temporarily flips settings flags
     so the existing background-loop logic runs exactly once for the requested ops."""
@@ -102,7 +133,12 @@ async def _run_auto_summarize_once(do_summary: bool = True, do_reply: bool = Tru
     settings["email_auto_calendar"] = bool(do_calendar)
     _save_settings(settings)
     try:
-        return await _auto_summarize_pass(days_back=days_back, progress_cb=progress_cb)
+        return await _auto_summarize_pass(
+            days_back=days_back,
+            account_id=account_id,
+            max_process=max_process,
+            progress_cb=progress_cb,
+        )
     finally:
         s2 = _load_settings()
         for k, v in prev.items():
@@ -140,7 +176,7 @@ def _latest_inbox_fallback_uids(conn, reconnect):
         return [], reconnect()
 
 
-async def _auto_summarize_pass(days_back: int = 1, account_id: str | None = None, progress_cb=None) -> str:
+async def _auto_summarize_pass(days_back: int = 1, account_id: str | None = None, max_process: int | None = None, progress_cb=None) -> str:
     """Single pass of the auto-summarize/reply scan.
 
     When account_id is None, iterates over every enabled account in
@@ -167,28 +203,41 @@ async def _auto_summarize_pass(days_back: int = 1, account_id: str | None = None
             names = {}
         if len(ids) <= 1:
             # Single-account (or zero rows — fallback to legacy settings.json lookup)
-            return await _auto_summarize_pass_single(days_back=days_back, account_id=(ids[0] if ids else None), progress_cb=progress_cb)
+            return await _auto_summarize_pass_single(
+                days_back=days_back,
+                account_id=(ids[0] if ids else None),
+                max_process=max_process,
+                progress_cb=progress_cb,
+            )
         outs = []
         for idx, aid in enumerate(ids, start=1):
             try:
                 await _emit_progress(progress_cb, f"{names.get(aid, aid[:8])}: starting ({idx}/{len(ids)})")
-                result = await _auto_summarize_pass_single(days_back=days_back, account_id=aid, progress_cb=progress_cb)
+                result = await _auto_summarize_pass_single(
+                    days_back=days_back,
+                    account_id=aid,
+                    max_process=max_process,
+                    progress_cb=progress_cb,
+                )
                 outs.append(f"[{names.get(aid, aid[:8])}] {result}")
             except Exception as e:
                 logger.warning(f"auto-summarize pass failed for account {aid}: {e}")
                 outs.append(f"[{names.get(aid, aid[:8])}] error: {e}")
         return "\n".join(outs)
-    return await _auto_summarize_pass_single(days_back=days_back, account_id=account_id, progress_cb=progress_cb)
+    return await _auto_summarize_pass_single(
+        days_back=days_back,
+        account_id=account_id,
+        max_process=max_process,
+        progress_cb=progress_cb,
+    )
 
 
-async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None = None, progress_cb=None) -> str:
+async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None = None, max_process: int | None = None, progress_cb=None) -> str:
     """Single pass of the auto-summarize/reply scan for ONE account.
     Reads current settings flags."""
     import asyncio
     import sqlite3 as _sql3
-    import requests as _req
-    from src.endpoint_resolver import resolve_endpoint
-    from src.llm_core import _uses_max_completion_tokens, _restricts_temperature
+    from src.llm_core import _uses_max_completion_tokens
 
     settings = _load_settings()
     auto_sum = settings.get("email_auto_summarize", False)
@@ -265,9 +314,15 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
         ).fetchall()}
         if auto_tag or auto_spam:
             if account_owner:
-                _tag_existing = {r[0] for r in _c.execute("SELECT message_id FROM email_tags WHERE owner=?", (account_owner,)).fetchall()}
+                _tag_existing = {r[0] for r in _c.execute(
+                    "SELECT message_id FROM email_tags WHERE owner=? AND (account_id=? OR account_id='' OR account_id IS NULL)",
+                    (account_owner, account_id or ""),
+                ).fetchall()}
             else:
-                _tag_existing = {r[0] for r in _c.execute("SELECT message_id FROM email_tags WHERE owner='' OR owner IS NULL").fetchall()}
+                _tag_existing = {r[0] for r in _c.execute(
+                    "SELECT message_id FROM email_tags WHERE (owner='' OR owner IS NULL) AND (account_id=? OR account_id='' OR account_id IS NULL)",
+                    (account_id or "",),
+                ).fetchall()}
         else:
             _tag_existing = set()
         _cal_existing = {r[0] for r in _c.execute(
@@ -296,11 +351,10 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
         if auto_spam and not spam_folder:
             logger.warning("Auto-spam enabled but no Junk/Spam folder detected — will classify but not move")
 
-        url, model, headers = resolve_endpoint("utility", owner=account_owner)
-        if not url:
-            url, model, headers = resolve_endpoint("default", owner=account_owner)
-        if not url or not model:
+        task_candidates = resolve_task_candidates(owner=account_owner)
+        if not task_candidates:
             return "No model configured"
+        url, model, headers = task_candidates[0]
 
         writing_style = settings.get("email_writing_style", "")
         processed = 0
@@ -314,7 +368,14 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
         _reply_failed = 0
         _detail_lines = []
         _current_folder = "INBOX"
-        _max_process = 5
+        # Calendar extraction is sequential and each row can involve a model
+        # call plus a calendar write. Keep the scheduled calendar-only pass
+        # below the 5-minute action budget instead of timing out mid-run.
+        _default_max_process = 3 if (auto_cal and not auto_sum and not auto_reply and not auto_tag and not auto_spam) else 5
+        try:
+            _max_process = max(1, int(max_process)) if max_process is not None else _default_max_process
+        except Exception:
+            _max_process = _default_max_process
         for _entry in uid_list:
             if processed >= _max_process:
                 break
@@ -406,48 +467,30 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                     req_headers.update(headers)
 
                 if need_sum:
-                    tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
-                    payload = {
-                        "model": model,
-                        "messages": [
-                            {"role": "system", "content": "You are an email summarizer. Format: 1-3 short bullet points (use '- '). Cover: main point, action items, deadlines. If the email has attachments (marked '--- ATTACHMENTS ---'), USE THEIR CONTENTS — pull out invoice totals, deadlines, key clauses, any concrete numbers/dates in PDFs/docs, and reflect them in the bullets. Be terse.\n\nOUTPUT FORMAT: Put ONLY the bullet points between these exact markers, each on its own line:\n<<<SUMMARY>>>\n- ...\n<<<END>>>\nAny reasoning or planning must come BEFORE <<<SUMMARY>>> (ideally inside <think>...</think>). Only the text between the markers is kept."},
-                            {"role": "user", "content": f"From: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}\n\n---\n\nSummarize the email. Output the bullets between <<<SUMMARY>>> and <<<END>>>."},
-                        ],
-                        tok_key: 16384,
-                        "temperature": 0.3,
-                        "stream": False,
-                    }
-                    # Reasoning models (o1/o3/o4/gpt-5) reject an explicit temperature.
-                    if _restricts_temperature(model):
-                        payload.pop("temperature", None)
                     try:
-                        # Use to_thread so this sync HTTP call doesn't freeze
-                        # the entire event loop while the LLM thinks (240s).
-                        resp = await asyncio.to_thread(
-                            _req.post, url, json=payload, headers=req_headers, timeout=240
+                        summary = await task_llm_call_async(
+                            messages=[
+                                {"role": "system", "content": "You are an email summarizer. Format: 1-3 short bullet points (use '- '). Cover: main point, action items, deadlines. If the email has attachments (marked '--- ATTACHMENTS ---'), USE THEIR CONTENTS — pull out invoice totals, deadlines, key clauses, any concrete numbers/dates in PDFs/docs, and reflect them in the bullets. Be terse.\n\nOUTPUT FORMAT: Put ONLY the bullet points between these exact markers, each on its own line:\n<<<SUMMARY>>>\n- ...\n<<<END>>>\nAny reasoning or planning must come BEFORE <<<SUMMARY>>> (ideally inside <think>...</think>). Only the text between the markers is kept."},
+                                {"role": "user", "content": f"From: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}\n\n---\n\nSummarize the email. Output the bullets between <<<SUMMARY>>> and <<<END>>>."},
+                            ],
+                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                            owner=account_owner or None,
+                            temperature=0.3, max_tokens=16384, timeout=240,
                         )
-                        if resp.ok:
-                            rdata = resp.json()
-                            m = (rdata.get("choices") or [{}])[0].get("message", {})
-                            summary = (m.get("content") or "").strip()
-                            summary = _extract_reply(summary)
-                            if not summary:
-                                rc = (m.get("reasoning_content") or "").strip()
-                                bullets = [ln.strip() for ln in rc.split("\n") if re.match(r"^[-•*]\s+|^\d+[.)]\s+", ln.strip())]
-                                summary = "\n".join(bullets) if bullets else ""
-                            if summary:
-                                _c = _sql3.connect(SCHEDULED_DB)
-                                _c.execute("""
-                                    INSERT OR REPLACE INTO email_summaries
-                                    (message_id, owner, uid, folder, subject, sender, summary, model_used, created_at)
-                                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
-                                """, (message_id, account_owner or "", uid.decode() if isinstance(uid, bytes) else str(uid), _folder, subject, sender, summary, model, datetime.utcnow().isoformat()))
-                                _c.commit()
-                                _c.close()
-                                _sum_existing.add(message_id)
-                                _summaries_created += 1
-                                _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
-                                _detail_lines.append(f"summary · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
+                        summary = _extract_reply((summary or "").strip())
+                        if summary:
+                            _c = _sql3.connect(SCHEDULED_DB)
+                            _c.execute("""
+                                INSERT OR REPLACE INTO email_summaries
+                                (message_id, owner, uid, folder, subject, sender, summary, model_used, created_at)
+                                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                            """, (message_id, account_owner or "", uid.decode() if isinstance(uid, bytes) else str(uid), _folder, subject, sender, summary, model, datetime.utcnow().isoformat()))
+                            _c.commit()
+                            _c.close()
+                            _sum_existing.add(message_id)
+                            _summaries_created += 1
+                            _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
+                            _detail_lines.append(f"summary · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
                     except Exception as e:
                         _uid_text = uid.decode() if isinstance(uid, bytes) else str(uid)
                         _detail_lines.append(f"summary failed · {_folder}#{_uid_text} · {subject or '(no subject)'} — {sender or '(unknown sender)'}")
@@ -468,14 +511,14 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                     if context_snippets:
                         sys_prompt += "\n\nRELEVANT CONTEXT FROM PAST EMAILS AND CONTACTS:\n" + "\n\n---\n\n".join(context_snippets[:5])
                     try:
-                        reply = await llm_call_async(
-                            url=url, model=model,
+                        reply = await task_llm_call_async(
                             messages=[
                                 {"role": "system", "content": sys_prompt},
                                 {"role": "user", "content": f"Original email:\nFrom: {sender}\nSubject: {subject}\n\n{body_for_llm[:12000]}\n\nDraft a reply. Return only the reply body text."},
                             ],
-                            temperature=0.7, max_tokens=1024,
-                            headers=req_headers, timeout=90,
+                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                            owner=account_owner or None,
+                            temperature=0.7, max_tokens=1024, timeout=90,
                         )
                         reply = _apply_email_style_mechanics(_extract_reply(reply or ""))
                         if reply:
@@ -502,6 +545,8 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                 # ── Calendar event extraction (independent of reply drafting) ──
                 if need_cal:
                     _cal_run_count = 0
+                    _cal_event_uids = []
+                    _cal_parse_ok = False
                     try:
                         # Pull a snapshot of upcoming events so the LLM can decide
                         # create vs update vs cancel based on what already exists.
@@ -510,8 +555,7 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                         _existing_summary = get_upcoming_events(_acct_owner, horizon_days=60, limit=40)
                         existing_json = json.dumps(_existing_summary)
                         is_sent = _folder.lower().startswith("sent") or "sent" in _folder.lower()
-                        cal_extract = await llm_call_async(
-                            url=url, model=model,
+                        cal_extract = await task_llm_call_async(
                             messages=[
                                 {"role": "system", "content": (
                                     "You are a calendar assistant. The user receives emails AND sends replies "
@@ -562,8 +606,9 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                                     f"{body[:4000]}"
                                 )},
                             ],
-                            temperature=0.1, max_tokens=16384,
-                            headers=req_headers, timeout=180,
+                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                            owner=account_owner or None,
+                            temperature=0.1, max_tokens=16384, timeout=75,
                         )
                         _raw_original = cal_extract or ""
                         cal_extract = _strip_think(_raw_original)
@@ -573,10 +618,10 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                             if matches:
                                 cal_extract = matches[-1].group()
                         logger.info(f"[cal-extract] uid={uid.decode() if isinstance(uid, bytes) else uid} folder={_folder} subj={subject[:50]!r} raw_len={len(cal_extract)} orig_len={len(_raw_original)} raw={cal_extract[:800]!r}")
-                        jm = re.search(r'\[.*\]', cal_extract, re.DOTALL)
-                        if jm:
+                        ops = _extract_json_array_from_text(cal_extract)
+                        if ops is not None:
                             try:
-                                ops = json.loads(jm.group())
+                                _cal_parse_ok = True
                                 logger.info(f"[cal-extract] parsed {len(ops)} op(s)")
                                 if isinstance(ops, list) and ops:
                                     from src.tool_implementations import do_manage_calendar
@@ -606,6 +651,8 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                                             r = await do_manage_calendar(json.dumps(args), owner=_acct_owner)
                                             if r.get("exit_code", 0) == 0:
                                                 logger.info(f"[cal-extract] Updated event uid={cuid} → {op.get('title')} {op['date']}")
+                                                if cuid and cuid not in _cal_event_uids:
+                                                    _cal_event_uids.append(cuid)
                                                 _cal_run_count += 1
                                             else:
                                                 logger.warning(f"[cal-extract] update failed: {r.get('error')}")
@@ -686,29 +733,41 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                                             r = await do_manage_calendar(cal_args, owner=_acct_owner)
                                             if r.get("exit_code", 0) == 0:
                                                 logger.info(f"[cal-extract] Created event: {op['title']} on {op['date']}")
+                                                _created_uid = (r.get("uid") or "").strip()
+                                                if _created_uid and _created_uid not in _cal_event_uids:
+                                                    _cal_event_uids.append(_created_uid)
                                                 _events_created += 1
                                                 _cal_run_count += 1
                                             else:
                                                 logger.warning(f"[cal-extract] create failed: {r.get('error')} args={cal_args[:200]}")
                             except Exception as je:
                                 logger.warning(f"[cal-extract] JSON parse failed: {je} on raw={cal_extract[:200]!r}")
+                        else:
+                            logger.warning(f"[cal-extract] no JSON array found on raw={cal_extract[:200]!r}")
                     except Exception as e:
                         logger.warning(f"[cal-extract] Meeting extraction LLM call failed for uid={uid}: {e}")
                     else:
-                        # Record we processed this email so we don't re-LLM next run.
-                        # Only mark as processed on success ? transient LLM failures
-                        # are retried on the next poll run (matches summary/reply pattern).
+                        # Record successfully parsed results so we don't re-LLM
+                        # no-op emails. Transient LLM failures are retried on
+                        # the next poll run.
                         try:
-                            _cc = _sql3.connect(SCHEDULED_DB)
-                            _cc.execute(
-                                "INSERT OR REPLACE INTO email_calendar_extractions "
-                                "(message_id, owner, uid, events_created, created_at) VALUES (?, ?, ?, ?, ?)",
-                                (message_id, account_owner or "", uid.decode() if isinstance(uid, bytes) else str(uid),
-                                 _cal_run_count, datetime.utcnow().isoformat())
-                            )
-                            _cc.commit()
-                            _cc.close()
-                            _cal_existing.add(message_id)
+                            if _cal_parse_ok:
+                                _cc = _sql3.connect(SCHEDULED_DB)
+                                _cc.execute(
+                                    "INSERT OR REPLACE INTO email_calendar_extractions "
+                                    "(message_id, owner, uid, event_uids, events_created, created_at) VALUES (?, ?, ?, ?, ?, ?)",
+                                    (
+                                        message_id,
+                                        account_owner or "",
+                                        uid.decode() if isinstance(uid, bytes) else str(uid),
+                                        json.dumps(_cal_event_uids),
+                                        _cal_run_count,
+                                        datetime.utcnow().isoformat(),
+                                    ),
+                                )
+                                _cc.commit()
+                                _cc.close()
+                                _cal_existing.add(message_id)
                         except Exception as ce:
                             logger.debug(f"Could not cache calendar extraction: {ce}")
 
@@ -742,9 +801,11 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                             "temperature": 0,
                             tok_key: 200,
                         }
-                        urg_raw = await llm_call_async(
-                            url=url, model=model, messages=payload["messages"],
-                            temperature=0, max_tokens=200, headers=req_headers, timeout=60,
+                        urg_raw = await task_llm_call_async(
+                            messages=payload["messages"],
+                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                            owner=account_owner or None,
+                            temperature=0, max_tokens=200, timeout=60,
                         )
                         urg_raw = _strip_think(urg_raw or "")
                         urg_raw = re.sub(r"^```(?:json)?\s*|\s*```$", "", urg_raw, flags=re.MULTILINE).strip()
@@ -845,8 +906,13 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                         class_sys = (
                             "Classify the email. Return ONLY a JSON object, no prose, no markdown fences. "
                             "Schema: {\"tags\": [\"tag1\"], \"spam\": false, \"reason\": \"short\"}. "
-                            "Pick 1-2 tags from: work, personal, finance, bills, receipt, travel, "
-                            "newsletter, promo, notification, security, social, shopping, calendar.\n\n"
+                            "Pick 1-3 tags from: work, personal, urgent, action-needed, finance, bills, "
+                            "receipt, legal, travel, newsletter, promo, notification, security, social, "
+                            "shopping, calendar, support.\n\n"
+                            "Use work for professional/company/client/operations messages. "
+                            "Use personal for friends/family/private-life messages. "
+                            "Use urgent for real time-sensitive consequences. "
+                            "Use action-needed when the user likely needs to reply, pay, sign, book, or decide.\n\n"
                             "Set spam=true for ANY of:\n"
                             "- Phishing, scams, chain mail, deceptive offers\n"
                             "- Marketing/promotional blasts (\"special offer\", \"limited time\", discount codes)\n"
@@ -863,70 +929,55 @@ async def _auto_summarize_pass_single(days_back: int = 1, account_id: str | None
                             "If it's a mass-mailed generic update with no personal CTA, mark spam=true even if from a legitimate service. "
                             "Reason should be 5-10 words."
                         )
-                        tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
-                        payload = {
-                            "model": model,
-                            "messages": [
+                        raw_out = await task_llm_call_async(
+                            messages=[
                                 {"role": "system", "content": class_sys},
                                 {"role": "user", "content": f"From: {sender}\nSubject: {subject}\n\n{body[:4000]}"},
                             ],
-                            tok_key: 512,
-                            "temperature": 0.1,
-                            "stream": False,
-                        }
-                        # Reasoning models (o1/o3/o4/gpt-5) reject an explicit temperature.
-                        if _restricts_temperature(model):
-                            payload.pop("temperature", None)
-                        # to_thread keeps the event loop responsive during the LLM call
-                        resp = await asyncio.to_thread(
-                            _req.post, url, json=payload, headers=req_headers, timeout=120
+                            fallback_url=url, fallback_model=model, fallback_headers=headers,
+                            owner=account_owner or None,
+                            temperature=0.1, max_tokens=512, timeout=120,
                         )
-                        if not resp.ok:
-                            logger.warning(f"Auto-classify {uid.decode() if isinstance(uid, bytes) else str(uid)} HTTP {resp.status_code}: {resp.text[:200]}")
-                        else:
-                            rdata = resp.json()
-                            m = (rdata.get("choices") or [{}])[0].get("message", {})
-                            raw_out = (m.get("content") or "").strip()
-                            raw_out = _strip_think(raw_out)
-                            raw_out = re.sub(r"^```(?:json)?\s*|\s*```$", "", raw_out, flags=re.MULTILINE).strip()
-                            jm = re.search(r'\{.*\}', raw_out, re.DOTALL)
-                            parsed = None
-                            if jm:
-                                try:
-                                    parsed = json.loads(jm.group(0))
-                                except Exception:
-                                    parsed = None
-                            if parsed is not None:
-                                _ALLOWED_TAGS = {"work","personal","finance","bills","receipt","travel",
-                                                 "newsletter","marketing","notification","security","social",
-                                                 "shopping","calendar"}
-                                raw_tags = parsed.get("tags") or []
-                                if isinstance(raw_tags, str):
-                                    raw_tags = [raw_tags]
-                                tags = [t.strip().lower().replace("_", "-") for t in raw_tags if isinstance(t, str)]
-                                tags = ["marketing" if t == "promo" else t for t in tags]
-                                tags = [t for t in tags if t in _ALLOWED_TAGS][:2]
-                                is_spam = bool(parsed.get("spam"))
-                                spam_reason = str(parsed.get("reason") or "")[:200]
+                        raw_out = _strip_think((raw_out or "").strip())
+                        raw_out = re.sub(r"^```(?:json)?\s*|\s*```$", "", raw_out, flags=re.MULTILINE).strip()
+                        jm = re.search(r'\{.*\}', raw_out, re.DOTALL)
+                        parsed = None
+                        if jm:
+                            try:
+                                parsed = json.loads(jm.group(0))
+                            except Exception:
+                                parsed = None
+                        if parsed is not None:
+                            _ALLOWED_TAGS = {"work","personal","urgent","action-needed","finance","bills",
+                                             "receipt","legal","travel","newsletter","marketing","notification",
+                                             "security","social","shopping","calendar","support"}
+                            raw_tags = parsed.get("tags") or []
+                            if isinstance(raw_tags, str):
+                                raw_tags = [raw_tags]
+                            tags = [t.strip().lower().replace("_", "-") for t in raw_tags if isinstance(t, str)]
+                            tags = ["marketing" if t == "promo" else t for t in tags]
+                            tags = [t for t in tags if t in _ALLOWED_TAGS][:3]
+                            is_spam = bool(parsed.get("spam"))
+                            spam_reason = str(parsed.get("reason") or "")[:200]
 
-                                moved_to = ""
-                                if is_spam and auto_spam and spam_folder:
-                                    if _imap_move(uid, spam_folder, account_id=account_id, owner=account_owner):
-                                        moved_to = spam_folder
-                                        logger.info(f"Auto-spam moved uid={uid.decode() if isinstance(uid, bytes) else str(uid)} to {spam_folder}: {spam_reason}")
+                            moved_to = ""
+                            if is_spam and auto_spam and spam_folder:
+                                if _imap_move(uid, spam_folder, account_id=account_id, owner=account_owner):
+                                    moved_to = spam_folder
+                                    logger.info(f"Auto-spam moved uid={uid.decode() if isinstance(uid, bytes) else str(uid)} to {spam_folder}: {spam_reason}")
 
-                                _c = _sql3.connect(SCHEDULED_DB)
-                                _c.execute("""
-                                    INSERT OR REPLACE INTO email_tags
-                                    (message_id, owner, uid, folder, subject, sender, tags, spam_verdict,
-                                     spam_reason, moved_to, model_used, created_at)
-                                    VALUES (?, ?, ?, 'INBOX', ?, ?, ?, ?, ?, ?, ?, ?)
-                                """, (message_id, account_owner or "", uid.decode() if isinstance(uid, bytes) else str(uid), subject, sender,
-                                      json.dumps(tags), 1 if is_spam else 0,
-                                      spam_reason, moved_to, model, datetime.utcnow().isoformat()))
-                                _c.commit()
-                                _c.close()
-                                _tag_existing.add(message_id)
+                            _c = _sql3.connect(SCHEDULED_DB)
+                            _c.execute("""
+                                INSERT OR REPLACE INTO email_tags
+                                (message_id, owner, account_id, uid, folder, subject, sender, tags, spam_verdict,
+                                 spam_reason, moved_to, model_used, created_at)
+                                VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                            """, (message_id, account_owner or "", account_id or "", uid.decode() if isinstance(uid, bytes) else str(uid), _folder, subject, sender,
+                                  json.dumps(tags), 1 if is_spam else 0,
+                                  spam_reason, moved_to, model, datetime.utcnow().isoformat()))
+                            _c.commit()
+                            _c.close()
+                            _tag_existing.add(message_id)
                     except Exception as e:
                         logger.warning(f"Auto-classify {uid} failed: {e}")
 

routes/gallery/gallery_helpers.py

@@ -99,6 +99,7 @@ def _image_to_dict(img: GalleryImage, session_name: str = None) -> Dict[str, Any
         "filename": img.filename,
         "url": f"/api/generated-image/{img.filename}",
         "prompt": img.prompt,
+        "caption": img.caption or "",
         "model": img.model,
         "size": img.size,
         "quality": img.quality,

routes/gallery/gallery_routes.py

@@ -77,6 +77,39 @@ def _normalize_image_endpoint_base(url: str) -> str:
     return base
 
 
+def _is_openai_api_base(url: str) -> bool:
+    """Return True only when url's hostname is exactly api.openai.com."""
+    from urllib.parse import urlsplit
+    try:
+        candidate = url if "://" in url else f"https://{url}"
+        return urlsplit(candidate).hostname == "api.openai.com"
+    except Exception:
+        return False
+
+
+_GALLERY_ENDPOINT_PATHS = frozenset({
+    "/images/edits",
+    "/images/generations",
+    "/images/harmonize",
+    "/images/img2img",
+    "/images/inpaint",
+    "/images/upscale",
+    "/images/variations",
+    "/sdapi/v1/img2img",
+})
+
+
+def _join_checked_gallery_endpoint(base: str, path: str) -> str:
+    """Append a known-constant gallery path suffix to a validated base URL.
+
+    Rejects paths not in the pre-approved list so arbitrary strings can never
+    be spliced into the URL passed to httpx.
+    """
+    if path not in _GALLERY_ENDPOINT_PATHS:
+        raise ValueError(f"Unexpected gallery path: {path!r}")
+    return base + path
+
+
 def _visible_image_endpoint_query(db, owner: str | None):
     from src.auth_helpers import owner_filter
     q = db.query(ModelEndpoint).filter(
@@ -255,9 +288,10 @@ def setup_gallery_routes() -> APIRouter:
                 pass
             try:
                 db.commit()
-            except Exception as e:
+            except Exception:
                 db.rollback()
-                raise HTTPException(500, f"DB commit failed: {e}")
+                logger.exception("gallery_replace: DB commit failed")
+                raise HTTPException(500, "Image update failed")
             return {"ok": True, "width": img.width, "height": img.height}
         finally:
             db.close()
@@ -385,8 +419,9 @@ def setup_gallery_routes() -> APIRouter:
                     return {"image": data.get("data", [{}])[0].get("b64_json", "")}
                 # Fallback: no upscale endpoint — return error
                 return {"error": f"Upscale endpoint not available ({resp.status_code})"}
-        except Exception as e:
-            return {"error": str(e)}
+        except Exception:
+            logger.exception("ai_upscale: request failed")
+            return {"error": "Upscale request failed"}
 
     # ---- POST /api/gallery/style-transfer ----
     @router.post("/api/gallery/style-transfer")
@@ -431,8 +466,9 @@ def setup_gallery_routes() -> APIRouter:
                     if img_data:
                         return {"image": img_data}
                 return {"error": f"Style transfer failed ({resp.status_code})"}
-        except Exception as e:
-            return {"error": str(e)}
+        except Exception:
+            logger.exception("style_transfer: request failed")
+            return {"error": "Style transfer failed"}
 
     # ---- GET /api/gallery/tags ----
     @router.get("/api/gallery/tags")
@@ -588,9 +624,9 @@ def setup_gallery_routes() -> APIRouter:
                 "tags": sorted(all_tags),
                 "models": all_models,
             }
-        except Exception as e:
-            logger.error(f"Failed to fetch gallery library: {e}")
-            raise HTTPException(500, f"Failed to fetch gallery library: {e}")
+        except Exception:
+            logger.exception("Failed to fetch gallery library")
+            raise HTTPException(500, "Failed to fetch gallery library")
         finally:
             db.close()
 
@@ -766,9 +802,10 @@ def setup_gallery_routes() -> APIRouter:
             return _image_to_dict(img)
         except HTTPException:
             raise
-        except Exception as e:
+        except Exception:
             db.rollback()
-            raise HTTPException(500, str(e))
+            logger.exception("patch_gallery_image: update failed")
+            raise HTTPException(500, "Image update failed")
         finally:
             db.close()
 
@@ -845,9 +882,10 @@ def setup_gallery_routes() -> APIRouter:
                     cleared += 1
             db.commit()
             return {"ok": True, "cleared": cleared}
-        except Exception as e:
+        except Exception:
             db.rollback()
-            raise HTTPException(500, str(e))
+            logger.exception("clear_gallery_user_tags: failed")
+            raise HTTPException(500, "Tag update failed")
         finally:
             db.close()
 
@@ -871,9 +909,10 @@ def setup_gallery_routes() -> APIRouter:
                     cleared += 1
             db.commit()
             return {"ok": True, "cleared": cleared}
-        except Exception as e:
+        except Exception:
             db.rollback()
-            raise HTTPException(500, str(e))
+            logger.exception("clear_gallery_ai_tags: failed")
+            raise HTTPException(500, "Tag update failed")
         finally:
             db.close()
 
@@ -909,9 +948,10 @@ def setup_gallery_routes() -> APIRouter:
                     img.tags = ', '.join(cleaned)
             db.commit()
             return {"ok": True, "rows_touched": rows_touched, "tags_removed": tags_removed}
-        except Exception as e:
+        except Exception:
             db.rollback()
-            raise HTTPException(500, str(e))
+            logger.exception("dedupe_gallery_tags: failed")
+            raise HTTPException(500, "Tag deduplication failed")
         finally:
             db.close()
 
@@ -1029,9 +1069,10 @@ def setup_gallery_routes() -> APIRouter:
             return {"status": "deleted", "id": image_id}
         except HTTPException:
             raise
-        except Exception as e:
+        except Exception:
             db.rollback()
-            raise HTTPException(500, str(e))
+            logger.exception("delete_gallery_image: failed")
+            raise HTTPException(500, "Image deletion failed")
         finally:
             db.close()
 
@@ -1044,21 +1085,22 @@ def setup_gallery_routes() -> APIRouter:
         import httpx
         user = require_privilege(request, "can_generate_images")
         body = await request.json()
-        # Use endpoint from request body (editor dropdown) or fall back to DB lookup
-        base = (body.pop("_endpoint", "") or "").rstrip("/")
+        # Use endpoint from request body (editor dropdown) or fall back to DB lookup.
+        # Store as requested_base to avoid carrying user input into the outbound request.
+        requested_base = (body.pop("_endpoint", "") or "").rstrip("/")
         # SSRF hardening: validate a client-supplied endpoint before any
         # outbound request (mirrors routes/embedding_routes.py).
-        if base:
+        if requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
-                base,
+                requested_base,
                 block_private=os.getenv("IMAGE_BLOCK_PRIVATE_IPS", "false").lower() == "true",
             )
             if not ok:
                 raise HTTPException(400, f"Rejected endpoint URL: {reason}")
         chosen_model = (body.pop("_model", "") or "").strip()
         api_key = None
-        if not base:
+        if not requested_base:
             db = SessionLocal()
             try:
                 ep = _first_visible_image_endpoint(db, user)
@@ -1069,32 +1111,23 @@ def setup_gallery_routes() -> APIRouter:
             finally:
                 db.close()
         else:
-            # Pull api_key from the matching DB row so OpenAI auth works.
-            # Users may have stored base_url with/without /v1 suffix and with/without
-            # trailing slash, so compare normalized forms.
-            def _norm_url(u: str) -> str:
-                if not u:
-                    return u
-                u = u.rstrip("/")
-                if u.endswith("/v1"):
-                    u = u[:-3]
-                return u
-            _target = _norm_url(base)
+            # Resolve the client-supplied base to a registered visible endpoint.
+            # Admins are not exempted — gallery proxy routes must use a DB row
+            # so the outbound URL never depends directly on request-body input.
             db = SessionLocal()
             try:
-                ep = _visible_image_endpoint_for_base(db, _target, user)
-                if ep:
-                    base = (ep.base_url or base).rstrip("/")
-                    api_key = ep.api_key
-                elif user and not _current_user_is_admin(request, user):
+                ep = _visible_image_endpoint_for_base(db, requested_base, user)
+                if not ep:
                     raise HTTPException(403, "Choose a registered image endpoint")
+                base = ep.base_url.rstrip("/")
+                api_key = ep.api_key
             finally:
                 db.close()
 
         if not base.endswith("/v1"):
             base += "/v1"
 
-        is_openai = "api.openai.com" in base
+        is_openai = _is_openai_api_base(base)
 
         if is_openai:
             # OpenAI path: /v1/images/edits with gpt-image-1.
@@ -1131,8 +1164,9 @@ def setup_gallery_routes() -> APIRouter:
                 mask_buf.seek(0)
             except HTTPException:
                 raise
-            except Exception as e:
-                raise HTTPException(400, f"Failed to prepare OpenAI request: {e}")
+            except Exception:
+                logger.exception("inpaint_proxy: failed to prepare OpenAI request")
+                raise HTTPException(400, "Failed to prepare inpaint request")
 
             width = int(body.get("width") or 1024)
             height = int(body.get("height") or 1024)
@@ -1163,9 +1197,10 @@ def setup_gallery_routes() -> APIRouter:
             headers = {"Authorization": f"Bearer {api_key}"}
             try:
                 async with httpx.AsyncClient(timeout=120) as client:
-                    r = await client.post(f"{base}/images/edits", headers=headers, data=data, files=files)
+                    r = await client.post(_join_checked_gallery_endpoint(base, "/images/edits"), headers=headers, data=data, files=files)
                     if r.status_code != 200:
-                        raise HTTPException(r.status_code, f"OpenAI edit failed: {r.text[:300]}")
+                        logger.error("inpaint_proxy OpenAI edit: status %s", r.status_code)
+                        raise HTTPException(r.status_code, "OpenAI edit failed")
                     result = r.json()
                     raw_b64 = None
                     if result.get("data"):
@@ -1212,16 +1247,18 @@ def setup_gallery_routes() -> APIRouter:
             if chosen_model:
                 body["model"] = chosen_model
             async with httpx.AsyncClient(timeout=120) as client:
-                r = await client.post(f"{base}/images/inpaint", json=body)
+                r = await client.post(_join_checked_gallery_endpoint(base, "/images/inpaint"), json=body)
                 if r.status_code != 200:
-                    raise HTTPException(r.status_code, f"Inpaint failed: {r.text[:200]}")
+                    logger.error("inpaint_proxy diffusion: status %s", r.status_code)
+                    raise HTTPException(r.status_code, "Inpaint request failed")
                 return r.json()
         except httpx.TimeoutException:
             raise HTTPException(504, "Inpaint request timed out (120s)")
         except HTTPException:
             raise
-        except Exception as e:
-            raise HTTPException(502, f"Inpaint error: {str(e)}")
+        except Exception:
+            logger.exception("inpaint_proxy: request failed")
+            raise HTTPException(502, "Inpaint request failed")
 
     # ---- POST /api/image/harmonize — proper img2img call ----
     # Earlier version routed through inpaint with a full-white mask, but
@@ -1243,24 +1280,23 @@ def setup_gallery_routes() -> APIRouter:
         if not image_b64:
             raise HTTPException(400, "No image provided")
 
-        endpoint = (body.get("_endpoint") or "").rstrip("/")
+        requested_base = (body.get("_endpoint") or "").rstrip("/")
         # SSRF hardening: a client-supplied endpoint is fetched server-side
         # below, so validate it first (mirrors routes/embedding_routes.py).
         # Local-first means loopback/LAN is allowed by default; the cloud
         # metadata range and non-HTTP(S) schemes are always rejected.
-        if endpoint:
+        if requested_base:
             from src.url_safety import check_outbound_url
             ok, reason = check_outbound_url(
-                endpoint,
+                requested_base,
                 block_private=os.getenv("IMAGE_BLOCK_PRIVATE_IPS", "false").lower() == "true",
             )
             if not ok:
                 raise HTTPException(400, f"Rejected endpoint URL: {reason}")
         model = (body.get("_model") or "").strip()
 
-        base = endpoint
         api_key = None
-        if not base:
+        if not requested_base:
             db = SessionLocal()
             try:
                 ep = _first_visible_image_endpoint(db, user)
@@ -1271,14 +1307,16 @@ def setup_gallery_routes() -> APIRouter:
             finally:
                 db.close()
         else:
+            # Resolve the client-supplied base to a registered visible endpoint.
+            # Admins are not exempted — gallery proxy routes must use a DB row
+            # so the outbound URL never depends directly on request-body input.
             db = SessionLocal()
             try:
-                ep = _visible_image_endpoint_for_base(db, base, user)
-                if ep:
-                    base = (ep.base_url or base).rstrip("/")
-                    api_key = ep.api_key
-                elif user and not _current_user_is_admin(request, user):
+                ep = _visible_image_endpoint_for_base(db, requested_base, user)
+                if not ep:
                     raise HTTPException(403, "Choose a registered image endpoint")
+                base = ep.base_url.rstrip("/")
+                api_key = ep.api_key
             finally:
                 db.close()
 
@@ -1313,7 +1351,7 @@ def setup_gallery_routes() -> APIRouter:
         # source. Earlier hack (alpha-blend the regen back at `strength`)
         # produced visibly broken results, so we refuse and tell the
         # user to spin up a real diffusion endpoint instead.
-        if "api.openai.com" in base:
+        if _is_openai_api_base(base):
             raise HTTPException(400,
                 "Harmonize needs a diffusion server that supports img2img "
                 "(SD WebUI / Forge / Comfy). OpenAI's API doesn't expose "
@@ -1378,14 +1416,16 @@ def setup_gallery_routes() -> APIRouter:
         # 1024×1024 inference pass on slower setups.
         async with httpx.AsyncClient(timeout=240) as client:
             for path, kind, payload in candidates:
-                target = base_root + path if path.startswith("/sdapi") else base + path
+                _effective_base = base_root if path.startswith("/sdapi") else base
+                target = _join_checked_gallery_endpoint(_effective_base, path)
                 try:
                     r = await client.post(target, json=payload, headers=headers)
                     if r.status_code == 404:
                         last_err = f"{path}: 404"
                         continue  # try next variant
                     if r.status_code != 200:
-                        last_err = f"{path}: {r.status_code} {r.text[:120]}"
+                        logger.warning("harmonize: %s returned %s", path, r.status_code)
+                        last_err = f"{path}: {r.status_code}"
                         continue
                     data = r.json()
                     # Normalise return shape.
@@ -1394,8 +1434,8 @@ def setup_gallery_routes() -> APIRouter:
                         # surface it now instead of trying the other routes
                         # (otherwise the real error gets buried under 404s).
                         if data.get("error") and not data.get("image"):
-                            raise HTTPException(502,
-                                f"Diffusion server error at {path}: {data['error']}")
+                            logger.warning("harmonize: server error at %s: %s", path, data.get("error"))
+                            raise HTTPException(502, f"Diffusion server error at {path}")
                         if data.get("image"):
                             return {"image": data["image"]}
                         if data.get("images") and isinstance(data["images"], list):
@@ -1415,15 +1455,15 @@ def setup_gallery_routes() -> APIRouter:
                                 if img_b64:
                                     return {"image": img_b64}
                     last_err = f"{path}: server returned no image"
-                except httpx.ConnectError as e:
-                    raise HTTPException(502, f"Can't reach diffusion server at {base}: {e}")
+                except httpx.ConnectError:
+                    logger.warning("harmonize: can't reach diffusion server at %s", base)
+                    raise HTTPException(502, "Can't reach diffusion server")
                 except httpx.TimeoutException:
                     raise HTTPException(504, "Harmonize timed out (240s) — restart the diffusion server or lower Color match / disable Seam fix")
         raise HTTPException(502,
-            f"None of the img2img routes worked on {base}. "
-            f"Last response: {last_err or 'unknown'}. "
-            "Your diffusion server needs to expose one of /v1/images/harmonize, "
-            "/v1/images/img2img, /v1/images/variations, or /sdapi/v1/img2img.")
+            "No supported img2img route responded. "
+            "Your diffusion server needs to expose one of: "
+            "/v1/images/harmonize, /v1/images/img2img, /v1/images/variations, /sdapi/v1/img2img.")
 
     # ---- POST /api/image/sharpen ----
     @router.post("/api/image/sharpen")
@@ -1467,8 +1507,8 @@ def setup_gallery_routes() -> APIRouter:
             import base64, io
             from PIL import Image
             import numpy as np
-        except ImportError as e:
-            raise HTTPException(500, f"Server missing dependency: {e}")
+        except ImportError:
+            raise HTTPException(500, "Server missing a required dependency")
         # Decode source image (RGB; Real-ESRGAN doesn't preserve alpha).
         img_bytes = base64.b64decode(image_b64)
         src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
@@ -1495,9 +1535,9 @@ def setup_gallery_routes() -> APIRouter:
             buf = io.BytesIO()
             out_img.save(buf, format="PNG")
             return {"image": base64.b64encode(buf.getvalue()).decode()}
-        except Exception as e:
-            logger.warning(f"Denoise failed: {e}")
-            return {"error": f"Denoise failed: {e}"}
+        except Exception:
+            logger.warning("Denoise failed", exc_info=True)
+            return {"error": "Denoise failed"}
 
     # ---- POST /api/image/upscale-local ----
     # Local Real-ESRGAN upscale (2× or 4×). Self-contained — no diffusion
@@ -1518,8 +1558,8 @@ def setup_gallery_routes() -> APIRouter:
             import base64, io
             from PIL import Image
             import numpy as np
-        except ImportError as e:
-            raise HTTPException(500, f"Server missing dependency: {e}")
+        except ImportError:
+            raise HTTPException(500, "Server missing a required dependency")
         img_bytes = base64.b64decode(image_b64)
         src = Image.open(io.BytesIO(img_bytes)).convert("RGB")
         try:
@@ -1543,9 +1583,9 @@ def setup_gallery_routes() -> APIRouter:
             buf = io.BytesIO()
             out_img.save(buf, format="PNG")
             return {"image": base64.b64encode(buf.getvalue()).decode()}
-        except Exception as e:
-            logger.warning(f"Upscale failed: {e}")
-            return {"error": f"Upscale failed: {e}"}
+        except Exception:
+            logger.warning("AI upscale failed", exc_info=True)
+            return {"error": "AI upscale failed"}
 
     # ---- POST /api/image/remove-bg ----
     @router.post("/api/image/remove-bg")
@@ -1703,8 +1743,9 @@ def setup_gallery_routes() -> APIRouter:
             buf = io.BytesIO()
             enhanced.save(buf, format="PNG")
             return {"image": base64.b64encode(buf.getvalue()).decode(), "method": "pil"}
-        except Exception as e:
-            raise HTTPException(500, f"Face enhancement failed: {str(e)}")
+        except Exception:
+            logger.exception("enhance_face: failed")
+            raise HTTPException(500, "Face enhancement failed")
 
     # ---- Album management (path-param routes) ----
 
@@ -1899,9 +1940,8 @@ def setup_gallery_routes() -> APIRouter:
             async with httpx.AsyncClient(timeout=60) as client:
                 resp = await client.post(chat_url, json=payload, headers=h)
                 if resp.status_code != 200:
-                    body = resp.text[:500]
-                    logger.error(f"Vision model {resp.status_code}: {body}")
-                    return {"error": f"Vision model returned {resp.status_code}: {body[:200]}"}
+                    logger.error("ai_tag vision model: status %s: %s", resp.status_code, resp.text[:500])
+                    return {"error": "Vision model request failed"}
                 data = resp.json()
                 # Anthropic returns content[0].text, OpenAI returns choices[0].message.content
                 if provider == "anthropic":
@@ -1917,9 +1957,9 @@ def setup_gallery_routes() -> APIRouter:
             return {"ok": True, "ai_tags": tag_str}
         except HTTPException:
             raise
-        except Exception as e:
-            logger.error(f"AI tagging failed: {e}")
-            return {"error": str(e)}
+        except Exception:
+            logger.exception("AI tagging failed")
+            return {"error": "Auto-tagging failed"}
         finally:
             db.close()
 

routes/gallery_helpers.py

@@ -1,9 +1,9 @@
-"""Backward-compat shim — canonical location is routes/gallery/gallery_helpers.py.
+"""Backward-compat shim - canonical location is routes/gallery/gallery_helpers.py.
 
 This module is replaced in ``sys.modules`` by the canonical module object so
 that ``import routes.gallery_helpers``, ``from routes.gallery_helpers import X``,
 ``importlib.import_module("routes.gallery_helpers")``, and
-``monkeypatch.setattr(routes.gallery_helpers, ...)`` all operate on the *same*
+``monkeypatch.setattr(routes.gallery_helpers, ...)`` all operate on the same
 object. Keeps existing import paths working after slice 2a (#4082/#4071).
 """
 

routes/gallery_routes.py

@@ -1,9 +1,9 @@
-"""Backward-compat shim — canonical location is routes/gallery/gallery_routes.py.
+"""Backward-compat shim - canonical location is routes/gallery/gallery_routes.py.
 
 This module is replaced in ``sys.modules`` by the canonical module object so
 that ``import routes.gallery_routes``, ``from routes.gallery_routes import X``,
 ``importlib.import_module("routes.gallery_routes")``, and
-``monkeypatch.setattr(routes.gallery_routes, ...)`` all operate on the *same*
+``monkeypatch.setattr(routes.gallery_routes, ...)`` all operate on the same
 object the application actually uses. Keeps existing import paths working
 after slice 2a (#4082/#4071). Source-introspection tests read the canonical
 file by path.

routes/history/__init__.py

@@ -0,0 +1,5 @@
+"""History route domain package (slice 2d, #4082/#4071).
+
+Contains history_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/history_routes.py re-exports from here.
+"""

routes/history/history_routes.py

@@ -0,0 +1,768 @@
+"""History routes — session history, truncation, fork, conversation topics."""
+
+import json
+import uuid
+import logging
+import re
+from typing import Dict, Any, Optional
+
+from fastapi import APIRouter, Request, HTTPException
+
+from core.models import ChatMessage
+from core.database import SessionLocal, ChatMessage as DbChatMessage, Session as DbSession
+from src.topic_analyzer import analyze_topics
+from routes.session_routes import (
+    _message_role,
+    _message_text,
+    _reject_compact_during_active_run,
+    _verify_session_owner,
+)
+
+logger = logging.getLogger(__name__)
+
+_HISTORY_INLINE_MEDIA_THRESHOLD = 200_000
+_DATA_IMAGE_RE = re.compile(r"data:image/[^;,\"]+;base64,[A-Za-z0-9+/=\s]+")
+
+
+def _history_display_content(content: Any) -> Any:
+    """Return a lightweight browser-display copy of stored message content.
+
+    Older multimodal user messages may be persisted as a JSON *string*
+    containing image_url blocks with inline base64 image bytes. Those bytes are
+    needed for model calls when the turn is first sent, but they should not be
+    sent back through /api/history every time the user opens the chat. The
+    attachment metadata already carries file ids/names for the UI cards.
+    """
+    if isinstance(content, list):
+        text_parts = []
+        omitted_media = 0
+        for block in content:
+            if not isinstance(block, dict):
+                continue
+            if block.get("type") == "text":
+                text = block.get("text")
+                if isinstance(text, str) and text:
+                    text_parts.append(text)
+            elif block.get("type") in {"image_url", "input_image", "audio", "input_audio"}:
+                omitted_media += 1
+        text = "\n".join(text_parts).strip()
+        if omitted_media and not text:
+            return f"[{omitted_media} media attachment{'s' if omitted_media != 1 else ''} omitted from history view]"
+        return text
+
+    if not isinstance(content, str):
+        return content
+    if len(content) < _HISTORY_INLINE_MEDIA_THRESHOLD and "data:image/" not in content:
+        return content
+
+    stripped = content.lstrip()
+    if stripped.startswith("["):
+        try:
+            blocks = json.loads(content)
+        except (json.JSONDecodeError, TypeError, ValueError):
+            blocks = None
+        if isinstance(blocks, list):
+            text_parts = []
+            for block in blocks:
+                if not isinstance(block, dict):
+                    continue
+                if block.get("type") == "text":
+                    text = block.get("text")
+                    if isinstance(text, str) and text:
+                        text_parts.append(text)
+            if text_parts:
+                return "\n".join(text_parts).strip()
+
+    if "data:image/" in content:
+        return _DATA_IMAGE_RE.sub("[inline image omitted from history view]", content)
+    return content
+
+
+def _merge_continue_rows_to_delete(db_messages, db1, db2):
+    """DB rows to delete when merging the last two assistant messages.
+
+    Always the second assistant message (db2), plus ONLY the single
+    intervening "continue" user message (the one carrying "previous response
+    was interrupted") — matching the in-memory merge. The previous code
+    deleted the whole index range between the two assistant rows, destroying
+    any tool/system/user messages in between and desyncing the DB from the
+    in-memory history.
+    """
+    to_delete = [db2]
+    i1 = next((i for i, m in enumerate(db_messages) if m is db1), None)
+    i2 = next((i for i, m in enumerate(db_messages) if m is db2), None)
+    if i1 is not None and i2 is not None and i2 - 1 > i1:
+        between = db_messages[i2 - 1]
+        if getattr(between, "role", "") == "user" and            "previous response was interrupted" in (getattr(between, "content", "") or ""):
+            to_delete.append(between)
+    return to_delete
+
+
+def setup_history_routes(session_manager) -> APIRouter:
+    router = APIRouter(tags=["history"])
+
+    def _db_history_entry(m: DbChatMessage) -> Dict[str, Any]:
+        entry = {"role": m.role, "content": _history_display_content(m.content)}
+        meta = {}
+        if m.meta_data:
+            try:
+                meta = json.loads(m.meta_data) or {}
+            except (json.JSONDecodeError, ValueError):
+                meta = {}
+        if m.timestamp and "timestamp" not in meta:
+            meta["timestamp"] = m.timestamp.isoformat() + "Z"
+        if meta:
+            entry["metadata"] = meta
+        return entry
+
+    @router.get("/api/history/{session_id}")
+    async def get_session_history(
+        request: Request,
+        session_id: str,
+        limit: Optional[int] = None,
+        offset: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        _verify_session_owner(request, session_id)
+        if limit is not None:
+            page_limit = max(1, min(int(limit), 100))
+            db = SessionLocal()
+            try:
+                db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
+                if db_session is None:
+                    raise HTTPException(404, f"Session '{session_id}' not found")
+
+                total = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id)
+                    .count()
+                )
+                page_offset = int(offset) if offset is not None else max(total - page_limit, 0)
+                page_offset = max(0, min(page_offset, total))
+                rows = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id)
+                    .order_by(DbChatMessage.timestamp)
+                    .offset(page_offset)
+                    .limit(page_limit)
+                    .all()
+                )
+                history_dict = [
+                    entry for entry in (_db_history_entry(m) for m in rows)
+                    if not (entry.get("metadata") or {}).get("hidden")
+                ]
+                return {
+                    "history": history_dict,
+                    "model": db_session.model,
+                    "endpoint_url": db_session.endpoint_url,
+                    "name": db_session.name,
+                    "offset": page_offset,
+                    "limit": page_limit,
+                    "total": total,
+                    "has_more_before": page_offset > 0,
+                    "has_more_after": page_offset + len(rows) < total,
+                }
+            finally:
+                db.close()
+
+        try:
+            session = session_manager.get_session(session_id)
+        except KeyError:
+            raise HTTPException(404, f"Session '{session_id}' not found")
+
+        history_dict = []
+        for msg in session.history:
+            if isinstance(msg, ChatMessage):
+                # Skip hidden messages (e.g. compaction summaries for AI context)
+                if msg.metadata and msg.metadata.get("hidden"):
+                    continue
+                entry = {"role": msg.role, "content": _history_display_content(msg.content)}
+                if msg.metadata:
+                    entry["metadata"] = msg.metadata
+                history_dict.append(entry)
+            elif isinstance(msg, dict):
+                if msg.get("metadata", {}).get("hidden"):
+                    continue
+                entry = {
+                    "role": msg.get("role", ""),
+                    "content": _history_display_content(msg.get("content", "")),
+                }
+                if msg.get("metadata"):
+                    entry["metadata"] = msg["metadata"]
+                history_dict.append(entry)
+
+        # Fallback: load from DB if in-memory is empty
+        if not history_dict:
+            db = SessionLocal()
+            try:
+                db_messages = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id)
+                    .order_by(DbChatMessage.timestamp)
+                    .all()
+                )
+                db_history = []
+                for m in db_messages:
+                    db_history.append(_db_history_entry(m))
+                if db_history:
+                    # Rebuild in-memory history from the full set so hidden
+                    # messages (e.g. compaction summaries) are kept for AI context.
+                    session.history = [
+                        ChatMessage(role=m["role"], content=m["content"], metadata=m.get("metadata"))
+                        for m in db_history
+                    ]
+                # Response excludes hidden messages, matching the in-memory path.
+                history_dict = [
+                    m for m in db_history
+                    if not (m.get("metadata") or {}).get("hidden")
+                ]
+            except Exception as e:
+                logger.error(f"DB fallback failed for {session_id}: {e}")
+            finally:
+                db.close()
+
+        return {
+            "history": history_dict,
+            "model": session.model,
+            "endpoint_url": session.endpoint_url,
+            "name": session.name,
+        }
+
+    @router.post("/api/session/{session_id}/truncate")
+    async def truncate_session(request: Request, session_id: str):
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            keep_count = body.get("keep_count", 0)
+            result = session_manager.truncate_messages(session_id, keep_count)
+            return {"status": "ok", "kept": keep_count, "truncated": result}
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except Exception as e:
+            logger.error(f"Truncate error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/message")
+    async def add_message(request: Request, session_id: str):
+        """Add a message to a session (for slash command persistence)."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            role = body.get("role", "assistant")
+            content = body.get("content", "")
+            if not content:
+                raise HTTPException(400, "content is required")
+            msg = ChatMessage(role=role, content=content, metadata=body.get("metadata"))
+            session_manager.add_message(session_id, msg)
+            return {"status": "ok"}
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+
+    @router.post("/api/session/{session_id}/delete-messages")
+    async def delete_messages(request: Request, session_id: str):
+        """Delete specific messages by DB ID (or legacy index)."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            msg_ids = body.get("msg_ids", [])
+            indices = body.get("indices")  # legacy fallback
+
+            session = session_manager.get_session(session_id)
+            db = SessionLocal()
+            try:
+                if msg_ids:
+                    # New ID-based delete
+                    deleted = 0
+                    for mid in msg_ids:
+                        db_msg = db.query(DbChatMessage).filter(
+                            DbChatMessage.id == mid,
+                            DbChatMessage.session_id == session_id,
+                        ).first()
+                        if db_msg:
+                            db.delete(db_msg)
+                            deleted += 1
+
+                    # Remove from in-memory history by matching _db_id
+                    def _get_db_id(m):
+                        meta = m.metadata if isinstance(m, ChatMessage) else (m.get('metadata') if isinstance(m, dict) else None)
+                        return meta.get('_db_id') if isinstance(meta, dict) else None
+                    session.history = [m for m in session.history if _get_db_id(m) not in msg_ids]
+                elif indices:
+                    # Legacy index-based delete
+                    indices = sorted(indices, reverse=True)
+                    db_messages = db.query(DbChatMessage).filter(
+                        DbChatMessage.session_id == session_id
+                    ).order_by(DbChatMessage.timestamp).all()
+
+                    deleted = 0
+                    for idx in indices:
+                        if 0 <= idx < len(db_messages):
+                            db.delete(db_messages[idx])
+                            deleted += 1
+                        if 0 <= idx < len(session.history):
+                            session.history.pop(idx)
+                else:
+                    return {"status": "ok", "deleted": 0}
+
+                session.message_count = len(session.history)
+                db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
+                if db_session:
+                    db_session.message_count = len(session.history)
+                    from datetime import datetime, timezone
+                    db_session.updated_at = datetime.now(timezone.utc)
+
+                db.commit()
+                return {"status": "ok", "deleted": deleted}
+            finally:
+                db.close()
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except Exception as e:
+            logger.error(f"Delete messages error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/edit-message")
+    async def edit_message(request: Request, session_id: str):
+        """Edit the content of a message by its database ID."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            msg_id = body.get("msg_id")
+            content = body.get("content")
+            if not msg_id or content is None:
+                raise HTTPException(400, "msg_id and content are required")
+
+            session = session_manager.get_session(session_id)
+            db = SessionLocal()
+            try:
+                db_msg = db.query(DbChatMessage).filter(
+                    DbChatMessage.id == msg_id,
+                    DbChatMessage.session_id == session_id,
+                ).first()
+                if not db_msg:
+                    raise HTTPException(404, "Message not found")
+
+                db_msg.content = content
+                meta = {}
+                if db_msg.meta_data:
+                    try: meta = json.loads(db_msg.meta_data)
+                    except (json.JSONDecodeError, ValueError): pass
+                meta['edited'] = True
+                db_msg.meta_data = json.dumps(meta)
+
+                # Update in-memory history by matching _db_id
+                for hmsg in session.history:
+                    hmeta = hmsg.metadata if isinstance(hmsg, ChatMessage) else hmsg.get('metadata')
+                    if isinstance(hmeta, dict) and hmeta.get('_db_id') == msg_id:
+                        if isinstance(hmsg, ChatMessage):
+                            hmsg.content = content
+                            hmsg.metadata['edited'] = True
+                        elif isinstance(hmsg, dict):
+                            hmsg['content'] = content
+                            hmsg['metadata']['edited'] = True
+                        break
+
+                db.commit()
+                return {"status": "ok"}
+            finally:
+                db.close()
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except HTTPException:
+            raise
+        except Exception as e:
+            logger.error(f"Edit message error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/mark-stopped")
+    async def mark_stopped(request: Request, session_id: str):
+        """Mark the last assistant message as stopped by user."""
+        _verify_session_owner(request, session_id)
+        try:
+            session = session_manager.get_session(session_id)
+            # Find last assistant message and add stopped metadata
+            for msg in reversed(session.history):
+                if (isinstance(msg, ChatMessage) and msg.role == 'assistant') or \
+                   (isinstance(msg, dict) and msg.get('role') == 'assistant'):
+                    if isinstance(msg, ChatMessage):
+                        if not msg.metadata:
+                            msg.metadata = {}
+                        msg.metadata['stopped'] = True
+                        if not msg.metadata.get('model'):
+                            msg.metadata['model'] = session.model
+                    else:
+                        if 'metadata' not in msg:
+                            msg['metadata'] = {}
+                        msg['metadata']['stopped'] = True
+                        if not msg['metadata'].get('model'):
+                            msg['metadata']['model'] = session.model
+                    break
+            # Also update in DB
+            db = SessionLocal()
+            try:
+                import json as _json
+                db_messages = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id, DbChatMessage.role == 'assistant')
+                    .order_by(DbChatMessage.timestamp.desc())
+                    .first()
+                )
+                if db_messages:
+                    meta = {}
+                    if db_messages.meta_data:
+                        try:
+                            meta = _json.loads(db_messages.meta_data)
+                        except (json.JSONDecodeError, ValueError):
+                            pass
+                    meta['stopped'] = True
+                    if not meta.get('model'):
+                        meta['model'] = session.model
+                    db_messages.meta_data = _json.dumps(meta)
+                    db.commit()
+            finally:
+                db.close()
+            session_manager.save_sessions()
+            return {"status": "ok"}
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except Exception as e:
+            logger.error(f"Mark stopped error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/update-last-meta")
+    async def update_last_meta(request: Request, session_id: str):
+        """Merge metadata into the last assistant message (e.g. save variants)."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            meta_update = body.get("metadata", {})
+            session = session_manager.get_session(session_id)
+
+            # Update in-memory
+            for msg in reversed(session.history):
+                if (isinstance(msg, ChatMessage) and msg.role == 'assistant') or \
+                   (isinstance(msg, dict) and msg.get('role') == 'assistant'):
+                    if isinstance(msg, ChatMessage):
+                        if not msg.metadata:
+                            msg.metadata = {}
+                        msg.metadata.update(meta_update)
+                    else:
+                        if 'metadata' not in msg:
+                            msg['metadata'] = {}
+                        msg['metadata'].update(meta_update)
+                    break
+
+            # Update in DB
+            db = SessionLocal()
+            try:
+                import json as _json
+                db_msg = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id, DbChatMessage.role == 'assistant')
+                    .order_by(DbChatMessage.timestamp.desc())
+                    .first()
+                )
+                if db_msg:
+                    meta = {}
+                    if db_msg.meta_data:
+                        try: meta = _json.loads(db_msg.meta_data)
+                        except (json.JSONDecodeError, ValueError): pass
+                    meta.update(meta_update)
+                    db_msg.meta_data = _json.dumps(meta)
+                    db.commit()
+            finally:
+                db.close()
+            session_manager.save_sessions()
+            return {"status": "ok"}
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except Exception as e:
+            logger.error(f"Update last meta error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/merge-last-assistant")
+    async def merge_last_assistant(request: Request, session_id: str):
+        """Merge the last two assistant messages into one (for continue)."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            separator = body.get("separator", "\n\n")
+            session = session_manager.get_session(session_id)
+
+            # Find last two assistant messages in-memory
+            ai_indices = []
+            for i, msg in enumerate(session.history):
+                role = msg.role if isinstance(msg, ChatMessage) else msg.get('role', '')
+                if role == 'assistant':
+                    ai_indices.append(i)
+
+            if len(ai_indices) < 2:
+                return {"status": "ok", "merged": False}
+
+            idx1, idx2 = ai_indices[-2], ai_indices[-1]
+            msg1, msg2 = session.history[idx1], session.history[idx2]
+
+            content1 = msg1.content if isinstance(msg1, ChatMessage) else msg1.get('content', '')
+            content2 = msg2.content if isinstance(msg2, ChatMessage) else msg2.get('content', '')
+            merged_content = content1 + separator + content2
+
+            # Merge metadata
+            meta1 = (msg1.metadata if isinstance(msg1, ChatMessage) else msg1.get('metadata')) or {}
+            meta2 = (msg2.metadata if isinstance(msg2, ChatMessage) else msg2.get('metadata')) or {}
+            merged_meta = {**meta1, **meta2}
+            merged_meta.pop('stopped', None)  # no longer stopped after continue
+
+            # Update first message, remove second
+            if isinstance(msg1, ChatMessage):
+                msg1.content = merged_content
+                msg1.metadata = merged_meta
+            else:
+                msg1['content'] = merged_content
+                msg1['metadata'] = merged_meta
+
+            # Also remove the hidden "continue" user message between them if present
+            # It's the message at idx2-1 if it's a user message with continue text
+            remove_indices = [idx2]
+            if idx2 - 1 > idx1:
+                between = session.history[idx2 - 1]
+                between_role = between.role if isinstance(between, ChatMessage) else between.get('role', '')
+                between_content = between.content if isinstance(between, ChatMessage) else between.get('content', '')
+                if between_role == 'user' and 'previous response was interrupted' in between_content:
+                    remove_indices.insert(0, idx2 - 1)
+
+            for ri in sorted(remove_indices, reverse=True):
+                session.history.pop(ri)
+
+            # Update DB
+            db = SessionLocal()
+            try:
+                import json as _json
+                db_messages = (
+                    db.query(DbChatMessage)
+                    .filter(DbChatMessage.session_id == session_id)
+                    .order_by(DbChatMessage.timestamp)
+                    .all()
+                )
+                # Find last two assistant messages in DB
+                ai_db = [(i, m) for i, m in enumerate(db_messages) if m.role == 'assistant']
+                if len(ai_db) >= 2:
+                    (_, db1), (_, db2) = ai_db[-2], ai_db[-1]
+                    db1.content = merged_content
+                    db1.meta_data = _json.dumps(merged_meta)
+
+                    # Mirror the in-memory deletion: remove the second assistant
+                    # message and ONLY the "continue" user message between them
+                    # (not arbitrary tool/system/user rows). The old
+                    # range-delete destroyed every row between the two assistant
+                    # messages, desyncing the DB from the in-memory history.
+                    for _row in _merge_continue_rows_to_delete(db_messages, db1, db2):
+                        db.delete(_row)
+
+                    db.commit()
+            finally:
+                db.close()
+            session_manager.save_sessions()
+            return {"status": "ok", "merged": True}
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        except Exception as e:
+            logger.error(f"Merge assistant error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.post("/api/session/{session_id}/fork")
+    async def fork_session(request: Request, session_id: str):
+        """Create a new session with messages copied up to keep_count."""
+        _verify_session_owner(request, session_id)
+        try:
+            body = await request.json()
+            keep_count = body.get("keep_count", 0)
+
+            # Get the source session
+            source = session_manager.sessions.get(session_id)
+            if not source:
+                raise HTTPException(404, "Session not found")
+
+            # Create new session
+            new_id = str(uuid.uuid4())
+            fork_name = f"\u2ADD {source.name}"
+            new_session = session_manager.create_session(
+                session_id=new_id,
+                name=fork_name,
+                endpoint_url=source.endpoint_url,
+                model=source.model,
+                rag=False,
+                owner=getattr(source, 'owner', None),
+            )
+
+            # Copy messages up to keep_count
+            msgs_to_copy = source.history[:keep_count]
+            for msg in msgs_to_copy:
+                # Copy the metadata dict. Sharing it would let the fork's
+                # persistence (add_message -> _persist_message stamps
+                # _db_id/timestamp onto the dict) mutate the SOURCE session's
+                # in-memory messages, corrupting their _db_id and breaking
+                # edit/delete-by-id on the original conversation.
+                meta = dict(msg.metadata) if isinstance(msg.metadata, dict) else None
+                new_session.add_message(ChatMessage(msg.role, msg.content, meta))
+            try:
+                from src.event_bus import fire_event
+                fire_event("session_created", getattr(source, 'owner', None))
+            except Exception:
+                logger.debug("session_created event dispatch failed", exc_info=True)
+
+            return {
+                "status": "ok",
+                "id": new_id,
+                "name": fork_name,
+                "kept": len(msgs_to_copy),
+            }
+        except HTTPException:
+            raise
+        except Exception as e:
+            logger.error(f"Fork error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    @router.get("/api/conversations/topics")
+    async def get_conversation_topics(request: Request) -> Dict[str, Any]:
+        from src.auth_helpers import require_user
+        user = require_user(request)
+        try:
+            return analyze_topics(session_manager, owner=user or None)
+        except Exception as e:
+            raise HTTPException(500, f"Topic analysis failed: {e}")
+
+    @router.post("/api/session/{session_id}/compact")
+    async def compact_session(request: Request, session_id: str):
+        """Manually trigger context compaction for a session."""
+        _verify_session_owner(request, session_id)
+        from src.auth_helpers import effective_user
+        owner = effective_user(request)
+        try:
+            session = session_manager.get_session(session_id)
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        _reject_compact_during_active_run(session_id)
+
+        try:
+            from src.model_context import estimate_tokens, get_context_length
+            from src.llm_core import llm_call_async
+            from src.endpoint_resolver import resolve_endpoint
+
+            if len(session.history) < 6:
+                return {"status": "ok", "message": "Not enough messages to compact"}
+
+            ctx_len = get_context_length(session.endpoint_url, session.model)
+            messages_before = session.get_context_messages()
+            used_before = estimate_tokens(messages_before)
+            pct_before = round((used_before / ctx_len) * 100, 1) if ctx_len else 0
+            msg_count_before = len(session.history)
+
+            # Keep only last 4 messages, summarize the rest
+            keep_count = 4
+            older = session.history[:-keep_count]
+            recent = session.history[-keep_count:]
+
+            # Build text to summarize
+            convo_text = "\n".join(
+                f"{_message_role(m).upper()}: "
+                f"{_message_text(m)[:2000]}"
+                for m in older
+            )
+
+            # Use utility model if available
+            util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner or None)
+            compact_url = util_url or session.endpoint_url
+            compact_model = util_model or session.model
+            compact_headers = util_headers if util_url else session.headers
+
+            from src.context_compactor import SELF_SUMMARY_SYSTEM_PROMPT
+            compaction_count = sum(1 for m in session.history if isinstance(m, ChatMessage) and "[Conversation summary" in (m.content or ""))
+            sys_prompt = SELF_SUMMARY_SYSTEM_PROMPT.replace("{count}", str(len(older))).replace("{n}", str(compaction_count + 1))
+            summary = await llm_call_async(
+                compact_url, compact_model,
+                [
+                    {"role": "system", "content": sys_prompt},
+                    {"role": "user", "content": convo_text},
+                ],
+                temperature=0.2, max_tokens=1024,
+                headers=compact_headers, timeout=30,
+            )
+
+            # Replace session history: summary as system message + recent messages
+            # System message holds the full summary for AI context
+            system_summary = ChatMessage(
+                role="system",
+                content=f"[Conversation summary — {len(older)} earlier messages were compacted]\n\n{summary}",
+                metadata={"compacted": True, "hidden": True},
+            )
+            # Visible assistant message just shows stats
+            summary_msg = ChatMessage(
+                role="assistant",
+                content=f"**Conversation compacted** — {len(older)} messages summarized, {len(recent)} kept.",
+                metadata={"compacted": True, "messages_removed": len(older)},
+            )
+            new_history = [system_summary, summary_msg] + list(recent)
+            session.history = new_history
+            session.message_count = len(session.history)
+            logger.info(f"Compact: session {session_id} history now has {len(session.history)} messages (was {msg_count_before})")
+
+            # Update DB: delete old messages, insert summary
+            db = SessionLocal()
+            try:
+                db_msgs = db.query(DbChatMessage).filter(
+                    DbChatMessage.session_id == session_id
+                ).order_by(DbChatMessage.timestamp).all()
+
+                # Delete all but the last keep_count
+                for m in db_msgs[:-keep_count]:
+                    db.delete(m)
+
+                # Insert system summary (hidden, for AI context) and visible summary
+                import json as _json
+                import uuid
+                from datetime import datetime, timezone
+                now = datetime.now(timezone.utc)
+                db_sys_summary = DbChatMessage(
+                    id=str(uuid.uuid4()),
+                    session_id=session_id,
+                    role="system",
+                    content=system_summary.content,
+                    meta_data=_json.dumps(system_summary.metadata),
+                    timestamp=now,
+                )
+                db.add(db_sys_summary)
+                db_summary = DbChatMessage(
+                    id=str(uuid.uuid4()),
+                    session_id=session_id,
+                    role="assistant",
+                    content=summary_msg.content,
+                    meta_data=_json.dumps(summary_msg.metadata),
+                    timestamp=now,
+                )
+                db.add(db_summary)
+
+                # Update session record
+                db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
+                if db_session:
+                    db_session.message_count = len(session.history)
+                    db_session.updated_at = datetime.now(timezone.utc)
+                db.commit()
+            finally:
+                db.close()
+
+            session_manager.save_sessions()
+
+            used_after = estimate_tokens(session.get_context_messages())
+            pct_after = round((used_after / ctx_len) * 100, 1) if ctx_len else 0
+
+            return {
+                "status": "ok",
+                "message": f"Compacted: {msg_count_before} msgs → {len(session.history)} msgs ({pct_before}% → {pct_after}%)",
+                "before": pct_before,
+                "after": pct_after,
+            }
+
+        except Exception as e:
+            logger.error(f"Manual compact error {session_id}: {e}")
+            raise HTTPException(500, str(e))
+
+    return router

routes/history_routes.py

@@ -1,662 +1,17 @@
-"""History routes — session history, truncation, fork, conversation topics."""
+"""Backward-compat shim — canonical location is routes/history/history_routes.py.
 
-import json
-import uuid
-import logging
-from typing import Dict, Any
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.history_routes``, ``from routes.history_routes import X``,
+``importlib.import_module("routes.history_routes")``, and the
+``import ... as history_routes`` + ``monkeypatch.setattr(history_routes, ...)``
+pattern used by test_history_compact_tool_calls.py / test_fork_session_metadata.py
+all operate on the *same* object the application actually uses. Keeps existing
+import paths working after slice 2d (#4082/#4071). Source-introspection tests
+read the canonical file by path.
+"""
 
-from fastapi import APIRouter, Request, HTTPException
+import sys as _sys
 
-from core.models import ChatMessage
-from core.database import SessionLocal, ChatMessage as DbChatMessage, Session as DbSession
-from src.topic_analyzer import analyze_topics
-from routes.session_routes import (
-    _message_role,
-    _message_text,
-    _reject_compact_during_active_run,
-    _verify_session_owner,
-)
+from routes.history import history_routes as _canonical  # noqa: F401
 
-logger = logging.getLogger(__name__)
-
-
-def _merge_continue_rows_to_delete(db_messages, db1, db2):
-    """DB rows to delete when merging the last two assistant messages.
-
-    Always the second assistant message (db2), plus ONLY the single
-    intervening "continue" user message (the one carrying "previous response
-    was interrupted") — matching the in-memory merge. The previous code
-    deleted the whole index range between the two assistant rows, destroying
-    any tool/system/user messages in between and desyncing the DB from the
-    in-memory history.
-    """
-    to_delete = [db2]
-    i1 = next((i for i, m in enumerate(db_messages) if m is db1), None)
-    i2 = next((i for i, m in enumerate(db_messages) if m is db2), None)
-    if i1 is not None and i2 is not None and i2 - 1 > i1:
-        between = db_messages[i2 - 1]
-        if getattr(between, "role", "") == "user" and            "previous response was interrupted" in (getattr(between, "content", "") or ""):
-            to_delete.append(between)
-    return to_delete
-
-
-def setup_history_routes(session_manager) -> APIRouter:
-    router = APIRouter(tags=["history"])
-
-    @router.get("/api/history/{session_id}")
-    async def get_session_history(request: Request, session_id: str) -> Dict[str, Any]:
-        _verify_session_owner(request, session_id)
-        try:
-            session = session_manager.get_session(session_id)
-        except KeyError:
-            raise HTTPException(404, f"Session '{session_id}' not found")
-
-        history_dict = []
-        for msg in session.history:
-            if isinstance(msg, ChatMessage):
-                # Skip hidden messages (e.g. compaction summaries for AI context)
-                if msg.metadata and msg.metadata.get("hidden"):
-                    continue
-                entry = {"role": msg.role, "content": msg.content}
-                if msg.metadata:
-                    entry["metadata"] = msg.metadata
-                history_dict.append(entry)
-            elif isinstance(msg, dict):
-                if msg.get("metadata", {}).get("hidden"):
-                    continue
-                entry = {
-                    "role": msg.get("role", ""),
-                    "content": msg.get("content", ""),
-                }
-                if msg.get("metadata"):
-                    entry["metadata"] = msg["metadata"]
-                history_dict.append(entry)
-
-        # Fallback: load from DB if in-memory is empty
-        if not history_dict:
-            db = SessionLocal()
-            try:
-                db_messages = (
-                    db.query(DbChatMessage)
-                    .filter(DbChatMessage.session_id == session_id)
-                    .order_by(DbChatMessage.timestamp)
-                    .all()
-                )
-                import json as _json
-                db_history = []
-                for m in db_messages:
-                    entry = {"role": m.role, "content": m.content}
-                    meta = {}
-                    if m.meta_data:
-                        try:
-                            meta = _json.loads(m.meta_data) or {}
-                        except (json.JSONDecodeError, ValueError):
-                            meta = {}
-                    if m.timestamp and "timestamp" not in meta:
-                        meta["timestamp"] = m.timestamp.isoformat() + "Z"
-                    if meta:
-                        entry["metadata"] = meta
-                    db_history.append(entry)
-                if db_history:
-                    # Rebuild in-memory history from the full set so hidden
-                    # messages (e.g. compaction summaries) are kept for AI context.
-                    session.history = [
-                        ChatMessage(role=m["role"], content=m["content"], metadata=m.get("metadata"))
-                        for m in db_history
-                    ]
-                # Response excludes hidden messages, matching the in-memory path.
-                history_dict = [
-                    m for m in db_history
-                    if not (m.get("metadata") or {}).get("hidden")
-                ]
-            except Exception as e:
-                logger.error(f"DB fallback failed for {session_id}: {e}")
-            finally:
-                db.close()
-
-        return {
-            "history": history_dict,
-            "model": session.model,
-            "endpoint_url": session.endpoint_url,
-            "name": session.name,
-        }
-
-    @router.post("/api/session/{session_id}/truncate")
-    async def truncate_session(request: Request, session_id: str):
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            keep_count = body.get("keep_count", 0)
-            result = session_manager.truncate_messages(session_id, keep_count)
-            return {"status": "ok", "kept": keep_count, "truncated": result}
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except Exception as e:
-            logger.error(f"Truncate error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/message")
-    async def add_message(request: Request, session_id: str):
-        """Add a message to a session (for slash command persistence)."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            role = body.get("role", "assistant")
-            content = body.get("content", "")
-            if not content:
-                raise HTTPException(400, "content is required")
-            msg = ChatMessage(role=role, content=content, metadata=body.get("metadata"))
-            session_manager.add_message(session_id, msg)
-            return {"status": "ok"}
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-
-    @router.post("/api/session/{session_id}/delete-messages")
-    async def delete_messages(request: Request, session_id: str):
-        """Delete specific messages by DB ID (or legacy index)."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            msg_ids = body.get("msg_ids", [])
-            indices = body.get("indices")  # legacy fallback
-
-            session = session_manager.get_session(session_id)
-            db = SessionLocal()
-            try:
-                if msg_ids:
-                    # New ID-based delete
-                    deleted = 0
-                    for mid in msg_ids:
-                        db_msg = db.query(DbChatMessage).filter(
-                            DbChatMessage.id == mid,
-                            DbChatMessage.session_id == session_id,
-                        ).first()
-                        if db_msg:
-                            db.delete(db_msg)
-                            deleted += 1
-
-                    # Remove from in-memory history by matching _db_id
-                    def _get_db_id(m):
-                        meta = m.metadata if isinstance(m, ChatMessage) else (m.get('metadata') if isinstance(m, dict) else None)
-                        return meta.get('_db_id') if isinstance(meta, dict) else None
-                    session.history = [m for m in session.history if _get_db_id(m) not in msg_ids]
-                elif indices:
-                    # Legacy index-based delete
-                    indices = sorted(indices, reverse=True)
-                    db_messages = db.query(DbChatMessage).filter(
-                        DbChatMessage.session_id == session_id
-                    ).order_by(DbChatMessage.timestamp).all()
-
-                    deleted = 0
-                    for idx in indices:
-                        if 0 <= idx < len(db_messages):
-                            db.delete(db_messages[idx])
-                            deleted += 1
-                        if 0 <= idx < len(session.history):
-                            session.history.pop(idx)
-                else:
-                    return {"status": "ok", "deleted": 0}
-
-                session.message_count = len(session.history)
-                db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
-                if db_session:
-                    db_session.message_count = len(session.history)
-                    from datetime import datetime, timezone
-                    db_session.updated_at = datetime.now(timezone.utc)
-
-                db.commit()
-                return {"status": "ok", "deleted": deleted}
-            finally:
-                db.close()
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except Exception as e:
-            logger.error(f"Delete messages error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/edit-message")
-    async def edit_message(request: Request, session_id: str):
-        """Edit the content of a message by its database ID."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            msg_id = body.get("msg_id")
-            content = body.get("content")
-            if not msg_id or content is None:
-                raise HTTPException(400, "msg_id and content are required")
-
-            session = session_manager.get_session(session_id)
-            db = SessionLocal()
-            try:
-                db_msg = db.query(DbChatMessage).filter(
-                    DbChatMessage.id == msg_id,
-                    DbChatMessage.session_id == session_id,
-                ).first()
-                if not db_msg:
-                    raise HTTPException(404, "Message not found")
-
-                db_msg.content = content
-                meta = {}
-                if db_msg.meta_data:
-                    try: meta = json.loads(db_msg.meta_data)
-                    except (json.JSONDecodeError, ValueError): pass
-                meta['edited'] = True
-                db_msg.meta_data = json.dumps(meta)
-
-                # Update in-memory history by matching _db_id
-                for hmsg in session.history:
-                    hmeta = hmsg.metadata if isinstance(hmsg, ChatMessage) else hmsg.get('metadata')
-                    if isinstance(hmeta, dict) and hmeta.get('_db_id') == msg_id:
-                        if isinstance(hmsg, ChatMessage):
-                            hmsg.content = content
-                            hmsg.metadata['edited'] = True
-                        elif isinstance(hmsg, dict):
-                            hmsg['content'] = content
-                            hmsg['metadata']['edited'] = True
-                        break
-
-                db.commit()
-                return {"status": "ok"}
-            finally:
-                db.close()
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except HTTPException:
-            raise
-        except Exception as e:
-            logger.error(f"Edit message error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/mark-stopped")
-    async def mark_stopped(request: Request, session_id: str):
-        """Mark the last assistant message as stopped by user."""
-        _verify_session_owner(request, session_id)
-        try:
-            session = session_manager.get_session(session_id)
-            # Find last assistant message and add stopped metadata
-            for msg in reversed(session.history):
-                if (isinstance(msg, ChatMessage) and msg.role == 'assistant') or \
-                   (isinstance(msg, dict) and msg.get('role') == 'assistant'):
-                    if isinstance(msg, ChatMessage):
-                        if not msg.metadata:
-                            msg.metadata = {}
-                        msg.metadata['stopped'] = True
-                        if not msg.metadata.get('model'):
-                            msg.metadata['model'] = session.model
-                    else:
-                        if 'metadata' not in msg:
-                            msg['metadata'] = {}
-                        msg['metadata']['stopped'] = True
-                        if not msg['metadata'].get('model'):
-                            msg['metadata']['model'] = session.model
-                    break
-            # Also update in DB
-            db = SessionLocal()
-            try:
-                import json as _json
-                db_messages = (
-                    db.query(DbChatMessage)
-                    .filter(DbChatMessage.session_id == session_id, DbChatMessage.role == 'assistant')
-                    .order_by(DbChatMessage.timestamp.desc())
-                    .first()
-                )
-                if db_messages:
-                    meta = {}
-                    if db_messages.meta_data:
-                        try:
-                            meta = _json.loads(db_messages.meta_data)
-                        except (json.JSONDecodeError, ValueError):
-                            pass
-                    meta['stopped'] = True
-                    if not meta.get('model'):
-                        meta['model'] = session.model
-                    db_messages.meta_data = _json.dumps(meta)
-                    db.commit()
-            finally:
-                db.close()
-            session_manager.save_sessions()
-            return {"status": "ok"}
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except Exception as e:
-            logger.error(f"Mark stopped error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/update-last-meta")
-    async def update_last_meta(request: Request, session_id: str):
-        """Merge metadata into the last assistant message (e.g. save variants)."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            meta_update = body.get("metadata", {})
-            session = session_manager.get_session(session_id)
-
-            # Update in-memory
-            for msg in reversed(session.history):
-                if (isinstance(msg, ChatMessage) and msg.role == 'assistant') or \
-                   (isinstance(msg, dict) and msg.get('role') == 'assistant'):
-                    if isinstance(msg, ChatMessage):
-                        if not msg.metadata:
-                            msg.metadata = {}
-                        msg.metadata.update(meta_update)
-                    else:
-                        if 'metadata' not in msg:
-                            msg['metadata'] = {}
-                        msg['metadata'].update(meta_update)
-                    break
-
-            # Update in DB
-            db = SessionLocal()
-            try:
-                import json as _json
-                db_msg = (
-                    db.query(DbChatMessage)
-                    .filter(DbChatMessage.session_id == session_id, DbChatMessage.role == 'assistant')
-                    .order_by(DbChatMessage.timestamp.desc())
-                    .first()
-                )
-                if db_msg:
-                    meta = {}
-                    if db_msg.meta_data:
-                        try: meta = _json.loads(db_msg.meta_data)
-                        except (json.JSONDecodeError, ValueError): pass
-                    meta.update(meta_update)
-                    db_msg.meta_data = _json.dumps(meta)
-                    db.commit()
-            finally:
-                db.close()
-            session_manager.save_sessions()
-            return {"status": "ok"}
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except Exception as e:
-            logger.error(f"Update last meta error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/merge-last-assistant")
-    async def merge_last_assistant(request: Request, session_id: str):
-        """Merge the last two assistant messages into one (for continue)."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            separator = body.get("separator", "\n\n")
-            session = session_manager.get_session(session_id)
-
-            # Find last two assistant messages in-memory
-            ai_indices = []
-            for i, msg in enumerate(session.history):
-                role = msg.role if isinstance(msg, ChatMessage) else msg.get('role', '')
-                if role == 'assistant':
-                    ai_indices.append(i)
-
-            if len(ai_indices) < 2:
-                return {"status": "ok", "merged": False}
-
-            idx1, idx2 = ai_indices[-2], ai_indices[-1]
-            msg1, msg2 = session.history[idx1], session.history[idx2]
-
-            content1 = msg1.content if isinstance(msg1, ChatMessage) else msg1.get('content', '')
-            content2 = msg2.content if isinstance(msg2, ChatMessage) else msg2.get('content', '')
-            merged_content = content1 + separator + content2
-
-            # Merge metadata
-            meta1 = (msg1.metadata if isinstance(msg1, ChatMessage) else msg1.get('metadata')) or {}
-            meta2 = (msg2.metadata if isinstance(msg2, ChatMessage) else msg2.get('metadata')) or {}
-            merged_meta = {**meta1, **meta2}
-            merged_meta.pop('stopped', None)  # no longer stopped after continue
-
-            # Update first message, remove second
-            if isinstance(msg1, ChatMessage):
-                msg1.content = merged_content
-                msg1.metadata = merged_meta
-            else:
-                msg1['content'] = merged_content
-                msg1['metadata'] = merged_meta
-
-            # Also remove the hidden "continue" user message between them if present
-            # It's the message at idx2-1 if it's a user message with continue text
-            remove_indices = [idx2]
-            if idx2 - 1 > idx1:
-                between = session.history[idx2 - 1]
-                between_role = between.role if isinstance(between, ChatMessage) else between.get('role', '')
-                between_content = between.content if isinstance(between, ChatMessage) else between.get('content', '')
-                if between_role == 'user' and 'previous response was interrupted' in between_content:
-                    remove_indices.insert(0, idx2 - 1)
-
-            for ri in sorted(remove_indices, reverse=True):
-                session.history.pop(ri)
-
-            # Update DB
-            db = SessionLocal()
-            try:
-                import json as _json
-                db_messages = (
-                    db.query(DbChatMessage)
-                    .filter(DbChatMessage.session_id == session_id)
-                    .order_by(DbChatMessage.timestamp)
-                    .all()
-                )
-                # Find last two assistant messages in DB
-                ai_db = [(i, m) for i, m in enumerate(db_messages) if m.role == 'assistant']
-                if len(ai_db) >= 2:
-                    (_, db1), (_, db2) = ai_db[-2], ai_db[-1]
-                    db1.content = merged_content
-                    db1.meta_data = _json.dumps(merged_meta)
-
-                    # Mirror the in-memory deletion: remove the second assistant
-                    # message and ONLY the "continue" user message between them
-                    # (not arbitrary tool/system/user rows). The old
-                    # range-delete destroyed every row between the two assistant
-                    # messages, desyncing the DB from the in-memory history.
-                    for _row in _merge_continue_rows_to_delete(db_messages, db1, db2):
-                        db.delete(_row)
-
-                    db.commit()
-            finally:
-                db.close()
-            session_manager.save_sessions()
-            return {"status": "ok", "merged": True}
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        except Exception as e:
-            logger.error(f"Merge assistant error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.post("/api/session/{session_id}/fork")
-    async def fork_session(request: Request, session_id: str):
-        """Create a new session with messages copied up to keep_count."""
-        _verify_session_owner(request, session_id)
-        try:
-            body = await request.json()
-            keep_count = body.get("keep_count", 0)
-
-            # Get the source session
-            source = session_manager.sessions.get(session_id)
-            if not source:
-                raise HTTPException(404, "Session not found")
-
-            # Create new session
-            new_id = str(uuid.uuid4())
-            fork_name = f"\u2ADD {source.name}"
-            new_session = session_manager.create_session(
-                session_id=new_id,
-                name=fork_name,
-                endpoint_url=source.endpoint_url,
-                model=source.model,
-                rag=False,
-                owner=getattr(source, 'owner', None),
-            )
-
-            # Copy messages up to keep_count
-            msgs_to_copy = source.history[:keep_count]
-            for msg in msgs_to_copy:
-                # Copy the metadata dict. Sharing it would let the fork's
-                # persistence (add_message -> _persist_message stamps
-                # _db_id/timestamp onto the dict) mutate the SOURCE session's
-                # in-memory messages, corrupting their _db_id and breaking
-                # edit/delete-by-id on the original conversation.
-                meta = dict(msg.metadata) if isinstance(msg.metadata, dict) else None
-                new_session.add_message(ChatMessage(msg.role, msg.content, meta))
-            try:
-                from src.event_bus import fire_event
-                fire_event("session_created", getattr(source, 'owner', None))
-            except Exception:
-                logger.debug("session_created event dispatch failed", exc_info=True)
-
-            return {
-                "status": "ok",
-                "id": new_id,
-                "name": fork_name,
-                "kept": len(msgs_to_copy),
-            }
-        except HTTPException:
-            raise
-        except Exception as e:
-            logger.error(f"Fork error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    @router.get("/api/conversations/topics")
-    async def get_conversation_topics(request: Request) -> Dict[str, Any]:
-        from src.auth_helpers import require_user
-        user = require_user(request)
-        try:
-            return analyze_topics(session_manager, owner=user or None)
-        except Exception as e:
-            raise HTTPException(500, f"Topic analysis failed: {e}")
-
-    @router.post("/api/session/{session_id}/compact")
-    async def compact_session(request: Request, session_id: str):
-        """Manually trigger context compaction for a session."""
-        _verify_session_owner(request, session_id)
-        from src.auth_helpers import effective_user
-        owner = effective_user(request)
-        try:
-            session = session_manager.get_session(session_id)
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        _reject_compact_during_active_run(session_id)
-
-        try:
-            from src.model_context import estimate_tokens, get_context_length
-            from src.llm_core import llm_call_async
-            from src.endpoint_resolver import resolve_endpoint
-
-            if len(session.history) < 6:
-                return {"status": "ok", "message": "Not enough messages to compact"}
-
-            ctx_len = get_context_length(session.endpoint_url, session.model)
-            messages_before = session.get_context_messages()
-            used_before = estimate_tokens(messages_before)
-            pct_before = round((used_before / ctx_len) * 100, 1) if ctx_len else 0
-            msg_count_before = len(session.history)
-
-            # Keep only last 4 messages, summarize the rest
-            keep_count = 4
-            older = session.history[:-keep_count]
-            recent = session.history[-keep_count:]
-
-            # Build text to summarize
-            convo_text = "\n".join(
-                f"{_message_role(m).upper()}: "
-                f"{_message_text(m)[:2000]}"
-                for m in older
-            )
-
-            # Use utility model if available
-            util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner or None)
-            compact_url = util_url or session.endpoint_url
-            compact_model = util_model or session.model
-            compact_headers = util_headers if util_url else session.headers
-
-            from src.context_compactor import SELF_SUMMARY_SYSTEM_PROMPT
-            compaction_count = sum(1 for m in session.history if isinstance(m, ChatMessage) and "[Conversation summary" in (m.content or ""))
-            sys_prompt = SELF_SUMMARY_SYSTEM_PROMPT.replace("{count}", str(len(older))).replace("{n}", str(compaction_count + 1))
-            summary = await llm_call_async(
-                compact_url, compact_model,
-                [
-                    {"role": "system", "content": sys_prompt},
-                    {"role": "user", "content": convo_text},
-                ],
-                temperature=0.2, max_tokens=1024,
-                headers=compact_headers, timeout=30,
-            )
-
-            # Replace session history: summary as system message + recent messages
-            # System message holds the full summary for AI context
-            system_summary = ChatMessage(
-                role="system",
-                content=f"[Conversation summary — {len(older)} earlier messages were compacted]\n\n{summary}",
-                metadata={"compacted": True, "hidden": True},
-            )
-            # Visible assistant message just shows stats
-            summary_msg = ChatMessage(
-                role="assistant",
-                content=f"**Conversation compacted** — {len(older)} messages summarized, {len(recent)} kept.",
-                metadata={"compacted": True, "messages_removed": len(older)},
-            )
-            new_history = [system_summary, summary_msg] + list(recent)
-            session.history = new_history
-            session.message_count = len(session.history)
-            logger.info(f"Compact: session {session_id} history now has {len(session.history)} messages (was {msg_count_before})")
-
-            # Update DB: delete old messages, insert summary
-            db = SessionLocal()
-            try:
-                db_msgs = db.query(DbChatMessage).filter(
-                    DbChatMessage.session_id == session_id
-                ).order_by(DbChatMessage.timestamp).all()
-
-                # Delete all but the last keep_count
-                for m in db_msgs[:-keep_count]:
-                    db.delete(m)
-
-                # Insert system summary (hidden, for AI context) and visible summary
-                import json as _json
-                import uuid
-                from datetime import datetime, timezone
-                now = datetime.now(timezone.utc)
-                db_sys_summary = DbChatMessage(
-                    id=str(uuid.uuid4()),
-                    session_id=session_id,
-                    role="system",
-                    content=system_summary.content,
-                    meta_data=_json.dumps(system_summary.metadata),
-                    timestamp=now,
-                )
-                db.add(db_sys_summary)
-                db_summary = DbChatMessage(
-                    id=str(uuid.uuid4()),
-                    session_id=session_id,
-                    role="assistant",
-                    content=summary_msg.content,
-                    meta_data=_json.dumps(summary_msg.metadata),
-                    timestamp=now,
-                )
-                db.add(db_summary)
-
-                # Update session record
-                db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
-                if db_session:
-                    db_session.message_count = len(session.history)
-                    db_session.updated_at = datetime.now(timezone.utc)
-                db.commit()
-            finally:
-                db.close()
-
-            session_manager.save_sessions()
-
-            used_after = estimate_tokens(session.get_context_messages())
-            pct_after = round((used_after / ctx_len) * 100, 1) if ctx_len else 0
-
-            return {
-                "status": "ok",
-                "message": f"Compacted: {msg_count_before} msgs → {len(session.history)} msgs ({pct_before}% → {pct_after}%)",
-                "before": pct_before,
-                "after": pct_after,
-            }
-
-        except Exception as e:
-            logger.error(f"Manual compact error {session_id}: {e}")
-            raise HTTPException(500, str(e))
-
-    return router
+_sys.modules[__name__] = _canonical

routes/hwfit_routes.py

@@ -191,7 +191,7 @@ def setup_hwfit_routes():
         return detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh)
 
     @router.get("/models")
-    def get_models(use_case: str = "", sort: str = "newest", limit: int = 50, search: str = "", host: str = "", quant: str = "", ctx: str = "", gpu_count: str = "", gpu_group: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, manual_mode: str = "", manual_gpu_count: str = "", manual_vram_gb: str = "", manual_ram_gb: str = "", manual_backend: str = "", ignore_detected_gpu: bool = False, ignore_detected_ram: bool = False, fit_only: bool = False):
+    def get_models(use_case: str = "", sort: str = "newest", limit: int = 50, search: str = "", host: str = "", quant: str = "", ctx: str = "", gpu_count: str = "", gpu_group: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, refresh_catalog: bool = False, manual_mode: str = "", manual_gpu_count: str = "", manual_vram_gb: str = "", manual_ram_gb: str = "", manual_backend: str = "", ignore_detected_gpu: bool = False, ignore_detected_ram: bool = False, fit_only: bool = False):
         """Rank LLM models against detected hardware and return scored results.
         gpu_count: override GPU count (0 = CPU only, 1-N = simulate N GPUs of the
             active group). gpu_group: index into system.gpu_groups (the homogeneous
@@ -200,11 +200,17 @@ def setup_hwfit_routes():
         fresh=true bypasses the hardware-detection cache."""
         from services.hwfit.hardware import detect_system
         from services.hwfit.fit import rank_models
-        from services.hwfit.models import get_models, model_catalog_path
+        from services.hwfit.models import get_models, model_catalog_path, refresh_dynamic_catalogs
         host, ssh_port = _validate_detection_target(host, ssh_port)
         system = deepcopy(detect_system(host=host, ssh_port=ssh_port, platform=platform, fresh=fresh))
         if system.get("error"):
             return {"system": system, "models": [], "error": system["error"]}
+        catalog_refresh = None
+        if refresh_catalog:
+            try:
+                catalog_refresh = refresh_dynamic_catalogs(force=True)
+            except Exception as e:
+                catalog_refresh = {"error": str(e)}
         if not get_models():
             return {
                 "system": system,
@@ -304,7 +310,10 @@ def setup_hwfit_routes():
             rank_kwargs.pop("target_context", None)
             rank_kwargs.pop("fit_only", None)
         results = rank_models(system, **rank_kwargs)
-        return {"system": system, "models": results}
+        payload = {"system": system, "models": results}
+        if catalog_refresh is not None:
+            payload["catalog_refresh"] = catalog_refresh
+        return payload
 
     @router.get("/profiles")
     def get_serve_profiles(model: str = "", model_path: str = "", host: str = "", ssh_port: str = "", platform: str = "", fresh: bool = False, serve_weights_gb: float = 0.0, serve_quant: str = ""):

routes/memory/__init__.py

@@ -0,0 +1,5 @@
+"""Memory route domain package (slice 2c, #4082/#4071).
+
+Contains memory_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/memory_routes.py re-exports from here.
+"""

routes/memory/memory_routes.py

@@ -0,0 +1,552 @@
+# routes/memory_routes.py
+from fastapi import APIRouter, Form, HTTPException, Request, UploadFile, File
+from typing import Dict, Any, Optional, List
+import json
+import os
+import re
+import tempfile
+import time
+from datetime import datetime
+import logging
+
+# Leading list-marker like "1.", "12)", or "3:" plus surrounding whitespace.
+# Strips one prefix per call so import-from-LLM-output doesn't leave the
+# numbering inside the saved memory text. Bullet markers (-, *, •) are
+# also peeled here for the same reason.
+_LIST_PREFIX_RE = re.compile(r"^\s*(?:\d{1,3}[.):]\s+|[-*•]\s+)")
+
+
+def _strip_list_prefix(text: str) -> str:
+    if not text:
+        return text
+    return _LIST_PREFIX_RE.sub("", text, count=1).strip()
+
+from services.memory import MemoryManager
+from core.session_manager import SessionManager
+from src.request_models import MemoryAddRequest
+from core.database import SessionLocal
+from src.llm_core import llm_call_async
+from services.memory.memory_extractor import audit_memories
+from src.auth_helpers import get_current_user, require_user
+from src.endpoint_resolver import resolve_endpoint
+from src.task_endpoint import resolve_task_endpoint
+from src.upload_limits import read_upload_limited, MEMORY_IMPORT_MAX_BYTES
+
+logger = logging.getLogger(__name__)
+
+
+def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionManager, memory_vector=None):
+    """Set up memory-related routes."""
+    router = APIRouter(prefix="/api/memory", tags=["memory"])
+
+    def _owner(request: Request) -> Optional[str]:
+        return get_current_user(request)
+
+    def _assert_session_owner(session_obj, user):
+        """SECURITY: 404 if the caller does not own this session.
+
+        SessionManager.get_session is NOT owner-scoped — it returns any
+        session by id. These routes accept a caller-supplied session id, so
+        without this gate a user could target another tenant's session and
+        leak their chat history, their session-scoped LLM credentials, or the
+        session title. Mirrors session_routes / webhook_routes ownership.
+        """
+        if user is not None and getattr(session_obj, "owner", None) != user:
+            raise HTTPException(404, "Session not found")
+
+    def _verify_memory_owner(memory: dict, user: Optional[str]):
+        """Raise 404 if user doesn't own this memory.
+
+        SECURITY: strict ownership — previously `mem_owner and mem_owner != user`
+        allowed any user to read/edit/delete memories with an empty/null owner
+        field, which leaked legacy data across the multi-user deploy.
+        """
+        if user is None:
+            return  # Auth disabled
+        if memory.get("owner") != user:
+            raise HTTPException(404, "Memory not found")
+
+    @router.post("/debug")
+    def debug_memory_relevance(request: Request, query: str = Form(...)):
+        """Debug which memories would be triggered for a query"""
+        user = _owner(request)
+        memories = memory_manager.load(owner=user)
+        relevant = memory_manager.get_relevant_memories(query, memories, threshold=0.05)
+
+        return {
+            "query": query,
+            "total_memories": len(memories),
+            "relevant_count": len(relevant),
+            "relevant_memories": [{"text": m["text"], "category": m.get("category", "unknown")}
+                                 for m in relevant]
+        }
+
+    @router.post("/add", response_model=Dict[str, Any])
+    async def api_add_memory(
+        request: Request,
+        memory_data: Optional[MemoryAddRequest] = None
+    ):
+        """Add a new memory entry with optional category, source, and session reference."""
+        from src.auth_helpers import require_privilege
+        require_privilege(request, "can_manage_memory")
+        if memory_data is None:
+            form = await request.form()
+            memory_data = MemoryAddRequest(
+                text=form.get("text"),
+                category=form.get("category", "fact"),
+                source=form.get("source", "user"),
+                session_id=form.get("session_id")
+            )
+
+        user = _owner(request)
+        text = (memory_data.text or "").strip()
+        if not text:
+            raise HTTPException(400, "empty memory")
+        user_mem = memory_manager.load(owner=user)
+        if memory_manager.find_duplicates(text, user_mem):
+            return {"ok": True, "count": len(user_mem), "message": "Memory already exists"}
+
+        if memory_data.session_id:
+            try:
+                session_obj = session_manager.get_session(memory_data.session_id)
+            except KeyError:
+                raise HTTPException(404, "Session not found")
+            _assert_session_owner(session_obj, user)
+
+        new_entry = memory_manager.add_entry(text, memory_data.source, memory_data.category, owner=user)
+        if memory_data.session_id:
+            new_entry["session_id"] = memory_data.session_id
+        all_mem = memory_manager.load_all()
+        all_mem.append(new_entry)
+        memory_manager.save(all_mem)
+        # Sync vector index
+        if memory_vector and memory_vector.healthy:
+            memory_vector.add(new_entry["id"], text)
+        try:
+            from src.event_bus import fire_event
+            fire_event("memory_added", user)
+        except Exception:
+            logger.debug("memory_added event dispatch failed", exc_info=True)
+        return {"ok": True, "count": len([m for m in all_mem if m.get("owner") == user])}
+
+    @router.get("")
+    def api_get_memory(request: Request):
+        """Return all memory entries with their metadata."""
+        user = _owner(request)
+        return {"memory": memory_manager.load(owner=user)}
+
+    @router.post("/search")
+    def search_memories(request: Request, query: str = Form(...), session_id: str = Form(None), category: str = Form(None)):
+        """Search across all memories with optional filters."""
+        user = _owner(request)
+        memories = memory_manager.load(owner=user)
+
+        if session_id:
+            memories = [m for m in memories if m.get("session_id") == session_id]
+
+        if category:
+            memories = [m for m in memories if category in m.get("categories", [m.get("category", "")])]
+
+        relevant = memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
+
+        return {"memories": relevant, "total": len(relevant), "query": query}
+
+    @router.get("/timeline")
+    def memory_timeline(request: Request):
+        """Get memories in chronological order with source session information."""
+        user = _owner(request)
+        memories = memory_manager.load(owner=user)
+        sorted_memories = sorted(memories, key=lambda x: x.get("timestamp", 0), reverse=True)
+
+        results = []
+        for memory in sorted_memories:
+            if "timestamp" in memory:
+                try:
+                    dt = datetime.fromtimestamp(memory["timestamp"])
+                    memory["timestamp_str"] = dt.strftime("%Y-%m-%d %H:%M:%S")
+                except (ValueError, OSError, OverflowError):
+                    memory["timestamp_str"] = "Unknown"
+            else:
+                memory["timestamp_str"] = "Unknown"
+
+            session_id = memory.get("session_id")
+            if session_id and session_id in session_manager.sessions:
+                try:
+                    session = session_manager.get_session(session_id)
+                    if session:
+                        _assert_session_owner(session, user)
+                    memory["session_name"] = session.name if session else f"Session {session_id[:6]}"
+                except KeyError:
+                    memory["session_name"] = "Unknown"
+                except HTTPException as exc:
+                    if exc.status_code != 404:
+                        raise
+                    memory["session_name"] = "Unknown"
+            else:
+                memory["session_name"] = "Unknown"
+
+            results.append(memory)
+
+        return {"timeline": results, "total": len(results)}
+
+    @router.get("/by-session/{session_id}")
+    def get_memory_by_session(request: Request, session_id: str):
+        """Get all memories associated with a specific session."""
+        user = _owner(request)
+        try:
+            _session_obj = session_manager.get_session(session_id)
+        except KeyError:
+            raise HTTPException(404, f"Session {session_id} not found")
+        _assert_session_owner(_session_obj, user)
+        memories = memory_manager.load(owner=user)
+        session_memories = [m for m in memories if m.get("session_id") == session_id]
+
+        session_memories.sort(key=lambda x: x.get("timestamp", 0), reverse=True)
+
+        try:
+            session = session_manager.get_session(session_id)
+            session_name = session.name if session else f"Session {session_id[:6]}"
+        except KeyError:
+            session_name = f"Session {session_id[:6]}"
+
+        for memory in session_memories:
+            memory["session_name"] = session_name
+
+        return {
+            "session_id": session_id,
+            "session_name": session_name,
+            "memory_count": len(session_memories),
+            "memories": session_memories
+        }
+
+    @router.post("/extract")
+    async def extract_memory(request: Request, session: str = Form(...)) -> Dict[str, List[str]]:
+        """Analyze a session's chat history and return memory suggestions."""
+        require_user(request)
+        try:
+            sess = session_manager.get_session(session)
+        except KeyError:
+            raise HTTPException(404, "Session not found")
+        _assert_session_owner(sess, _owner(request))
+
+        system_msg = {
+            "role": "system",
+            "content": (
+                "You are a helpful assistant. Analyze the entire conversation history provided and extract any "
+                "useful factual statements, contacts, addresses, phone numbers, or other information that the user "
+                "might want to remember for future interactions. Return each piece of information as a JSON object "
+                "with a 'text' field. For example: [{'text': 'Alice lives at 123 Main St'}, {'text': 'Bob works at Acme Corp'}]. "
+                "Only include information that is specific and likely to be useful later."
+            ),
+        }
+        messages = [system_msg] + sess.get_context_messages()
+
+        t_url, t_model, t_headers = resolve_task_endpoint(
+            sess.endpoint_url, sess.model, sess.headers, owner=_owner(request)
+        )
+
+        try:
+            suggestion_text = await llm_call_async(
+                t_url,
+                t_model,
+                messages,
+                temperature=0.2,
+                max_tokens=500,
+                headers=t_headers,
+            )
+            try:
+                suggestions = json.loads(suggestion_text)
+                if isinstance(suggestions, list):
+                    suggestions = [s if isinstance(s, str) else s.get("text", "") for s in suggestions]
+                else:
+                    suggestions = []
+            except json.JSONDecodeError:
+                suggestions = [line.strip() for line in suggestion_text.splitlines() if line.strip()]
+
+            return {"suggestions": [s for s in suggestions if s]}
+        except Exception as e:
+            logger.error(f"LLM memory extraction failed (session {session}): {e}")
+            fallback = memory_manager.extract_memory_from_chat(sess.history, session)
+            return {"suggestions": [item["text"] for item in fallback]}
+
+    @router.post("/audit")
+    async def api_audit_memories(request: Request, session: str = Form(None)):
+        """Deduplicate and consolidate memories via LLM.
+
+        Uses task/utility/default settings through the shared resolver, with
+        the active session as fallback when no task or utility model is set.
+        Returns before and after memory counts.
+        """
+        user = _owner(request)
+        fallback_url = fallback_model = None
+        fallback_headers = None
+        if session:
+            try:
+                sess = session_manager.get_session(session)
+                _assert_session_owner(sess, user)
+                fallback_url = sess.endpoint_url
+                fallback_model = sess.model
+                fallback_headers = sess.headers
+            except KeyError:
+                pass
+
+        endpoint_url, model, headers = resolve_task_endpoint(
+            fallback_url, fallback_model, fallback_headers, owner=user
+        )
+
+        if not endpoint_url or not model:
+            raise HTTPException(400, "No default model configured — set one in Settings")
+
+        result = await audit_memories(
+            memory_manager,
+            memory_vector,
+            endpoint_url,
+            model,
+            headers,
+            owner=user,
+        )
+
+        if "error" in result and "before" not in result:
+            raise HTTPException(502, f"Audit failed: {result['error']}")
+
+        return {
+            "ok": "error" not in result,
+            "before": result.get("before", 0),
+            "after": result.get("after", 0),
+            "removed": result.get("before", 0) - result.get("after", 0),
+            # True when the audit skipped the LLM because nothing changed
+            # since the last tidy. Frontend already says "Already clean"
+            # for removed==0, so this is here for future use / debugging.
+            "already_tidy": bool(result.get("already_tidy")),
+        }
+
+    @router.post("/import")
+    async def import_memories_from_file(
+        request: Request,
+        session: str | None = Form(None),
+        file: UploadFile = File(...)
+    ):
+        """Extract memory suggestions from an uploaded file (PDF, TXT, MD, etc.)."""
+        from src.auth_helpers import require_privilege
+        require_privilege(request, "can_manage_memory")
+
+        endpoint_url = None
+        model = None
+        headers = {}
+
+        user = _owner(request)
+
+        if session:
+            try:
+                sess = session_manager.get_session(session)
+                _assert_session_owner(sess, user)
+            except KeyError:
+                sess = None
+            except HTTPException as exc:
+                if exc.status_code != 404:
+                    raise
+                sess = None
+
+            if sess is None:
+                logger.warning("Session %s not found or inaccessible, falling back to utility endpoint", session)
+                endpoint_url, model, headers = resolve_endpoint("utility", owner=user)
+            else:
+                endpoint_url, model, headers = resolve_task_endpoint(
+                    sess.endpoint_url, sess.model, sess.headers, owner=user
+                )
+        else:
+            endpoint_url, model, headers = resolve_task_endpoint(owner=user)
+    
+        if not endpoint_url or not model:
+            raise HTTPException(400, "No LLM model configured. Set a default model in Settings.")
+
+        content = await read_upload_limited(file, MEMORY_IMPORT_MAX_BYTES, "Memory import")
+        filename = file.filename or "upload"
+        _, ext = os.path.splitext(filename.lower())
+
+        allowed = {".txt", ".md", ".pdf", ".csv", ".log", ".json", ".py", ".js", ".html"}
+        if ext not in allowed:
+            raise HTTPException(400, f"Unsupported file type: {ext}")
+
+        # Extract text based on file type
+        if ext == ".pdf":
+            from src.document_processor import _process_pdf
+            with tempfile.NamedTemporaryFile(suffix=".pdf", delete=False) as tmp:
+                tmp.write(content)
+                tmp_path = tmp.name
+            try:
+                text = _process_pdf(tmp_path, owner=_owner(request))
+            finally:
+                os.unlink(tmp_path)
+        else:
+            try:
+                text = content.decode("utf-8")
+            except UnicodeDecodeError:
+                from charset_normalizer import detect
+                encoding = (detect(content) or {}).get("encoding") or "utf-8"
+                text = content.decode(encoding, errors="replace")
+
+        if not text.strip():
+            return {"suggestions": [], "message": "No readable content found"}
+
+        # Fast path: a .json upload that already looks like a memories export
+        # (list of {text, category, ...} dicts, or list of strings) round-trips
+        # directly without spending an LLM call to re-extract its own output.
+        # Without this, re-importing a memories.json from another account
+        # ran the file through the extractor, which often re-emitted the
+        # entries as a numbered list (and the numbering leaked into the
+        # `text` field).
+        if ext == ".json":
+            try:
+                parsed = json.loads(text)
+            except json.JSONDecodeError:
+                parsed = None
+            if isinstance(parsed, list) and parsed:
+                direct = []
+                for item in parsed:
+                    if isinstance(item, dict) and item.get("text"):
+                        direct.append({
+                            "text": _strip_list_prefix(str(item["text"])),
+                            "category": item.get("category") or "fact",
+                        })
+                    elif isinstance(item, str) and item.strip():
+                        direct.append({
+                            "text": _strip_list_prefix(item.strip()),
+                            "category": "fact",
+                        })
+                if direct:
+                    return {"suggestions": direct, "filename": filename}
+
+        # Truncate very long documents
+        if len(text) > 15000:
+            text = text[:15000] + "\n[Truncated]"
+
+        # Send to LLM for memory extraction
+        import_prompt = (
+            "You are a memory extraction assistant. The user uploaded a document. "
+            "Analyze the text below and extract specific, useful facts — things like "
+            "names, preferences, jobs, locations, relationships, opinions, projects, "
+            "goals, contacts, or any other personal details worth remembering.\n\n"
+            "Rules:\n"
+            "- Each fact should be a short, self-contained statement\n"
+            "- Do NOT extract generic knowledge\n"
+            "- Focus on personal, memorable information\n"
+            "- If there are no useful facts, return an empty array\n\n"
+            "Return a JSON array of objects with 'text' and 'category' fields.\n"
+            "Categories: 'identity', 'preference', 'fact', 'contact', 'project', 'goal'\n\n"
+            "Return ONLY valid JSON, no markdown fences."
+        )
+
+        try:
+            raw = await llm_call_async(
+                endpoint_url,
+                model,
+                [
+                    {"role": "system", "content": import_prompt},
+                    {"role": "user", "content": f"Document: {filename}\n\n{text}"},
+                ],
+                temperature=0.2,
+                max_tokens=2000,
+                headers=headers,
+            )
+
+            # Parse JSON
+            raw = raw.strip()
+            if raw.startswith("```"):
+                raw = raw.split("\n", 1)[-1].rsplit("```", 1)[0].strip()
+
+            suggestions = json.loads(raw)
+            if isinstance(suggestions, list):
+                normalized = []
+                for s in suggestions:
+                    if not s:
+                        continue
+                    if isinstance(s, dict):
+                        s = dict(s)
+                        if s.get("text"):
+                            s["text"] = _strip_list_prefix(str(s["text"]))
+                        normalized.append(s)
+                    else:
+                        normalized.append({"text": _strip_list_prefix(str(s)), "category": "fact"})
+                suggestions = normalized
+            else:
+                suggestions = []
+
+            return {"suggestions": suggestions, "filename": filename}
+
+        except json.JSONDecodeError:
+            # Fallback: split by lines, stripping any "1.", "2)" markdown-list
+            # numbering the model added so saved memories don't keep the prefix.
+            lines = [_strip_list_prefix(l.strip()) for l in raw.splitlines() if l.strip() and len(l.strip()) > 5]
+            return {"suggestions": [{"text": l, "category": "fact"} for l in lines[:20]], "filename": filename}
+        except Exception as e:
+            logger.error(f"Memory import extraction failed: {e}")
+            raise HTTPException(502, f"LLM extraction failed: {str(e)}")
+
+    @router.post("/{memory_id}/pin")
+    def pin_memory(request: Request, memory_id: str, pinned: bool = Form(True)):
+        """Pin or unpin a memory. Pinned memories are always included in context."""
+        user = _owner(request)
+        all_mem = memory_manager.load_all()
+        for i, memory in enumerate(all_mem):
+            if memory["id"] == memory_id:
+                _verify_memory_owner(memory, user)
+                all_mem[i]["pinned"] = pinned
+                memory_manager.save(all_mem)
+                return {"ok": True, "pinned": pinned}
+        raise HTTPException(404, f"Memory item {memory_id} not found")
+
+    # Wildcard routes MUST come last — otherwise they swallow /import, /search, etc.
+    @router.get("/{memory_id}")
+    def get_memory_item(request: Request, memory_id: str):
+        """Get a specific memory item by ID."""
+        user = _owner(request)
+        memories = memory_manager.load(owner=user)
+        for memory in memories:
+            if memory["id"] == memory_id:
+                return {"memory": memory}
+
+        raise HTTPException(404, "Memory not found")
+
+    @router.put("/{memory_id}")
+    def update_memory(request: Request, memory_id: str, text: str = Form(...), category: str = Form(None)):
+        """Update an existing memory item with new text and optional category."""
+        user = _owner(request)
+        all_mem = memory_manager.load_all()
+        for i, memory in enumerate(all_mem):
+            if memory["id"] == memory_id:
+                _verify_memory_owner(memory, user)
+                all_mem[i]["text"] = text.strip()
+                if category:
+                    all_mem[i]["category"] = category
+                all_mem[i]["timestamp"] = int(time.time())
+
+                memory_manager.save(all_mem)
+                # Sync vector index (remove old, add updated)
+                if memory_vector and memory_vector.healthy:
+                    memory_vector.remove(memory_id)
+                    memory_vector.add(memory_id, text.strip())
+                return {"ok": True, "message": "Memory updated successfully"}
+
+        raise HTTPException(404, f"Memory item {memory_id} not found")
+
+    @router.delete("/{memory_id}")
+    def delete_memory(request: Request, memory_id: str):
+        """Delete a memory item by its ID."""
+        user = _owner(request)
+        all_mem = memory_manager.load_all()
+
+        # Find and verify ownership before deleting
+        target = next((m for m in all_mem if m["id"] == memory_id), None)
+        if not target:
+            raise HTTPException(404, f"Memory item {memory_id} not found")
+        _verify_memory_owner(target, user)
+
+        all_mem = [m for m in all_mem if m["id"] != memory_id]
+        memory_manager.save(all_mem)
+        # Sync vector index
+        if memory_vector and memory_vector.healthy:
+            memory_vector.remove(memory_id)
+        return {"ok": True, "message": "Memory deleted successfully"}
+
+    return router

routes/memory_routes.py

@@ -1,552 +1,18 @@
-# routes/memory_routes.py
-from fastapi import APIRouter, Form, HTTPException, Request, UploadFile, File
-from typing import Dict, Any, Optional, List
-import json
-import os
-import re
-import tempfile
-import time
-from datetime import datetime
-import logging
+"""Backward-compat shim — canonical location is routes/memory/memory_routes.py.
 
-# Leading list-marker like "1.", "12)", or "3:" plus surrounding whitespace.
-# Strips one prefix per call so import-from-LLM-output doesn't leave the
-# numbering inside the saved memory text. Bullet markers (-, *, •) are
-# also peeled here for the same reason.
-_LIST_PREFIX_RE = re.compile(r"^\s*(?:\d{1,3}[.):]\s+|[-*•]\s+)")
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.memory_routes``, ``from routes.memory_routes import X``,
+``importlib.import_module("routes.memory_routes")``, and
+``monkeypatch.setattr(routes.memory_routes, "ATTR", ...)`` (used by
+test_memory_routes_session_owner.py and test_memory_owner_isolation.py via
+``import ... as mr`` + ``setattr(mr, ...)``) all operate on the *same* object
+the application actually uses. Keeps existing import paths working after
+slice 2c (#4082/#4071). Source-introspection tests read the canonical file
+by path.
+"""
 
+import sys as _sys
 
-def _strip_list_prefix(text: str) -> str:
-    if not text:
-        return text
-    return _LIST_PREFIX_RE.sub("", text, count=1).strip()
+from routes.memory import memory_routes as _canonical  # noqa: F401
 
-from services.memory import MemoryManager
-from core.session_manager import SessionManager
-from src.request_models import MemoryAddRequest
-from core.database import SessionLocal
-from src.llm_core import llm_call_async
-from services.memory.memory_extractor import audit_memories
-from src.auth_helpers import get_current_user, require_user
-from src.endpoint_resolver import resolve_endpoint
-from src.task_endpoint import resolve_task_endpoint
-from src.upload_limits import read_upload_limited, MEMORY_IMPORT_MAX_BYTES
-
-logger = logging.getLogger(__name__)
-
-
-def setup_memory_routes(memory_manager: MemoryManager, session_manager: SessionManager, memory_vector=None):
-    """Set up memory-related routes."""
-    router = APIRouter(prefix="/api/memory", tags=["memory"])
-
-    def _owner(request: Request) -> Optional[str]:
-        return get_current_user(request)
-
-    def _assert_session_owner(session_obj, user):
-        """SECURITY: 404 if the caller does not own this session.
-
-        SessionManager.get_session is NOT owner-scoped — it returns any
-        session by id. These routes accept a caller-supplied session id, so
-        without this gate a user could target another tenant's session and
-        leak their chat history, their session-scoped LLM credentials, or the
-        session title. Mirrors session_routes / webhook_routes ownership.
-        """
-        if user is not None and getattr(session_obj, "owner", None) != user:
-            raise HTTPException(404, "Session not found")
-
-    def _verify_memory_owner(memory: dict, user: Optional[str]):
-        """Raise 404 if user doesn't own this memory.
-
-        SECURITY: strict ownership — previously `mem_owner and mem_owner != user`
-        allowed any user to read/edit/delete memories with an empty/null owner
-        field, which leaked legacy data across the multi-user deploy.
-        """
-        if user is None:
-            return  # Auth disabled
-        if memory.get("owner") != user:
-            raise HTTPException(404, "Memory not found")
-
-    @router.post("/debug")
-    def debug_memory_relevance(request: Request, query: str = Form(...)):
-        """Debug which memories would be triggered for a query"""
-        user = _owner(request)
-        memories = memory_manager.load(owner=user)
-        relevant = memory_manager.get_relevant_memories(query, memories, threshold=0.05)
-
-        return {
-            "query": query,
-            "total_memories": len(memories),
-            "relevant_count": len(relevant),
-            "relevant_memories": [{"text": m["text"], "category": m.get("category", "unknown")}
-                                 for m in relevant]
-        }
-
-    @router.post("/add", response_model=Dict[str, Any])
-    async def api_add_memory(
-        request: Request,
-        memory_data: Optional[MemoryAddRequest] = None
-    ):
-        """Add a new memory entry with optional category, source, and session reference."""
-        from src.auth_helpers import require_privilege
-        require_privilege(request, "can_manage_memory")
-        if memory_data is None:
-            form = await request.form()
-            memory_data = MemoryAddRequest(
-                text=form.get("text"),
-                category=form.get("category", "fact"),
-                source=form.get("source", "user"),
-                session_id=form.get("session_id")
-            )
-
-        user = _owner(request)
-        text = (memory_data.text or "").strip()
-        if not text:
-            raise HTTPException(400, "empty memory")
-        user_mem = memory_manager.load(owner=user)
-        if memory_manager.find_duplicates(text, user_mem):
-            return {"ok": True, "count": len(user_mem), "message": "Memory already exists"}
-
-        if memory_data.session_id:
-            try:
-                session_obj = session_manager.get_session(memory_data.session_id)
-            except KeyError:
-                raise HTTPException(404, "Session not found")
-            _assert_session_owner(session_obj, user)
-
-        new_entry = memory_manager.add_entry(text, memory_data.source, memory_data.category, owner=user)
-        if memory_data.session_id:
-            new_entry["session_id"] = memory_data.session_id
-        all_mem = memory_manager.load_all()
-        all_mem.append(new_entry)
-        memory_manager.save(all_mem)
-        # Sync vector index
-        if memory_vector and memory_vector.healthy:
-            memory_vector.add(new_entry["id"], text)
-        try:
-            from src.event_bus import fire_event
-            fire_event("memory_added", user)
-        except Exception:
-            logger.debug("memory_added event dispatch failed", exc_info=True)
-        return {"ok": True, "count": len([m for m in all_mem if m.get("owner") == user])}
-
-    @router.get("")
-    def api_get_memory(request: Request):
-        """Return all memory entries with their metadata."""
-        user = _owner(request)
-        return {"memory": memory_manager.load(owner=user)}
-
-    @router.post("/search")
-    def search_memories(request: Request, query: str = Form(...), session_id: str = Form(None), category: str = Form(None)):
-        """Search across all memories with optional filters."""
-        user = _owner(request)
-        memories = memory_manager.load(owner=user)
-
-        if session_id:
-            memories = [m for m in memories if m.get("session_id") == session_id]
-
-        if category:
-            memories = [m for m in memories if category in m.get("categories", [m.get("category", "")])]
-
-        relevant = memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
-
-        return {"memories": relevant, "total": len(relevant), "query": query}
-
-    @router.get("/timeline")
-    def memory_timeline(request: Request):
-        """Get memories in chronological order with source session information."""
-        user = _owner(request)
-        memories = memory_manager.load(owner=user)
-        sorted_memories = sorted(memories, key=lambda x: x.get("timestamp", 0), reverse=True)
-
-        results = []
-        for memory in sorted_memories:
-            if "timestamp" in memory:
-                try:
-                    dt = datetime.fromtimestamp(memory["timestamp"])
-                    memory["timestamp_str"] = dt.strftime("%Y-%m-%d %H:%M:%S")
-                except (ValueError, OSError, OverflowError):
-                    memory["timestamp_str"] = "Unknown"
-            else:
-                memory["timestamp_str"] = "Unknown"
-
-            session_id = memory.get("session_id")
-            if session_id and session_id in session_manager.sessions:
-                try:
-                    session = session_manager.get_session(session_id)
-                    if session:
-                        _assert_session_owner(session, user)
-                    memory["session_name"] = session.name if session else f"Session {session_id[:6]}"
-                except KeyError:
-                    memory["session_name"] = "Unknown"
-                except HTTPException as exc:
-                    if exc.status_code != 404:
-                        raise
-                    memory["session_name"] = "Unknown"
-            else:
-                memory["session_name"] = "Unknown"
-
-            results.append(memory)
-
-        return {"timeline": results, "total": len(results)}
-
-    @router.get("/by-session/{session_id}")
-    def get_memory_by_session(request: Request, session_id: str):
-        """Get all memories associated with a specific session."""
-        user = _owner(request)
-        try:
-            _session_obj = session_manager.get_session(session_id)
-        except KeyError:
-            raise HTTPException(404, f"Session {session_id} not found")
-        _assert_session_owner(_session_obj, user)
-        memories = memory_manager.load(owner=user)
-        session_memories = [m for m in memories if m.get("session_id") == session_id]
-
-        session_memories.sort(key=lambda x: x.get("timestamp", 0), reverse=True)
-
-        try:
-            session = session_manager.get_session(session_id)
-            session_name = session.name if session else f"Session {session_id[:6]}"
-        except KeyError:
-            session_name = f"Session {session_id[:6]}"
-
-        for memory in session_memories:
-            memory["session_name"] = session_name
-
-        return {
-            "session_id": session_id,
-            "session_name": session_name,
-            "memory_count": len(session_memories),
-            "memories": session_memories
-        }
-
-    @router.post("/extract")
-    async def extract_memory(request: Request, session: str = Form(...)) -> Dict[str, List[str]]:
-        """Analyze a session's chat history and return memory suggestions."""
-        require_user(request)
-        try:
-            sess = session_manager.get_session(session)
-        except KeyError:
-            raise HTTPException(404, "Session not found")
-        _assert_session_owner(sess, _owner(request))
-
-        system_msg = {
-            "role": "system",
-            "content": (
-                "You are a helpful assistant. Analyze the entire conversation history provided and extract any "
-                "useful factual statements, contacts, addresses, phone numbers, or other information that the user "
-                "might want to remember for future interactions. Return each piece of information as a JSON object "
-                "with a 'text' field. For example: [{'text': 'Alice lives at 123 Main St'}, {'text': 'Bob works at Acme Corp'}]. "
-                "Only include information that is specific and likely to be useful later."
-            ),
-        }
-        messages = [system_msg] + sess.get_context_messages()
-
-        t_url, t_model, t_headers = resolve_task_endpoint(
-            sess.endpoint_url, sess.model, sess.headers, owner=_owner(request)
-        )
-
-        try:
-            suggestion_text = await llm_call_async(
-                t_url,
-                t_model,
-                messages,
-                temperature=0.2,
-                max_tokens=500,
-                headers=t_headers,
-            )
-            try:
-                suggestions = json.loads(suggestion_text)
-                if isinstance(suggestions, list):
-                    suggestions = [s if isinstance(s, str) else s.get("text", "") for s in suggestions]
-                else:
-                    suggestions = []
-            except json.JSONDecodeError:
-                suggestions = [line.strip() for line in suggestion_text.splitlines() if line.strip()]
-
-            return {"suggestions": [s for s in suggestions if s]}
-        except Exception as e:
-            logger.error(f"LLM memory extraction failed (session {session}): {e}")
-            fallback = memory_manager.extract_memory_from_chat(sess.history, session)
-            return {"suggestions": [item["text"] for item in fallback]}
-
-    @router.post("/audit")
-    async def api_audit_memories(request: Request, session: str = Form(None)):
-        """Deduplicate and consolidate memories via LLM.
-
-        Uses task/utility/default settings through the shared resolver, with
-        the active session as fallback when no task or utility model is set.
-        Returns before and after memory counts.
-        """
-        user = _owner(request)
-        fallback_url = fallback_model = None
-        fallback_headers = None
-        if session:
-            try:
-                sess = session_manager.get_session(session)
-                _assert_session_owner(sess, user)
-                fallback_url = sess.endpoint_url
-                fallback_model = sess.model
-                fallback_headers = sess.headers
-            except KeyError:
-                pass
-
-        endpoint_url, model, headers = resolve_task_endpoint(
-            fallback_url, fallback_model, fallback_headers, owner=user
-        )
-
-        if not endpoint_url or not model:
-            raise HTTPException(400, "No default model configured — set one in Settings")
-
-        result = await audit_memories(
-            memory_manager,
-            memory_vector,
-            endpoint_url,
-            model,
-            headers,
-            owner=user,
-        )
-
-        if "error" in result and "before" not in result:
-            raise HTTPException(502, f"Audit failed: {result['error']}")
-
-        return {
-            "ok": "error" not in result,
-            "before": result.get("before", 0),
-            "after": result.get("after", 0),
-            "removed": result.get("before", 0) - result.get("after", 0),
-            # True when the audit skipped the LLM because nothing changed
-            # since the last tidy. Frontend already says "Already clean"
-            # for removed==0, so this is here for future use / debugging.
-            "already_tidy": bool(result.get("already_tidy")),
-        }
-
-    @router.post("/import")
-    async def import_memories_from_file(
-        request: Request,
-        session: str | None = Form(None),
-        file: UploadFile = File(...)
-    ):
-        """Extract memory suggestions from an uploaded file (PDF, TXT, MD, etc.)."""
-        from src.auth_helpers import require_privilege
-        require_privilege(request, "can_manage_memory")
-
-        endpoint_url = None
-        model = None
-        headers = {}
-
-        user = _owner(request)
-
-        if session:
-            try:
-                sess = session_manager.get_session(session)
-                _assert_session_owner(sess, user)
-            except KeyError:
-                sess = None
-            except HTTPException as exc:
-                if exc.status_code != 404:
-                    raise
-                sess = None
-
-            if sess is None:
-                logger.warning("Session %s not found or inaccessible, falling back to utility endpoint", session)
-                endpoint_url, model, headers = resolve_endpoint("utility", owner=user)
-            else:
-                endpoint_url, model, headers = resolve_task_endpoint(
-                    sess.endpoint_url, sess.model, sess.headers, owner=user
-                )
-        else:
-            endpoint_url, model, headers = resolve_task_endpoint(owner=user)
-    
-        if not endpoint_url or not model:
-            raise HTTPException(400, "No LLM model configured. Set a default model in Settings.")
-
-        content = await read_upload_limited(file, MEMORY_IMPORT_MAX_BYTES, "Memory import")
-        filename = file.filename or "upload"
-        _, ext = os.path.splitext(filename.lower())
-
-        allowed = {".txt", ".md", ".pdf", ".csv", ".log", ".json", ".py", ".js", ".html"}
-        if ext not in allowed:
-            raise HTTPException(400, f"Unsupported file type: {ext}")
-
-        # Extract text based on file type
-        if ext == ".pdf":
-            from src.document_processor import _process_pdf
-            with tempfile.NamedTemporaryFile(suffix=".pdf", delete=False) as tmp:
-                tmp.write(content)
-                tmp_path = tmp.name
-            try:
-                text = _process_pdf(tmp_path, owner=_owner(request))
-            finally:
-                os.unlink(tmp_path)
-        else:
-            try:
-                text = content.decode("utf-8")
-            except UnicodeDecodeError:
-                from charset_normalizer import detect
-                encoding = (detect(content) or {}).get("encoding") or "utf-8"
-                text = content.decode(encoding, errors="replace")
-
-        if not text.strip():
-            return {"suggestions": [], "message": "No readable content found"}
-
-        # Fast path: a .json upload that already looks like a memories export
-        # (list of {text, category, ...} dicts, or list of strings) round-trips
-        # directly without spending an LLM call to re-extract its own output.
-        # Without this, re-importing a memories.json from another account
-        # ran the file through the extractor, which often re-emitted the
-        # entries as a numbered list (and the numbering leaked into the
-        # `text` field).
-        if ext == ".json":
-            try:
-                parsed = json.loads(text)
-            except json.JSONDecodeError:
-                parsed = None
-            if isinstance(parsed, list) and parsed:
-                direct = []
-                for item in parsed:
-                    if isinstance(item, dict) and item.get("text"):
-                        direct.append({
-                            "text": _strip_list_prefix(str(item["text"])),
-                            "category": item.get("category") or "fact",
-                        })
-                    elif isinstance(item, str) and item.strip():
-                        direct.append({
-                            "text": _strip_list_prefix(item.strip()),
-                            "category": "fact",
-                        })
-                if direct:
-                    return {"suggestions": direct, "filename": filename}
-
-        # Truncate very long documents
-        if len(text) > 15000:
-            text = text[:15000] + "\n[Truncated]"
-
-        # Send to LLM for memory extraction
-        import_prompt = (
-            "You are a memory extraction assistant. The user uploaded a document. "
-            "Analyze the text below and extract specific, useful facts — things like "
-            "names, preferences, jobs, locations, relationships, opinions, projects, "
-            "goals, contacts, or any other personal details worth remembering.\n\n"
-            "Rules:\n"
-            "- Each fact should be a short, self-contained statement\n"
-            "- Do NOT extract generic knowledge\n"
-            "- Focus on personal, memorable information\n"
-            "- If there are no useful facts, return an empty array\n\n"
-            "Return a JSON array of objects with 'text' and 'category' fields.\n"
-            "Categories: 'identity', 'preference', 'fact', 'contact', 'project', 'goal'\n\n"
-            "Return ONLY valid JSON, no markdown fences."
-        )
-
-        try:
-            raw = await llm_call_async(
-                endpoint_url,
-                model,
-                [
-                    {"role": "system", "content": import_prompt},
-                    {"role": "user", "content": f"Document: {filename}\n\n{text}"},
-                ],
-                temperature=0.2,
-                max_tokens=2000,
-                headers=headers,
-            )
-
-            # Parse JSON
-            raw = raw.strip()
-            if raw.startswith("```"):
-                raw = raw.split("\n", 1)[-1].rsplit("```", 1)[0].strip()
-
-            suggestions = json.loads(raw)
-            if isinstance(suggestions, list):
-                normalized = []
-                for s in suggestions:
-                    if not s:
-                        continue
-                    if isinstance(s, dict):
-                        s = dict(s)
-                        if s.get("text"):
-                            s["text"] = _strip_list_prefix(str(s["text"]))
-                        normalized.append(s)
-                    else:
-                        normalized.append({"text": _strip_list_prefix(str(s)), "category": "fact"})
-                suggestions = normalized
-            else:
-                suggestions = []
-
-            return {"suggestions": suggestions, "filename": filename}
-
-        except json.JSONDecodeError:
-            # Fallback: split by lines, stripping any "1.", "2)" markdown-list
-            # numbering the model added so saved memories don't keep the prefix.
-            lines = [_strip_list_prefix(l.strip()) for l in raw.splitlines() if l.strip() and len(l.strip()) > 5]
-            return {"suggestions": [{"text": l, "category": "fact"} for l in lines[:20]], "filename": filename}
-        except Exception as e:
-            logger.error(f"Memory import extraction failed: {e}")
-            raise HTTPException(502, f"LLM extraction failed: {str(e)}")
-
-    @router.post("/{memory_id}/pin")
-    def pin_memory(request: Request, memory_id: str, pinned: bool = Form(True)):
-        """Pin or unpin a memory. Pinned memories are always included in context."""
-        user = _owner(request)
-        all_mem = memory_manager.load_all()
-        for i, memory in enumerate(all_mem):
-            if memory["id"] == memory_id:
-                _verify_memory_owner(memory, user)
-                all_mem[i]["pinned"] = pinned
-                memory_manager.save(all_mem)
-                return {"ok": True, "pinned": pinned}
-        raise HTTPException(404, f"Memory item {memory_id} not found")
-
-    # Wildcard routes MUST come last — otherwise they swallow /import, /search, etc.
-    @router.get("/{memory_id}")
-    def get_memory_item(request: Request, memory_id: str):
-        """Get a specific memory item by ID."""
-        user = _owner(request)
-        memories = memory_manager.load(owner=user)
-        for memory in memories:
-            if memory["id"] == memory_id:
-                return {"memory": memory}
-
-        raise HTTPException(404, "Memory not found")
-
-    @router.put("/{memory_id}")
-    def update_memory(request: Request, memory_id: str, text: str = Form(...), category: str = Form(None)):
-        """Update an existing memory item with new text and optional category."""
-        user = _owner(request)
-        all_mem = memory_manager.load_all()
-        for i, memory in enumerate(all_mem):
-            if memory["id"] == memory_id:
-                _verify_memory_owner(memory, user)
-                all_mem[i]["text"] = text.strip()
-                if category:
-                    all_mem[i]["category"] = category
-                all_mem[i]["timestamp"] = int(time.time())
-
-                memory_manager.save(all_mem)
-                # Sync vector index (remove old, add updated)
-                if memory_vector and memory_vector.healthy:
-                    memory_vector.remove(memory_id)
-                    memory_vector.add(memory_id, text.strip())
-                return {"ok": True, "message": "Memory updated successfully"}
-
-        raise HTTPException(404, f"Memory item {memory_id} not found")
-
-    @router.delete("/{memory_id}")
-    def delete_memory(request: Request, memory_id: str):
-        """Delete a memory item by its ID."""
-        user = _owner(request)
-        all_mem = memory_manager.load_all()
-
-        # Find and verify ownership before deleting
-        target = next((m for m in all_mem if m["id"] == memory_id), None)
-        if not target:
-            raise HTTPException(404, f"Memory item {memory_id} not found")
-        _verify_memory_owner(target, user)
-
-        all_mem = [m for m in all_mem if m["id"] != memory_id]
-        memory_manager.save(all_mem)
-        # Sync vector index
-        if memory_vector and memory_vector.healthy:
-            memory_vector.remove(memory_id)
-        return {"ok": True, "message": "Memory deleted successfully"}
-
-    return router
+_sys.modules[__name__] = _canonical

routes/model_routes.py

@@ -19,6 +19,7 @@ from fastapi.responses import StreamingResponse
 from core.database import SessionLocal, ModelEndpoint, Session as DbSession
 from core.log_safety import redact_url as _redact_url_for_log
 from core.middleware import require_admin
+from src.constants import COOKBOOK_STATE_FILE
 from src.llm_core import _detect_provider, _host_match, ANTHROPIC_MODELS
 from src.tls_overrides import llm_verify
 from src.settings import load_settings as _load_settings, save_settings as _save_settings
@@ -112,6 +113,67 @@ def _clear_endpoint_settings_for_endpoint(settings: dict, ep_id: str, *, include
     return cleared
 
 
+_COOKBOOK_ACTIVE_SERVE_STATUSES = {
+    "starting", "loading", "ready", "running", "restarting",
+}
+
+
+def _active_cookbook_endpoint_ids() -> set[str]:
+    """Endpoint IDs owned by active Cookbook serve tasks.
+
+    Cookbook auto-registers endpoints with ids like ``local-*``. Those rows are
+    managed lifecycle state, not durable user configuration. If a tmux stream is
+    stopped or an old task lingers, the row must stop participating in model
+    selection and defaults.
+    """
+    try:
+        if not os.path.exists(COOKBOOK_STATE_FILE):
+            return set()
+        with open(COOKBOOK_STATE_FILE, "r", encoding="utf-8") as fh:
+            raw = fh.read()
+        state = json.loads(raw)
+    except Exception:
+        return set()
+    out: set[str] = set()
+    for task in state.get("tasks") or []:
+        if not isinstance(task, dict) or task.get("type") != "serve":
+            continue
+        if str(task.get("status") or "").lower() not in _COOKBOOK_ACTIVE_SERVE_STATUSES:
+            continue
+        ep_id = task.get("_endpointId") or task.get("endpointId") or task.get("endpoint_id")
+        if ep_id:
+            out.add(str(ep_id))
+    return out
+
+
+def _disable_stale_cookbook_local_endpoints(db) -> int:
+    """Disable enabled cookbook endpoints whose serve task is no longer active."""
+    active_ids = _active_cookbook_endpoint_ids()
+    if not active_ids:
+        return 0
+    stale = (
+        db.query(ModelEndpoint)
+        .filter(ModelEndpoint.is_enabled == True)  # noqa: E712
+        .filter(ModelEndpoint.id.like("local-%"))
+        .filter(~ModelEndpoint.id.in_(active_ids))
+        .all()
+    )
+    if not stale:
+        return 0
+    settings = _load_settings()
+    touched_settings = False
+    for ep in stale:
+        ep.is_enabled = False
+        ep.model_refresh_mode = "disabled"
+        if _clear_endpoint_settings_for_endpoint(settings, ep.id):
+            touched_settings = True
+        logger.info("Disabled stale Cookbook endpoint %s (%s @ %s)", ep.id, ep.name, ep.base_url)
+    if touched_settings:
+        _save_settings(settings)
+    db.commit()
+    return len(stale)
+
+
 def _clear_user_pref_endpoint_refs(all_prefs: dict, ep_id: str) -> int:
     """Remove endpoint references from scoped or legacy-flat user preferences."""
     if not isinstance(all_prefs, dict):
@@ -125,7 +187,24 @@ def _clear_user_pref_endpoint_refs(all_prefs: dict, ep_id: str) -> int:
     return cleared_users
 
 
-def _default_endpoint_needs_assignment(current_default_id: str, enabled_endpoint_ids) -> bool:
+def _endpoint_visible_model_ids(ep: Any) -> List[str]:
+    """Known visible model ids for an endpoint, including pinned/manual ids."""
+    if ep is None:
+        return []
+    return _visible_models(
+        getattr(ep, "cached_models", None),
+        getattr(ep, "hidden_models", None),
+        getattr(ep, "pinned_models", None),
+    )
+
+
+def _default_endpoint_needs_assignment(
+    current_default_id: str,
+    enabled_endpoint_ids,
+    *,
+    current_default_endpoint: Any = None,
+    current_default_model: str = "",
+) -> bool:
     """Whether the global default chat endpoint should be (re)assigned.
 
     True when nothing is configured yet, or the configured default no longer
@@ -137,7 +216,14 @@ def _default_endpoint_needs_assignment(current_default_id: str, enabled_endpoint
     """
     if not current_default_id:
         return True
-    return current_default_id not in enabled_endpoint_ids
+    if current_default_id not in enabled_endpoint_ids:
+        return True
+    if current_default_endpoint is None:
+        return False
+    if not (current_default_model or "").strip():
+        return True
+    visible = _endpoint_visible_model_ids(current_default_endpoint)
+    return bool(visible and current_default_model not in visible)
 
 
 # Loopback hosts a user might type for a local model server (LM Studio,
@@ -1066,6 +1152,36 @@ def _merge_model_ids(*lists):
     return out
 
 
+def _is_mlx_deepseek_v4_repo_id(model_id: str) -> bool:
+    m = str(model_id or "").lower()
+    return "mlx-community/deepseek-v4" in m
+
+
+def _is_mlx_deepseek_v4_shim_id(model_id: str) -> bool:
+    m = str(model_id or "").lower()
+    return "/.cache/odysseus/mlx-shims/deepseek-v4" in m
+
+
+def _filter_mlx_deepseek_v4_repo_when_shimmed(model_ids):
+    """Hide the broken MLX repo id when a launch-specific shim id is available.
+
+    mlx_lm.server may advertise the original HF repo id even though generation
+    only works through Odysseus' sanitized local shim. Keep the shim as the
+    submitted model id and remove the raw repo id from the picker/default list.
+    """
+    ids = list(model_ids or [])
+    has_shim = any(_is_mlx_deepseek_v4_shim_id(m) for m in ids)
+    if not has_shim:
+        return ids
+    return [m for m in ids if not _is_mlx_deepseek_v4_repo_id(m)]
+
+
+def _model_display_name(model_id: str) -> str:
+    if _is_mlx_deepseek_v4_shim_id(model_id):
+        return str(model_id or "").rstrip("/").split("/")[-1] or "DeepSeek-V4-Flash-4bit"
+    return str(model_id or "").split("/")[-1]
+
+
 def _visible_models(cached_models, hidden_models, pinned_models=None):
     """Merge cached + pinned model IDs, then filter out hidden ones.
 
@@ -1079,6 +1195,7 @@ def _visible_models(cached_models, hidden_models, pinned_models=None):
         _normalize_model_ids(cached_models),
         _normalize_model_ids(pinned_models),
     )
+    merged = _filter_mlx_deepseek_v4_repo_when_shimmed(merged)
     if not hidden_models:
         return merged
     hidden = set(_normalize_model_ids(hidden_models))
@@ -1194,6 +1311,8 @@ def setup_model_routes(model_discovery):
                 db = SessionLocal()
                 changed = False
                 try:
+                    if _disable_stale_cookbook_local_endpoints(db):
+                        changed = True
                     endpoints = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True).all()
                     now = _time.time()
                     groups: Dict[str, Dict[str, Any]] = {}
@@ -1267,6 +1386,8 @@ def setup_model_routes(model_discovery):
 
         db = SessionLocal()
         try:
+            if _disable_stale_cookbook_local_endpoints(db):
+                _invalidate_models_cache()
             q = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)
             if owner and not is_admin:
                 # Regular users see: their own endpoints + null-owner
@@ -1306,9 +1427,9 @@ def setup_model_routes(model_discovery):
                     "port": 0,
                     "url": chat_url,
                     "models": curated,
-                    "models_display": [mid.split("/")[-1] for mid in curated],
+                    "models_display": [_model_display_name(mid) for mid in curated],
                     "models_extra": extra,
-                    "models_extra_display": [mid.split("/")[-1] for mid in extra],
+                    "models_extra_display": [_model_display_name(mid) for mid in extra],
                     "endpoint_id": ep.id,
                     "endpoint_name": ep.name,
                     "category": category,
@@ -1336,7 +1457,7 @@ def setup_model_routes(model_discovery):
         return {"hosts": [], "items": items}
 
     @router.get("/models")
-    def api_models(request: Request, refresh: bool = False):
+    def api_models(request: Request, refresh: bool = False, background: bool = False):
         """Get available models — per-user (caller sees only their endpoints +
         legacy/shared null-owner rows). Cached per-user for 30s."""
         # Require auth; "" is the unconfigured single-user mode, treated as
@@ -1378,8 +1499,11 @@ def setup_model_routes(model_discovery):
             return cache_entry["data"]
         result = _fetch_models(owner=owner, is_admin=_is_admin)
         _models_cache[_cache_key] = {"data": result, "time": now}
-        # Kick off background refresh to update caches from live endpoints
-        _refresh_caches_bg(force=refresh)
+        # Kick off background refresh to update caches from live endpoints.
+        # Page boot can opt out with background=false so opening Odysseus does
+        # not start endpoint probes against slow/offline model servers.
+        if background or refresh:
+            _refresh_caches_bg(force=refresh)
         return result
 
     # Brief cache for local-probe results so picker-open doesn't hammer
@@ -1388,6 +1512,7 @@ def setup_model_routes(model_discovery):
     # within ~8s of the user noticing.
     _LOCAL_PROBE_TTL = 8.0
     _local_probe_cache: Dict[str, Any] = {"data": None, "time": 0.0}
+    _local_probe_inflight: Dict[str, Any] = {"task": None}
 
     @router.get("/model-endpoints/probe-local")
     async def probe_local_endpoints(request: Request):
@@ -1402,58 +1527,72 @@ def setup_model_routes(model_discovery):
                 (now - _local_probe_cache["time"]) < _LOCAL_PROBE_TTL):
             return _local_probe_cache["data"]
 
-        db = SessionLocal()
-        try:
-            endpoints = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True).all()
-            local_eps = []
-            for ep in endpoints:
-                base = _normalize_base(ep.base_url)
-                kind = _effective_endpoint_kind(ep, base)
-                if _classify_endpoint(base, kind) == "local":
-                    local_eps.append((ep.id, base, ep.api_key))
-        finally:
-            db.close()
-
-        grouped: Dict[str, Dict[str, Any]] = {}
-        for ep_id, base, api_key in local_eps:
-            key = _refresh_key(base, api_key)
-            grouped.setdefault(key, {"base": base, "api_key": api_key, "endpoint_ids": []})["endpoint_ids"].append(ep_id)
-
-        async def _probe_one(data: Dict[str, Any]) -> Dict[str, Any]:
-            t0 = _time.time()
-            try:
-                import asyncio as _asyncio
-                # Bumped 1.5s → 3.5s. The previous 1.5s budget was clipping
-                # local vLLM endpoints on Tailscale links where the model
-                # server is still loading (Qwen3.5-122B takes 2–3 min to
-                # warm); /v1/models can take 500–2500 ms on a busy box,
-                # which pushed _ping_endpoint's full path-discovery sweep
-                # past the cap and marked the row offline despite the
-                # user actively chatting with it.
-                ping = await _asyncio.to_thread(_ping_endpoint, data["base"], data.get("api_key"), 3.5)
-                lat = round((_time.time() - t0) * 1000)
-                return {
-                    "alive": bool(ping.get("reachable")),
-                    "latency_ms": lat,
-                    "status_code": ping.get("status_code"),
-                    "error": ping.get("error"),
-                }
-            except Exception as e:
-                return {"alive": False, "latency_ms": None, "status_code": None, "error": str(e)[:120]}
-
         import asyncio as _asyncio
-        results_list = await _asyncio.gather(
-            *[_probe_one(data) for data in grouped.values()],
-            return_exceptions=False,
-        )
-        results: Dict[str, Any] = {}
-        for data, r in zip(grouped.values(), results_list):
-            for eid in data["endpoint_ids"]:
-                results[eid] = r
+        task = _local_probe_inflight.get("task")
+        if task is not None and not task.done():
+            return await task
 
-        _local_probe_cache["data"] = results
-        _local_probe_cache["time"] = now
-        return results
+        async def _compute_local_probe() -> Dict[str, Any]:
+            db = SessionLocal()
+            try:
+                if _disable_stale_cookbook_local_endpoints(db):
+                    _invalidate_models_cache()
+                endpoints = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True).all()
+                local_eps = []
+                for ep in endpoints:
+                    base = _normalize_base(ep.base_url)
+                    kind = _effective_endpoint_kind(ep, base)
+                    if _classify_endpoint(base, kind) == "local":
+                        local_eps.append((ep.id, base, ep.api_key))
+            finally:
+                db.close()
+
+            grouped: Dict[str, Dict[str, Any]] = {}
+            for ep_id, base, api_key in local_eps:
+                key = _refresh_key(base, api_key)
+                grouped.setdefault(key, {"base": base, "api_key": api_key, "endpoint_ids": []})["endpoint_ids"].append(ep_id)
+
+            async def _probe_one(data: Dict[str, Any]) -> Dict[str, Any]:
+                t0 = _time.time()
+                try:
+                    # Bumped 1.5s → 3.5s. The previous 1.5s budget was clipping
+                    # local vLLM endpoints on Tailscale links where the model
+                    # server is still loading (Qwen3.5-122B takes 2–3 min to
+                    # warm); /v1/models can take 500–2500 ms on a busy box,
+                    # which pushed _ping_endpoint's full path-discovery sweep
+                    # past the cap and marked the row offline despite the
+                    # user actively chatting with it.
+                    ping = await _asyncio.to_thread(_ping_endpoint, data["base"], data.get("api_key"), 3.5)
+                    lat = round((_time.time() - t0) * 1000)
+                    return {
+                        "alive": bool(ping.get("reachable")),
+                        "latency_ms": lat,
+                        "status_code": ping.get("status_code"),
+                        "error": ping.get("error"),
+                    }
+                except Exception as e:
+                    return {"alive": False, "latency_ms": None, "status_code": None, "error": str(e)[:120]}
+
+            results_list = await _asyncio.gather(
+                *[_probe_one(data) for data in grouped.values()],
+                return_exceptions=False,
+            )
+            results: Dict[str, Any] = {}
+            for data, r in zip(grouped.values(), results_list):
+                for eid in data["endpoint_ids"]:
+                    results[eid] = r
+
+            _local_probe_cache["data"] = results
+            _local_probe_cache["time"] = _time.time()
+            return results
+
+        task = _asyncio.create_task(_compute_local_probe())
+        _local_probe_inflight["task"] = task
+        try:
+            return await task
+        finally:
+            if _local_probe_inflight.get("task") is task:
+                _local_probe_inflight["task"] = None
 
     @router.get("/ping")
     def ping_endpoints(request: Request):
@@ -1636,6 +1775,8 @@ def setup_model_routes(model_discovery):
         require_admin(request)
         db = SessionLocal()
         try:
+            if _disable_stale_cookbook_local_endpoints(db):
+                _invalidate_models_cache()
             rows = db.query(ModelEndpoint).order_by(ModelEndpoint.created_at).all()
             results = []
             for r in rows:
@@ -1643,67 +1784,11 @@ def setup_model_routes(model_discovery):
                 hidden = _hidden_model_ids(r)
                 pinned = _normalize_model_ids(getattr(r, "pinned_models", None))
                 visible = _visible_models(all_models, r.hidden_models, pinned)
-                # Endpoint counts as reachable if it has any model — including
-                # admin-pinned IDs that a probe would never surface.
-                status = "online" if (all_models or pinned) else "offline"
+                # Keep the list route cache-only. It feeds Settings →
+                # Added Models and must render immediately; explicit
+                # Refresh/Probe endpoints do the network work.
+                status = "online" if (all_models or pinned) else ("empty" if r.is_enabled else "offline")
                 ping = None
-                # When cached_models is empty, do a quick reachability probe.
-                # Bumped 1.0s → 3.5s because the user reported endpoints they
-                # were ACTIVELY chatting with showed "offline" — the previous
-                # 1s timeout was clipping live cloud endpoints (DeepSeek can
-                # take 1.5–2.5s on /v1/models when their region is under load,
-                # vLLM on a remote GPU box behind SSH can also push past 1s).
-                # 3.5s still keeps the picker render snappy in the common
-                # "everything's already cached" path because this branch only
-                # runs for endpoints with an empty cached_models.
-                if not all_models and not pinned and r.is_enabled:
-                    base_for_ping = _normalize_base(r.base_url)
-                    kind_for_ping = _effective_endpoint_kind(r, base_for_ping)
-                    ping_timeout = 10.0 if _classify_endpoint(base_for_ping, kind_for_ping) == "local" else 3.5
-                    ping = _ping_endpoint(r.base_url, r.api_key, timeout=ping_timeout)
-                    if ping.get("reachable"):
-                        status = "loading" if ping.get("loading") else "empty"
-                        if ping.get("loading"):
-                            base = _normalize_base(r.base_url)
-                            kind = _effective_endpoint_kind(r, base)
-                            results.append({
-                                "id": r.id,
-                                "name": r.name,
-                                "base_url": r.base_url,
-                                "has_key": bool(r.api_key),
-                                "api_key_fingerprint": _api_key_fingerprint(r.api_key),
-                                "is_enabled": r.is_enabled,
-                                "models": visible,
-                                "pinned_models": pinned,
-                                "hidden_count": len(hidden),
-                                "online": True,
-                                "status": status,
-                                "ping_error": (ping or {}).get("error") if ping else None,
-                                "model_type": getattr(r, "model_type", None) or "llm",
-                                "supports_tools": getattr(r, "supports_tools", None),
-                                "endpoint_kind": kind,
-                                "category": _classify_endpoint(base, kind),
-                                "model_refresh_mode": _endpoint_refresh_mode(r, kind),
-                                "model_refresh_interval": getattr(r, "model_refresh_interval", None),
-                                "model_refresh_timeout": getattr(r, "model_refresh_timeout", None),
-                            })
-                            continue
-                        # Best-effort: if the probe came back reachable, try
-                        # to populate cached_models in the background so the
-                        # NEXT picker load shows "online" instead of "empty".
-                        # Failure here is silent — we already returned the
-                        # "empty" status, and the existing background refresh
-                        # path will eventually fill it in too.
-                        try:
-                            probed = _probe_endpoint(r.base_url, r.api_key, timeout=max(5, int(ping_timeout)))
-                            if probed:
-                                r.cached_models = json.dumps(probed)
-                                db.commit()
-                                all_models = probed
-                                visible = _visible_models(all_models, r.hidden_models, pinned)
-                                status = "online"
-                        except Exception as _refill_err:
-                            logger.debug(f"opportunistic cached_models refill failed for {r.id}: {_refill_err!r}")
                 base = _normalize_base(r.base_url)
                 kind = _effective_endpoint_kind(r, base)
                 results.append({
@@ -1833,7 +1918,14 @@ def setup_model_routes(model_discovery):
                 if api_key.strip() and not existing.api_key:
                     existing.api_key = api_key.strip()
                     changed = True
-                if should_probe:
+                # Keep duplicate endpoint registration cheap. This path is hit
+                # by Cookbook/browser auto-register flows and can run while the
+                # user is sending a chat message. Probing a stale LAN endpoint
+                # here used to hold the request open for tens of seconds and
+                # contend with session creation, making "send" feel blocked.
+                # Explicit "require models" calls still probe; normal refresh
+                # belongs to /model-endpoints/{id}/models or /probe.
+                if require_model_list:
                     probed_models = _probe_endpoint(
                         base_url,
                         (api_key.strip() or existing.api_key or None),
@@ -1920,7 +2012,18 @@ def setup_model_routes(model_discovery):
                     ModelEndpoint.is_enabled == True  # noqa: E712
                 ).all()
             }
-            if _default_endpoint_needs_assignment(settings.get("default_endpoint_id") or "", enabled_ids):
+            current_default_id = settings.get("default_endpoint_id") or ""
+            current_default_ep = None
+            if current_default_id:
+                current_default_ep = db.query(ModelEndpoint).filter(
+                    ModelEndpoint.id == current_default_id
+                ).first()
+            if _default_endpoint_needs_assignment(
+                current_default_id,
+                enabled_ids,
+                current_default_endpoint=current_default_ep,
+                current_default_model=settings.get("default_model") or "",
+            ):
                 from src.endpoint_resolver import _first_chat_model
                 settings["default_endpoint_id"] = ep.id
                 settings["default_model"] = _first_chat_model(model_ids) or ""

routes/note_routes.py

@@ -483,11 +483,22 @@ async def dispatch_reminder(
                 api_key = intg.get("api_key", "")
                 if api_key:
                     hdrs["Authorization"] = f"Bearer {api_key}"
-                async with httpx.AsyncClient(timeout=10.0) as client:
-                    resp = await client.post(f"{base}/{topic}", content=ntfy_body, headers=hdrs)
-                    ntfy_sent = resp.is_success
-                    if not ntfy_sent:
-                        ntfy_error = f"ntfy returned HTTP {resp.status_code}"
+                # SSRF guard — same check (and env knob) as the webhook branch
+                # above: link-local / metadata addresses are always rejected;
+                # REMINDER_WEBHOOK_BLOCK_PRIVATE_IPS=true also blocks RFC-1918
+                # so a ntfy base_url can't be pointed at internal services.
+                import os as _os
+                from src.url_safety import check_outbound_url as _chk
+                _block = _os.getenv("REMINDER_WEBHOOK_BLOCK_PRIVATE_IPS", "false").lower() == "true"
+                _ok, _reason = _chk(f"{base}/{topic}", block_private=_block)
+                if not _ok:
+                    ntfy_error = f"ntfy URL rejected: {_reason}"
+                else:
+                    async with httpx.AsyncClient(timeout=10.0) as client:
+                        resp = await client.post(f"{base}/{topic}", content=ntfy_body, headers=hdrs)
+                        ntfy_sent = resp.is_success
+                        if not ntfy_sent:
+                            ntfy_error = f"ntfy returned HTTP {resp.status_code}"
             else:
                 ntfy_error = "No enabled ntfy integration"
         except Exception as e:

routes/research/__init__.py

@@ -0,0 +1,5 @@
+"""Research route domain package (slice 2b, #4082/#4071).
+
+Contains research_routes.py, migrated from the flat routes/ directory.
+Backward-compat shim at routes/research_routes.py re-exports from here.
+"""

routes/research/research_routes.py

@@ -0,0 +1,783 @@
+"""Research background task routes — /api/research/*."""
+
+import asyncio
+import json
+import logging
+import re
+import uuid
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+from fastapi import APIRouter, HTTPException, Query, Request
+from fastapi.responses import HTMLResponse, StreamingResponse
+from pydantic import BaseModel, Field
+from core.middleware import INTERNAL_TOOL_USER
+from src.endpoint_resolver import resolve_endpoint
+from src.auth_helpers import _auth_disabled, get_current_user
+from core.auth import RESERVED_USERNAMES
+from src.constants import DEEP_RESEARCH_DIR
+
+_SESSION_ID_RE = re.compile(r"^[a-zA-Z0-9-]{1,128}$")
+
+
+def _validate_session_id(session_id: str) -> str:
+    if not _SESSION_ID_RE.fullmatch(session_id):
+        raise HTTPException(400, "Invalid session ID format")
+    return session_id
+
+
+def _research_storage_root() -> Path:
+    return Path(DEEP_RESEARCH_DIR).resolve()
+
+
+def _find_research_path(session_id: str) -> Path | None:
+    """Find a persisted research file without deriving its path from input."""
+    expected_name = f"{_validate_session_id(session_id)}.json"
+    root = _research_storage_root()
+    for stored_path in root.glob("*.json"):
+        if stored_path.name != expected_name:
+            continue
+        resolved = stored_path.resolve()
+        try:
+            resolved.relative_to(root)
+        except ValueError:
+            return None
+        if not resolved.is_file():
+            return None
+        return resolved
+    return None
+
+
+def _require_research_path(session_id: str) -> Path:
+    path = _find_research_path(session_id)
+    if path is None:
+        raise HTTPException(404, "Research not found")
+    return path
+
+
+def _find_owned_research_path(session_id: str, user: str) -> Path | None:
+    path = _find_research_path(session_id)
+    if path is None:
+        return None
+    try:
+        owner = json.loads(path.read_text(encoding="utf-8")).get("owner")
+    except Exception:
+        return None
+    if owner != user:
+        return None
+    return path
+
+
+logger = logging.getLogger(__name__)
+
+# Model-name substrings that are NOT chat/generation models — research must
+# never pick these as its model. An OpenAI-style endpoint often lists
+# `text-embedding-ada-002` etc. first in its model list, which is why research
+# was failing with "Cannot reach model 'text-embedding-ada-002'".
+_NON_CHAT_MODEL = (
+    "text-embedding", "embedding", "tts-", "whisper", "dall-e",
+    "moderation", "rerank", "reranker", "clip", "stable-diffusion",
+)
+
+_RESEARCH_IMAGE_BLOCKLIST = {
+    "cdn.shopify.com/s/files/1/0179/4388/7926/files/icon.png",
+}
+
+
+def _is_research_icon_or_logo_url(url: str) -> bool:
+    path = url.lower().split("?")[0]
+    return any(token in path for token in (
+        "/logo", "logo_", "-logo", "favicon", "apple-touch-icon",
+        "sprite", "icon-", "_icon", "/icons/", "badge",
+    ))
+
+
+def _research_thumbnail(data: dict) -> str:
+    """Pick the same first visible image the visual report uses as hero."""
+    hidden = set(data.get("hidden_images") or [])
+    seen = set()
+
+    def usable(image: str) -> bool:
+        image = str(image or "").strip()
+        if not image or image in seen or image in hidden:
+            return False
+        if not image.startswith("https://"):
+            return False
+        if image.endswith((".svg", ".ico", ".gif")):
+            return False
+        if any(blocked in image for blocked in _RESEARCH_IMAGE_BLOCKLIST):
+            return False
+        if _is_research_icon_or_logo_url(image):
+            return False
+        return True
+
+    for source in data.get("sources") or []:
+        if not isinstance(source, dict):
+            continue
+        image = str(source.get("image") or source.get("og_image") or "").strip()
+        if usable(image):
+            seen.add(image)
+            return image
+    for finding in data.get("raw_findings") or data.get("findings") or []:
+        if not isinstance(finding, dict):
+            continue
+        image = str(finding.get("image") or finding.get("og_image") or "").strip()
+        if usable(image):
+            seen.add(image)
+            return image
+    return ""
+
+
+def _first_chat_model(models) -> str:
+    """First model that isn't an embedding/tts/etc. — falls back to models[0]."""
+    for m in (models or []):
+        if not any(p in str(m).lower() for p in _NON_CHAT_MODEL):
+            return m
+    return (models[0] if models else "")
+
+
+def _resolve_research_endpoint(sess, owner: Optional[str] = None) -> tuple:
+    """Return (endpoint_url, model, headers) for Deep Research, checking admin overrides."""
+    owner = owner or getattr(sess, "owner", None) or None
+    url, model, headers = resolve_endpoint(
+        "research",
+        fallback_url=sess.endpoint_url,
+        fallback_model=sess.model,
+        fallback_headers=sess.headers,
+        owner=owner,
+    )
+    return url, model, headers
+
+
+def _owned_enabled_endpoint(db, owner, endpoint_id=None):
+    """An enabled ModelEndpoint VISIBLE to `owner` (their own rows + legacy
+    null-owner "shared" rows), optionally narrowed to a specific endpoint_id;
+    None if nothing visible matches.
+
+    Owner-scoped on purpose. ModelEndpoint is per-user (core/database.py: non-null
+    owner = private, "the model picker only shows the endpoint to that user") and
+    holds a decrypted `api_key`. /api/research/start feeds the resolved row's
+    api_key + base_url into research_handler.start_research(llm_endpoint=,
+    llm_headers=), so an UNSCOPED lookup — by the caller-supplied endpoint_id, or
+    via the bare first-enabled fallback — would let a research-privileged user
+    spend ANOTHER user's API key/quota and reach whatever internal base_url they
+    configured. Mirrors webhook_routes._first_enabled_endpoint and
+    session_routes._owned_endpoint. A null/empty owner is a no-op (single-user /
+    legacy mode).
+    """
+    from src.database import ModelEndpoint
+    from src.auth_helpers import owner_filter
+    q = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)  # noqa: E712
+    if endpoint_id:
+        q = q.filter(ModelEndpoint.id == endpoint_id)
+    return owner_filter(q, ModelEndpoint, owner).first()
+
+
+def _resolve_endpoint_runtime(ep, owner=None, model: Optional[str] = None):
+    """Resolve a ModelEndpoint row into (chat_url, model, headers).
+
+    Mirrors endpoint_resolver.resolve_endpoint's provider-auth handling for
+    panel-selected research endpoints. ChatGPT Subscription endpoints keep
+    OAuth tokens in ProviderAuthSession, so ep.api_key is intentionally empty.
+    """
+    from src.endpoint_resolver import (
+        build_chat_url,
+        build_headers,
+        resolve_endpoint_runtime as resolve_model_endpoint_runtime,
+    )
+
+    try:
+        base, api_key = resolve_model_endpoint_runtime(ep, owner=owner)
+    except Exception as e:
+        logger.warning("Could not resolve endpoint credentials for research: %s", e)
+        return None
+
+    ep_model = (model or "").strip()
+    if not ep_model:
+        try:
+            models = json.loads(ep.cached_models) if ep.cached_models else []
+            if models:
+                ep_model = _first_chat_model(models)
+        except Exception:
+            pass
+    if not ep_model:
+        return None
+    return build_chat_url(base), ep_model, build_headers(api_key, base)
+
+
+def setup_research_routes(research_handler, session_manager=None) -> APIRouter:
+    router = APIRouter(tags=["research"])
+
+    def _require_user(request: Request) -> str:
+        """All research endpoints require an authenticated user. Research
+        data isn't owner-scoped in the on-disk JSON yet, so we at least
+        block anonymous access. Multi-tenant deploys should additionally
+        verify the session belongs to this user."""
+        user = get_current_user(request)
+        if not user:
+            if _auth_disabled():
+                return ""
+            raise HTTPException(401, "Not authenticated")
+        return user
+
+    def _owns_in_memory(session_id: str, user: str) -> bool:
+        """Ownership check for an in-flight (in-memory) research task.
+        Falls back to the on-disk JSON if the task has already finished."""
+        entry = research_handler._active_tasks.get(session_id)
+        if entry is not None:
+            return entry.get("owner", "") == user
+        # Task no longer in memory — check the persisted JSON.
+        try:
+            return _find_owned_research_path(session_id, user) is not None
+        except HTTPException:
+            return False
+
+    def _require_owned_or_active_research_path(session_id: str, user: str) -> Path | None:
+        """Validate ownership once and return the completed on-disk path.
+
+        Active running research has no completed disk path yet. Completed
+        tasks can remain in _active_tasks after persistence, so prefer their
+        owned disk path when available. Completed disk lookups still reuse the
+        path after the ownership gate.
+        """
+        entry = research_handler._active_tasks.get(session_id)
+        if entry is not None:
+            if entry.get("owner", "") != user:
+                raise HTTPException(404, "No research found for this session")
+            if entry.get("status") != "running":
+                path = _find_owned_research_path(session_id, user)
+                if path is not None:
+                    return path
+            return None
+
+        path = _find_owned_research_path(session_id, user)
+        if path is None:
+            raise HTTPException(404, "No research found for this session")
+        return path
+
+    @router.get("/api/research/active")
+    async def research_active(request: Request):
+        """List all currently active (running) research tasks."""
+        user = _require_user(request)
+        active = []
+        for sid, entry in research_handler._active_tasks.items():
+            # SECURITY: only show this user's running tasks.
+            if entry.get("owner", "") != user:
+                continue
+            if entry.get("status") == "running":
+                active.append({
+                    "session_id": sid,
+                    "query": entry.get("query", ""),
+                    "status": "running",
+                    "progress": entry.get("progress", {}),
+                    "started_at": entry.get("started_at", 0),
+                })
+        return {"active": active}
+
+    @router.get("/api/research/status/{session_id}")
+    async def research_status(session_id: str, request: Request):
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        if not _owns_in_memory(session_id, user):
+            raise HTTPException(404, "No research found for this session")
+        status = research_handler.get_status(session_id)
+        if status is None:
+            raise HTTPException(404, "No research found for this session")
+        return status
+
+    @router.post("/api/research/cancel/{session_id}")
+    async def research_cancel(session_id: str, request: Request):
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        if not _owns_in_memory(session_id, user):
+            raise HTTPException(404, "No research found for this session")
+        cancelled = research_handler.cancel_research(session_id)
+        return {"cancelled": cancelled}
+
+    @router.post("/api/research/result/{session_id}")
+    async def research_result(session_id: str, request: Request):
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        if not _owns_in_memory(session_id, user):
+            raise HTTPException(404, "No research result available")
+        result = research_handler.get_result(session_id)
+        if result is None:
+            raise HTTPException(404, "No research result available")
+        sources = research_handler.get_sources(session_id) or []
+        raw_findings = research_handler.get_raw_findings(session_id) or []
+        research_handler.clear_result(session_id)
+        return {"result": result, "sources": sources, "raw_findings": raw_findings}
+
+    def _assert_owns_research(session_id: str, user: str) -> None:
+        """404-not-403 ownership gate for a research session's on-disk JSON.
+        Use BEFORE returning any data or mutating the file."""
+        path = _require_research_path(session_id)
+        try:
+            owner = json.loads(path.read_text(encoding="utf-8")).get("owner")
+        except Exception:
+            raise HTTPException(404, "Research not found")
+        if owner != user:
+            raise HTTPException(404, "Research not found")
+
+    @router.get("/api/research/report/{session_id}")
+    async def research_report(session_id: str, request: Request):
+        """Serve the visual HTML report for a completed research session."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        _assert_owns_research(session_id, user)
+        logger.info(f"Visual report requested for session {session_id}")
+        try:
+            html_content = research_handler.get_report_html(session_id)
+        except Exception as e:
+            logger.error(f"Visual report generation error: {e}", exc_info=True)
+            raise HTTPException(500, f"Report generation failed: {e}")
+        if html_content is None:
+            logger.warning(f"No report data found for session {session_id}")
+            raise HTTPException(404, "No visual report available for this session")
+        return HTMLResponse(content=html_content)
+
+    class HideImageRequest(BaseModel):
+        url: str
+
+    @router.post("/api/research/{session_id}/hide-image")
+    async def research_hide_image(session_id: str, body: HideImageRequest, request: Request):
+        """Mark an image URL as hidden for this research's visual report.
+        Persisted to the research JSON so subsequent /report renders skip it."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        _assert_owns_research(session_id, user)
+        ok = research_handler.hide_image(session_id, body.url)
+        if not ok:
+            raise HTTPException(404, "Research not found")
+        return {"ok": True}
+
+    @router.post("/api/research/{session_id}/unhide-images")
+    async def research_unhide_images(session_id: str, request: Request):
+        """Clear the hidden-images list for a research session."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        _assert_owns_research(session_id, user)
+        ok = research_handler.unhide_all_images(session_id)
+        if not ok:
+            raise HTTPException(404, "Research not found")
+        return {"ok": True}
+
+    @router.get("/api/research/library")
+    async def research_library(
+        request: Request,
+        search: Optional[str] = Query(None),
+        sort: str = Query("recent"),
+        limit: int = Query(50),
+        archived: bool = Query(False),
+    ):
+        user = _require_user(request)
+        """List all completed research for the Library panel."""
+        data_dir = Path(DEEP_RESEARCH_DIR)
+        items = []
+        for p in data_dir.glob("*.json"):
+            try:
+                d = json.loads(p.read_text(encoding="utf-8"))
+                # SECURITY: only show research belonging to this user. Legacy
+                # JSONs without an `owner` field are hidden — auth was the only
+                # gate before, so every user saw every other user's reports.
+                if d.get("owner") != user:
+                    continue
+                # Archived view shows ONLY archived reports; default hides them.
+                if bool(d.get("archived")) != archived:
+                    continue
+                query = d.get("query", "")
+                if search and search.lower() not in query.lower():
+                    continue
+                sources = d.get("sources", [])
+                items.append({
+                    "id": p.stem,
+                    "query": query,
+                    "category": d.get("category") or "",
+                    "source_count": len(sources),
+                    "status": d.get("status", "done"),
+                    "duration": d.get("stats", {}).get("Duration", ""),
+                    "rounds": d.get("stats", {}).get("Rounds", ""),
+                    "started_at": d.get("started_at", 0),
+                    "completed_at": d.get("completed_at", 0),
+                    "archived": bool(d.get("archived")),
+                    "thumbnail": _research_thumbnail(d),
+                })
+            except Exception:
+                continue
+
+        # Sort
+        if sort == "recent":
+            items.sort(key=lambda x: x["completed_at"] or 0, reverse=True)
+        elif sort == "oldest":
+            items.sort(key=lambda x: x["completed_at"] or 0)
+        elif sort == "most-messages":
+            items.sort(key=lambda x: x["source_count"], reverse=True)
+        elif sort == "alpha":
+            items.sort(key=lambda x: x["query"].lower())
+
+        return {"research": items[:limit], "total": len(items)}
+
+    @router.get("/api/research/detail/{session_id}")
+    async def research_detail(session_id: str, request: Request):
+        """Return the full JSON for a single research result — sources,
+        summary, stats — used by the Library preview panel."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        path = _require_research_path(session_id)
+        try:
+            data = json.loads(path.read_text(encoding="utf-8"))
+        except Exception as e:
+            raise HTTPException(500, f"Failed to read research: {e}")
+        # SECURITY: 404 (not 403) so we don't leak that the report exists.
+        if data.get("owner") != user:
+            raise HTTPException(404, "Research not found")
+        return data
+
+    @router.post("/api/research/{session_id}/archive")
+    async def research_archive(session_id: str, request: Request, archived: bool = Query(True)):
+        """Soft-archive / restore a research report (sets `archived` in its JSON)."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        path = _require_research_path(session_id)
+        try:
+            data = json.loads(path.read_text(encoding="utf-8"))
+            if data.get("owner") != user:
+                raise HTTPException(404, "Research not found")
+            data["archived"] = bool(archived)
+            path.write_text(json.dumps(data), encoding="utf-8")
+        except HTTPException:
+            raise
+        except Exception as e:
+            raise HTTPException(500, f"Failed to update research: {e}")
+        return {"ok": True, "id": session_id, "archived": bool(archived)}
+
+    @router.delete("/api/research/{session_id}")
+    async def research_delete(session_id: str, request: Request):
+        """Delete a research result from disk."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        json_path = _find_research_path(session_id)
+        deleted = False
+        if json_path is not None:
+            # SECURITY: verify ownership before letting the caller delete it.
+            try:
+                data = json.loads(json_path.read_text(encoding="utf-8"))
+                if data.get("owner") != user:
+                    raise HTTPException(404, "Research not found")
+            except HTTPException:
+                raise
+            except Exception:
+                raise HTTPException(404, "Research not found")
+            json_path.unlink()
+            deleted = True
+        return {"deleted": deleted}
+
+    # ------------------------------------------------------------------
+    # Panel endpoints — launch research without a chat session
+    # ------------------------------------------------------------------
+
+    class ResearchStartRequest(BaseModel):
+        query: str
+        # max_rounds=0 means "Auto" — let the AI decide when to stop, capped at 20.
+        max_rounds: int = Field(default=0, ge=0, le=20)
+        search_provider: Optional[str] = None
+        endpoint_id: Optional[str] = None
+        model: Optional[str] = None
+        max_time: int = Field(default=300, ge=60, le=1800)
+        extraction_timeout: Optional[int] = Field(default=None, ge=15, le=3600)
+        extraction_concurrency: Optional[int] = Field(default=None, ge=1, le=12)
+        category: Optional[str] = None
+
+    @router.post("/api/research/start")
+    async def research_start(body: ResearchStartRequest, request: Request):
+        """Launch a research job from the dedicated panel."""
+        from src.auth_helpers import require_privilege
+        user = require_privilege(request, "can_use_research")
+        if user == INTERNAL_TOOL_USER:
+            tool_owner = (request.headers.get("X-Odysseus-Owner") or "").strip()
+            if tool_owner and tool_owner not in RESERVED_USERNAMES:
+                auth_mgr = getattr(request.app.state, "auth_manager", None)
+                if auth_mgr is not None and getattr(auth_mgr, "is_configured", False):
+                    try:
+                        privs = auth_mgr.get_privileges(tool_owner) or {}
+                        if not privs.get("can_use_research", True):
+                            raise HTTPException(403, f"Your account is not allowed to can use research.")
+                    except HTTPException:
+                        raise
+                    except Exception:
+                        pass
+                user = tool_owner
+        session_id = f"rp-{uuid.uuid4().hex[:12]}"
+
+        if body.endpoint_id:
+            from src.database import SessionLocal
+            db = SessionLocal()
+            try:
+                # Owner-scoped: never resolve another user's private endpoint
+                # (and its decrypted api_key / internal base_url). A scoped miss
+                # reads as 404 so the endpoint's existence isn't revealed.
+                ep = _owned_enabled_endpoint(db, user, body.endpoint_id)
+                if not ep:
+                    raise HTTPException(404, "Endpoint not found or disabled")
+                resolved = _resolve_endpoint_runtime(ep, owner=user, model=body.model)
+                if not resolved:
+                    raise HTTPException(400, "Endpoint is not configured with a usable model.")
+                ep_url, ep_model, ep_headers = resolved
+            finally:
+                db.close()
+        else:
+            ep_url, ep_model, ep_headers = resolve_endpoint("research", owner=user)
+            if not ep_url:
+                ep_url, ep_model, ep_headers = resolve_endpoint("utility", owner=user)
+            # When neither research nor utility is configured, use the user's
+            # configured DEFAULT model (default_endpoint_id/default_model) rather
+            # than arbitrarily grabbing the first enabled endpoint's first model
+            # (which surfaced gpt-3.5). "Default" should mean the default model.
+            if not ep_url:
+                ep_url, ep_model, ep_headers = resolve_endpoint("default", owner=user)
+            if not ep_url:
+                ep_url, ep_model, ep_headers = resolve_endpoint("chat", owner=user)
+            if not ep_url:
+                from src.database import SessionLocal
+                db = SessionLocal()
+                try:
+                    # Owner-scoped first-enabled fallback: the caller's own rows
+                    # + legacy null-owner shared rows only — never borrow another
+                    # user's private endpoint/api_key. Same fix as the
+                    # /api/v1/chat fallback (webhook_routes._first_enabled_endpoint).
+                    ep = _owned_enabled_endpoint(db, user)
+                    if ep:
+                        resolved = _resolve_endpoint_runtime(ep, owner=user)
+                        if resolved:
+                            ep_url, ep_model, ep_headers = resolved
+                finally:
+                    db.close()
+            if not ep_url:
+                raise HTTPException(400, "No endpoints configured. Add one in Settings first.")
+            if body.model:
+                ep_model = body.model
+
+        # max_rounds=0 → "Auto", let AI decide; pass 20 as the safety cap.
+        effective_max_rounds = body.max_rounds if body.max_rounds > 0 else 20
+        research_handler.start_research(
+            session_id=session_id,
+            query=body.query,
+            llm_endpoint=ep_url,
+            llm_model=ep_model,
+            max_time=body.max_time,
+            llm_headers=ep_headers,
+            max_rounds=effective_max_rounds,
+            search_provider=body.search_provider or None,
+            category=body.category or None,
+            extraction_timeout=body.extraction_timeout,
+            extraction_concurrency=body.extraction_concurrency,
+            owner=user,
+        )
+        return {"session_id": session_id, "status": "running", "query": body.query}
+
+    @router.get("/api/research/stream/{session_id}")
+    async def research_stream(session_id: str, request: Request):
+        """SSE stream of research progress events."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        if not _owns_in_memory(session_id, user):
+            raise HTTPException(404, "No research found for this session")
+        async def _generate():
+            last_progress = None
+            while True:
+                status = research_handler.get_status(session_id)
+                if status is None:
+                    yield f"data: {json.dumps({'status': 'not_found'})}\n\n"
+                    return
+                st = status.get("status", "")
+                progress = status.get("progress", {})
+                if progress != last_progress:
+                    last_progress = progress
+                    yield f"data: {json.dumps({**progress, 'status': st})}\n\n"
+                if st != "running":
+                    final = {'status': st, 'final': True}
+                    task = research_handler._active_tasks.get(session_id, {})
+                    if st == "error" and task.get("result"):
+                        final['error'] = str(task["result"])[:500]
+                    yield f"data: {json.dumps(final)}\n\n"
+                    return
+                await asyncio.sleep(1.5)
+
+        return StreamingResponse(
+            _generate(),
+            media_type="text/event-stream",
+            headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},
+        )
+
+    @router.post("/api/research/result-peek/{session_id}")
+    async def research_result_peek(session_id: str, request: Request):
+        """Get research result without clearing it (for panel use)."""
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        owned_disk_path = _require_owned_or_active_research_path(session_id, user)
+        result = research_handler.get_result(session_id)
+        if result is None:
+            p = owned_disk_path
+            if p is not None:
+                d = json.loads(p.read_text(encoding="utf-8"))
+                return {
+                    "result": d.get("result", ""),
+                    "sources": d.get("sources", []),
+                    "raw_findings": d.get("raw_findings", []),
+                    "category": d.get("category") or "",
+                }
+            raise HTTPException(404, "No research result available")
+        sources = research_handler.get_sources(session_id) or []
+        raw_findings = research_handler.get_raw_findings(session_id) or []
+        return {"result": result, "sources": sources, "raw_findings": raw_findings, "category": ""}
+
+    @router.post("/api/research/spinoff/{session_id}")
+    async def research_spinoff(session_id: str, request: Request):
+        """Create a new chat session pre-seeded with this research as context.
+
+        Reads the persisted research result + sources for `session_id`, creates
+        a fresh session (inheriting endpoint/model/headers from the source
+        session if available, otherwise from the resolved chat endpoint), and
+        injects a single system message containing the report and sources so
+        the user can ask follow-up questions in a clean conversation.
+        """
+        user = _require_user(request)
+        _validate_session_id(session_id)
+        # SECURITY: gate on ownership before reading the persisted research —
+        # otherwise any authenticated user could spin off (and thereby read)
+        # another user's report by guessing its session ID. Mirrors every other
+        # endpoint in this file (see result_peek above).
+        owned_disk_path = _require_owned_or_active_research_path(session_id, user)
+        if session_manager is None:
+            raise HTTPException(500, "session_manager not configured")
+
+        # Load research data — prefer in-memory result, fall back to disk
+        result = research_handler.get_result(session_id)
+        sources = research_handler.get_sources(session_id) or []
+        query = ""
+
+        path = owned_disk_path
+        if path is not None:
+            try:
+                disk = json.loads(path.read_text(encoding="utf-8"))
+                if not result:
+                    result = disk.get("result")
+                if not sources:
+                    sources = disk.get("sources", []) or []
+                query = disk.get("query", "") or ""
+            except Exception as e:
+                logger.warning(f"Could not read research JSON for spinoff: {e}")
+
+        if not result:
+            raise HTTPException(404, "No research result available for this session")
+
+        # Inherit endpoint/model/headers from the source session when possible.
+        # For panel-launched research (rp-* IDs), there is no chat session, so
+        # fall back through the same chain as /api/research/start: research →
+        # utility → first enabled endpoint in the DB.
+        ep_url, ep_model, ep_headers = "", "", {}
+        try:
+            src_sess = session_manager.get_session(session_id)
+            ep_url = src_sess.endpoint_url or ""
+            ep_model = src_sess.model or ""
+            ep_headers = dict(src_sess.headers or {})
+        except KeyError:
+            pass
+
+        def _merge(r_url, r_model, r_headers):
+            nonlocal ep_url, ep_model, ep_headers
+            if not ep_url and r_url:
+                ep_url = r_url
+            if not ep_model and r_model:
+                ep_model = r_model
+            if not ep_headers and r_headers:
+                ep_headers = dict(r_headers)
+
+        if not ep_url or not ep_model:
+            _merge(*resolve_endpoint("chat", owner=user))
+        if not ep_url or not ep_model:
+            _merge(*resolve_endpoint("research", owner=user))
+        if not ep_url or not ep_model:
+            _merge(*resolve_endpoint("utility", owner=user))
+        if not ep_url or not ep_model:
+            # Last resort: this user's enabled endpoint, plus legacy shared rows.
+            from src.database import SessionLocal
+            from src.endpoint_resolver import normalize_base, build_chat_url, build_headers
+            db = SessionLocal()
+            try:
+                ep = _owned_enabled_endpoint(db, user)
+                if ep:
+                    base = normalize_base(ep.base_url)
+                    fallback_url = build_chat_url(base)
+                    fallback_headers = build_headers(ep.api_key, base)
+                    fallback_model = ""
+                    if ep.cached_models:
+                        try:
+                            models = json.loads(ep.cached_models)
+                            if models:
+                                fallback_model = _first_chat_model(models)
+                        except Exception:
+                            pass
+                    _merge(fallback_url, fallback_model, fallback_headers)
+            finally:
+                db.close()
+
+        if not ep_url or not ep_model:
+            raise HTTPException(400, "No endpoint configured — add one in Settings first")
+
+        # Create new session
+        new_sid = str(uuid.uuid4())
+
+        title_query = (query or "research").strip()
+        if len(title_query) > 60:
+            title_query = title_query[:57] + "…"
+        new_name = f"Follow-up: {title_query}"
+
+        new_sess = session_manager.create_session(
+            session_id=new_sid,
+            name=new_name,
+            endpoint_url=ep_url,
+            model=ep_model,
+            rag=False,
+            owner=user,
+        )
+        if ep_headers:
+            new_sess.headers = ep_headers
+            session_manager.save_sessions()
+        try:
+            from src.event_bus import fire_event
+            fire_event("session_created", user)
+        except Exception:
+            logger.debug("session_created event dispatch failed", exc_info=True)
+
+        # Build the priming system message — report only, no sources injected.
+        # The user can open the visual report for source details; keeping sources
+        # out of the chat context saves tokens and avoids the AI fabricating
+        # citations.
+        date_str = datetime.utcnow().strftime("%Y-%m-%d")
+        primer = (
+            f"[Research context — {date_str}]\n\n"
+            f"The user previously ran a deep research investigation. Use the "
+            f"report below as your primary knowledge base when answering "
+            f"follow-up questions. If the user asks something not covered, "
+            f"say so plainly rather than guessing.\n\n"
+            f"=== ORIGINAL QUERY ===\n{query or '(not recorded)'}\n\n"
+            f"=== REPORT ===\n{result}"
+        )
+
+        from core.models import ChatMessage
+        new_sess.add_message(ChatMessage(
+            role="system",
+            content=primer,
+            metadata={"research_spinoff_from": session_id},
+        ))
+        session_manager.save_sessions()
+
+        return {
+            "session_id": new_sid,
+            "name": new_name,
+            "source_count": len(sources),
+        }
+
+    return router

routes/research_routes.py

@@ -1,678 +1,17 @@
-"""Research background task routes — /api/research/*."""
+"""Backward-compat shim — canonical location is routes/research/research_routes.py.
 
-import asyncio
-import json
-import logging
-import re
-import uuid
-from datetime import datetime
-from pathlib import Path
-from typing import Optional
+This module is replaced in ``sys.modules`` by the canonical module object so
+that ``import routes.research_routes``, ``from routes.research_routes import X``,
+``importlib.import_module("routes.research_routes")``, and
+``monkeypatch.setattr("routes.research_routes.ATTR", ...)`` (string-targeted
+patch used by ``test_research_owner_scope_routes.py``) all operate on the
+*same* object the application actually uses. Keeps existing import paths
+working after slice 2b (#4082/#4071). Source-introspection tests read the
+canonical file by path.
+"""
 
-from fastapi import APIRouter, HTTPException, Query, Request
-from fastapi.responses import HTMLResponse, StreamingResponse
-from pydantic import BaseModel, Field
-from core.middleware import INTERNAL_TOOL_USER
-from src.endpoint_resolver import resolve_endpoint
-from src.auth_helpers import _auth_disabled, get_current_user
-from core.auth import RESERVED_USERNAMES
-from src.constants import DEEP_RESEARCH_DIR
+import sys as _sys
 
-_SESSION_ID_RE = re.compile(r"^[a-zA-Z0-9-]{1,128}$")
+from routes.research import research_routes as _canonical  # noqa: F401
 
-logger = logging.getLogger(__name__)
-
-# Model-name substrings that are NOT chat/generation models — research must
-# never pick these as its model. An OpenAI-style endpoint often lists
-# `text-embedding-ada-002` etc. first in its model list, which is why research
-# was failing with "Cannot reach model 'text-embedding-ada-002'".
-_NON_CHAT_MODEL = (
-    "text-embedding", "embedding", "tts-", "whisper", "dall-e",
-    "moderation", "rerank", "reranker", "clip", "stable-diffusion",
-)
-
-
-def _first_chat_model(models) -> str:
-    """First model that isn't an embedding/tts/etc. — falls back to models[0]."""
-    for m in (models or []):
-        if not any(p in str(m).lower() for p in _NON_CHAT_MODEL):
-            return m
-    return (models[0] if models else "")
-
-
-def _resolve_research_endpoint(sess, owner: Optional[str] = None) -> tuple:
-    """Return (endpoint_url, model, headers) for Deep Research, checking admin overrides."""
-    owner = owner or getattr(sess, "owner", None) or None
-    url, model, headers = resolve_endpoint(
-        "research",
-        fallback_url=sess.endpoint_url,
-        fallback_model=sess.model,
-        fallback_headers=sess.headers,
-        owner=owner,
-    )
-    return url, model, headers
-
-
-def _owned_enabled_endpoint(db, owner, endpoint_id=None):
-    """An enabled ModelEndpoint VISIBLE to `owner` (their own rows + legacy
-    null-owner "shared" rows), optionally narrowed to a specific endpoint_id;
-    None if nothing visible matches.
-
-    Owner-scoped on purpose. ModelEndpoint is per-user (core/database.py: non-null
-    owner = private, "the model picker only shows the endpoint to that user") and
-    holds a decrypted `api_key`. /api/research/start feeds the resolved row's
-    api_key + base_url into research_handler.start_research(llm_endpoint=,
-    llm_headers=), so an UNSCOPED lookup — by the caller-supplied endpoint_id, or
-    via the bare first-enabled fallback — would let a research-privileged user
-    spend ANOTHER user's API key/quota and reach whatever internal base_url they
-    configured. Mirrors webhook_routes._first_enabled_endpoint and
-    session_routes._owned_endpoint. A null/empty owner is a no-op (single-user /
-    legacy mode).
-    """
-    from src.database import ModelEndpoint
-    from src.auth_helpers import owner_filter
-    q = db.query(ModelEndpoint).filter(ModelEndpoint.is_enabled == True)  # noqa: E712
-    if endpoint_id:
-        q = q.filter(ModelEndpoint.id == endpoint_id)
-    return owner_filter(q, ModelEndpoint, owner).first()
-
-
-def _resolve_endpoint_runtime(ep, owner=None, model: Optional[str] = None):
-    """Resolve a ModelEndpoint row into (chat_url, model, headers).
-
-    Mirrors endpoint_resolver.resolve_endpoint's provider-auth handling for
-    panel-selected research endpoints. ChatGPT Subscription endpoints keep
-    OAuth tokens in ProviderAuthSession, so ep.api_key is intentionally empty.
-    """
-    from src.endpoint_resolver import (
-        build_chat_url,
-        build_headers,
-        resolve_endpoint_runtime as resolve_model_endpoint_runtime,
-    )
-
-    try:
-        base, api_key = resolve_model_endpoint_runtime(ep, owner=owner)
-    except Exception as e:
-        logger.warning("Could not resolve endpoint credentials for research: %s", e)
-        return None
-
-    ep_model = (model or "").strip()
-    if not ep_model:
-        try:
-            models = json.loads(ep.cached_models) if ep.cached_models else []
-            if models:
-                ep_model = _first_chat_model(models)
-        except Exception:
-            pass
-    if not ep_model:
-        return None
-    return build_chat_url(base), ep_model, build_headers(api_key, base)
-
-
-def setup_research_routes(research_handler, session_manager=None) -> APIRouter:
-    router = APIRouter(tags=["research"])
-
-    def _require_user(request: Request) -> str:
-        """All research endpoints require an authenticated user. Research
-        data isn't owner-scoped in the on-disk JSON yet, so we at least
-        block anonymous access. Multi-tenant deploys should additionally
-        verify the session belongs to this user."""
-        user = get_current_user(request)
-        if not user:
-            if _auth_disabled():
-                return ""
-            raise HTTPException(401, "Not authenticated")
-        return user
-
-    def _validate_session_id(session_id: str) -> None:
-        if not _SESSION_ID_RE.fullmatch(session_id):
-            raise HTTPException(400, "Invalid session ID format")
-
-    def _owns_in_memory(session_id: str, user: str) -> bool:
-        """Ownership check for an in-flight (in-memory) research task.
-        Falls back to the on-disk JSON if the task has already finished."""
-        entry = research_handler._active_tasks.get(session_id)
-        if entry is not None:
-            return entry.get("owner", "") == user
-        # Task no longer in memory — check the persisted JSON.
-        path = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-        if not path.exists():
-            return False
-        try:
-            return json.loads(path.read_text(encoding="utf-8")).get("owner") == user
-        except Exception:
-            return False
-
-    @router.get("/api/research/active")
-    async def research_active(request: Request):
-        """List all currently active (running) research tasks."""
-        user = _require_user(request)
-        active = []
-        for sid, entry in research_handler._active_tasks.items():
-            # SECURITY: only show this user's running tasks.
-            if entry.get("owner", "") != user:
-                continue
-            if entry.get("status") == "running":
-                active.append({
-                    "session_id": sid,
-                    "query": entry.get("query", ""),
-                    "status": "running",
-                    "progress": entry.get("progress", {}),
-                    "started_at": entry.get("started_at", 0),
-                })
-        return {"active": active}
-
-    @router.get("/api/research/status/{session_id}")
-    async def research_status(session_id: str, request: Request):
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research found for this session")
-        status = research_handler.get_status(session_id)
-        if status is None:
-            raise HTTPException(404, "No research found for this session")
-        return status
-
-    @router.post("/api/research/cancel/{session_id}")
-    async def research_cancel(session_id: str, request: Request):
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research found for this session")
-        cancelled = research_handler.cancel_research(session_id)
-        return {"cancelled": cancelled}
-
-    @router.post("/api/research/result/{session_id}")
-    async def research_result(session_id: str, request: Request):
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research result available")
-        result = research_handler.get_result(session_id)
-        if result is None:
-            raise HTTPException(404, "No research result available")
-        sources = research_handler.get_sources(session_id) or []
-        raw_findings = research_handler.get_raw_findings(session_id) or []
-        research_handler.clear_result(session_id)
-        return {"result": result, "sources": sources, "raw_findings": raw_findings}
-
-    def _assert_owns_research(session_id: str, user: str) -> None:
-        """404-not-403 ownership gate for a research session's on-disk JSON.
-        Use BEFORE returning any data or mutating the file."""
-        path = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-        if not path.exists():
-            raise HTTPException(404, "Research not found")
-        try:
-            owner = json.loads(path.read_text(encoding="utf-8")).get("owner")
-        except Exception:
-            raise HTTPException(404, "Research not found")
-        if owner != user:
-            raise HTTPException(404, "Research not found")
-
-    @router.get("/api/research/report/{session_id}")
-    async def research_report(session_id: str, request: Request):
-        """Serve the visual HTML report for a completed research session."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        _assert_owns_research(session_id, user)
-        logger.info(f"Visual report requested for session {session_id}")
-        try:
-            html_content = research_handler.get_report_html(session_id)
-        except Exception as e:
-            logger.error(f"Visual report generation error: {e}", exc_info=True)
-            raise HTTPException(500, f"Report generation failed: {e}")
-        if html_content is None:
-            logger.warning(f"No report data found for session {session_id}")
-            raise HTTPException(404, "No visual report available for this session")
-        return HTMLResponse(content=html_content)
-
-    class HideImageRequest(BaseModel):
-        url: str
-
-    @router.post("/api/research/{session_id}/hide-image")
-    async def research_hide_image(session_id: str, body: HideImageRequest, request: Request):
-        """Mark an image URL as hidden for this research's visual report.
-        Persisted to the research JSON so subsequent /report renders skip it."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        _assert_owns_research(session_id, user)
-        ok = research_handler.hide_image(session_id, body.url)
-        if not ok:
-            raise HTTPException(404, "Research not found")
-        return {"ok": True}
-
-    @router.post("/api/research/{session_id}/unhide-images")
-    async def research_unhide_images(session_id: str, request: Request):
-        """Clear the hidden-images list for a research session."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        _assert_owns_research(session_id, user)
-        ok = research_handler.unhide_all_images(session_id)
-        if not ok:
-            raise HTTPException(404, "Research not found")
-        return {"ok": True}
-
-    @router.get("/api/research/library")
-    async def research_library(
-        request: Request,
-        search: Optional[str] = Query(None),
-        sort: str = Query("recent"),
-        limit: int = Query(50),
-        archived: bool = Query(False),
-    ):
-        user = _require_user(request)
-        """List all completed research for the Library panel."""
-        data_dir = Path(DEEP_RESEARCH_DIR)
-        items = []
-        for p in data_dir.glob("*.json"):
-            try:
-                d = json.loads(p.read_text(encoding="utf-8"))
-                # SECURITY: only show research belonging to this user. Legacy
-                # JSONs without an `owner` field are hidden — auth was the only
-                # gate before, so every user saw every other user's reports.
-                if d.get("owner") != user:
-                    continue
-                # Archived view shows ONLY archived reports; default hides them.
-                if bool(d.get("archived")) != archived:
-                    continue
-                query = d.get("query", "")
-                if search and search.lower() not in query.lower():
-                    continue
-                sources = d.get("sources", [])
-                items.append({
-                    "id": p.stem,
-                    "query": query,
-                    "category": d.get("category") or "",
-                    "source_count": len(sources),
-                    "status": d.get("status", "done"),
-                    "duration": d.get("stats", {}).get("Duration", ""),
-                    "rounds": d.get("stats", {}).get("Rounds", ""),
-                    "started_at": d.get("started_at", 0),
-                    "completed_at": d.get("completed_at", 0),
-                    "archived": bool(d.get("archived")),
-                })
-            except Exception:
-                continue
-
-        # Sort
-        if sort == "recent":
-            items.sort(key=lambda x: x["completed_at"] or 0, reverse=True)
-        elif sort == "oldest":
-            items.sort(key=lambda x: x["completed_at"] or 0)
-        elif sort == "most-messages":
-            items.sort(key=lambda x: x["source_count"], reverse=True)
-        elif sort == "alpha":
-            items.sort(key=lambda x: x["query"].lower())
-
-        return {"research": items[:limit], "total": len(items)}
-
-    @router.get("/api/research/detail/{session_id}")
-    async def research_detail(session_id: str, request: Request):
-        """Return the full JSON for a single research result — sources,
-        summary, stats — used by the Library preview panel."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        path = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-        if not path.exists():
-            raise HTTPException(404, "Research not found")
-        try:
-            data = json.loads(path.read_text(encoding="utf-8"))
-        except Exception as e:
-            raise HTTPException(500, f"Failed to read research: {e}")
-        # SECURITY: 404 (not 403) so we don't leak that the report exists.
-        if data.get("owner") != user:
-            raise HTTPException(404, "Research not found")
-        return data
-
-    @router.post("/api/research/{session_id}/archive")
-    async def research_archive(session_id: str, request: Request, archived: bool = Query(True)):
-        """Soft-archive / restore a research report (sets `archived` in its JSON)."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        path = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-        if not path.exists():
-            raise HTTPException(404, "Research not found")
-        try:
-            data = json.loads(path.read_text(encoding="utf-8"))
-            if data.get("owner") != user:
-                raise HTTPException(404, "Research not found")
-            data["archived"] = bool(archived)
-            path.write_text(json.dumps(data), encoding="utf-8")
-        except HTTPException:
-            raise
-        except Exception as e:
-            raise HTTPException(500, f"Failed to update research: {e}")
-        return {"ok": True, "id": session_id, "archived": bool(archived)}
-
-    @router.delete("/api/research/{session_id}")
-    async def research_delete(session_id: str, request: Request):
-        """Delete a research result from disk."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        data_dir = Path(DEEP_RESEARCH_DIR)
-        json_path = data_dir / f"{session_id}.json"
-        deleted = False
-        if json_path.exists():
-            # SECURITY: verify ownership before letting the caller delete it.
-            try:
-                data = json.loads(json_path.read_text(encoding="utf-8"))
-                if data.get("owner") != user:
-                    raise HTTPException(404, "Research not found")
-            except HTTPException:
-                raise
-            except Exception:
-                raise HTTPException(404, "Research not found")
-            json_path.unlink()
-            deleted = True
-        return {"deleted": deleted}
-
-    # ------------------------------------------------------------------
-    # Panel endpoints — launch research without a chat session
-    # ------------------------------------------------------------------
-
-    class ResearchStartRequest(BaseModel):
-        query: str
-        # max_rounds=0 means "Auto" — let the AI decide when to stop, capped at 20.
-        max_rounds: int = Field(default=0, ge=0, le=20)
-        search_provider: Optional[str] = None
-        endpoint_id: Optional[str] = None
-        model: Optional[str] = None
-        max_time: int = Field(default=300, ge=60, le=1800)
-        extraction_timeout: Optional[int] = Field(default=None, ge=15, le=3600)
-        extraction_concurrency: Optional[int] = Field(default=None, ge=1, le=12)
-        category: Optional[str] = None
-
-    @router.post("/api/research/start")
-    async def research_start(body: ResearchStartRequest, request: Request):
-        """Launch a research job from the dedicated panel."""
-        from src.auth_helpers import require_privilege
-        user = require_privilege(request, "can_use_research")
-        if user == INTERNAL_TOOL_USER:
-            tool_owner = (request.headers.get("X-Odysseus-Owner") or "").strip()
-            if tool_owner and tool_owner not in RESERVED_USERNAMES:
-                auth_mgr = getattr(request.app.state, "auth_manager", None)
-                if auth_mgr is not None and getattr(auth_mgr, "is_configured", False):
-                    try:
-                        privs = auth_mgr.get_privileges(tool_owner) or {}
-                        if not privs.get("can_use_research", True):
-                            raise HTTPException(403, f"Your account is not allowed to can use research.")
-                    except HTTPException:
-                        raise
-                    except Exception:
-                        pass
-                user = tool_owner
-        session_id = f"rp-{uuid.uuid4().hex[:12]}"
-
-        if body.endpoint_id:
-            from src.database import SessionLocal
-            db = SessionLocal()
-            try:
-                # Owner-scoped: never resolve another user's private endpoint
-                # (and its decrypted api_key / internal base_url). A scoped miss
-                # reads as 404 so the endpoint's existence isn't revealed.
-                ep = _owned_enabled_endpoint(db, user, body.endpoint_id)
-                if not ep:
-                    raise HTTPException(404, "Endpoint not found or disabled")
-                resolved = _resolve_endpoint_runtime(ep, owner=user, model=body.model)
-                if not resolved:
-                    raise HTTPException(400, "Endpoint is not configured with a usable model.")
-                ep_url, ep_model, ep_headers = resolved
-            finally:
-                db.close()
-        else:
-            ep_url, ep_model, ep_headers = resolve_endpoint("research", owner=user)
-            if not ep_url:
-                ep_url, ep_model, ep_headers = resolve_endpoint("utility", owner=user)
-            # When neither research nor utility is configured, use the user's
-            # configured DEFAULT model (default_endpoint_id/default_model) rather
-            # than arbitrarily grabbing the first enabled endpoint's first model
-            # (which surfaced gpt-3.5). "Default" should mean the default model.
-            if not ep_url:
-                ep_url, ep_model, ep_headers = resolve_endpoint("default", owner=user)
-            if not ep_url:
-                ep_url, ep_model, ep_headers = resolve_endpoint("chat", owner=user)
-            if not ep_url:
-                from src.database import SessionLocal
-                db = SessionLocal()
-                try:
-                    # Owner-scoped first-enabled fallback: the caller's own rows
-                    # + legacy null-owner shared rows only — never borrow another
-                    # user's private endpoint/api_key. Same fix as the
-                    # /api/v1/chat fallback (webhook_routes._first_enabled_endpoint).
-                    ep = _owned_enabled_endpoint(db, user)
-                    if ep:
-                        resolved = _resolve_endpoint_runtime(ep, owner=user)
-                        if resolved:
-                            ep_url, ep_model, ep_headers = resolved
-                finally:
-                    db.close()
-            if not ep_url:
-                raise HTTPException(400, "No endpoints configured. Add one in Settings first.")
-            if body.model:
-                ep_model = body.model
-
-        # max_rounds=0 → "Auto", let AI decide; pass 20 as the safety cap.
-        effective_max_rounds = body.max_rounds if body.max_rounds > 0 else 20
-        research_handler.start_research(
-            session_id=session_id,
-            query=body.query,
-            llm_endpoint=ep_url,
-            llm_model=ep_model,
-            max_time=body.max_time,
-            llm_headers=ep_headers,
-            max_rounds=effective_max_rounds,
-            search_provider=body.search_provider or None,
-            category=body.category or None,
-            extraction_timeout=body.extraction_timeout,
-            extraction_concurrency=body.extraction_concurrency,
-            owner=user,
-        )
-        return {"session_id": session_id, "status": "running", "query": body.query}
-
-    @router.get("/api/research/stream/{session_id}")
-    async def research_stream(session_id: str, request: Request):
-        """SSE stream of research progress events."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research found for this session")
-        async def _generate():
-            last_progress = None
-            while True:
-                status = research_handler.get_status(session_id)
-                if status is None:
-                    yield f"data: {json.dumps({'status': 'not_found'})}\n\n"
-                    return
-                st = status.get("status", "")
-                progress = status.get("progress", {})
-                if progress != last_progress:
-                    last_progress = progress
-                    yield f"data: {json.dumps({**progress, 'status': st})}\n\n"
-                if st != "running":
-                    final = {'status': st, 'final': True}
-                    task = research_handler._active_tasks.get(session_id, {})
-                    if st == "error" and task.get("result"):
-                        final['error'] = str(task["result"])[:500]
-                    yield f"data: {json.dumps(final)}\n\n"
-                    return
-                await asyncio.sleep(1.5)
-
-        return StreamingResponse(
-            _generate(),
-            media_type="text/event-stream",
-            headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},
-        )
-
-    @router.post("/api/research/result-peek/{session_id}")
-    async def research_result_peek(session_id: str, request: Request):
-        """Get research result without clearing it (for panel use)."""
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research found for this session")
-        result = research_handler.get_result(session_id)
-        if result is None:
-            p = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-            if p.exists():
-                d = json.loads(p.read_text(encoding="utf-8"))
-                return {
-                    "result": d.get("result", ""),
-                    "sources": d.get("sources", []),
-                    "raw_findings": d.get("raw_findings", []),
-                    "category": d.get("category") or "",
-                }
-            raise HTTPException(404, "No research result available")
-        sources = research_handler.get_sources(session_id) or []
-        raw_findings = research_handler.get_raw_findings(session_id) or []
-        return {"result": result, "sources": sources, "raw_findings": raw_findings, "category": ""}
-
-    @router.post("/api/research/spinoff/{session_id}")
-    async def research_spinoff(session_id: str, request: Request):
-        """Create a new chat session pre-seeded with this research as context.
-
-        Reads the persisted research result + sources for `session_id`, creates
-        a fresh session (inheriting endpoint/model/headers from the source
-        session if available, otherwise from the resolved chat endpoint), and
-        injects a single system message containing the report and sources so
-        the user can ask follow-up questions in a clean conversation.
-        """
-        user = _require_user(request)
-        _validate_session_id(session_id)
-        # SECURITY: gate on ownership before reading the persisted research —
-        # otherwise any authenticated user could spin off (and thereby read)
-        # another user's report by guessing its session ID. Mirrors every other
-        # endpoint in this file (see result_peek above).
-        if not _owns_in_memory(session_id, user):
-            raise HTTPException(404, "No research found for this session")
-        if session_manager is None:
-            raise HTTPException(500, "session_manager not configured")
-
-        # Load research data — prefer in-memory result, fall back to disk
-        result = research_handler.get_result(session_id)
-        sources = research_handler.get_sources(session_id) or []
-        query = ""
-
-        path = Path(DEEP_RESEARCH_DIR) / f"{session_id}.json"
-        if path.exists():
-            try:
-                disk = json.loads(path.read_text(encoding="utf-8"))
-                if not result:
-                    result = disk.get("result")
-                if not sources:
-                    sources = disk.get("sources", []) or []
-                query = disk.get("query", "") or ""
-            except Exception as e:
-                logger.warning(f"Could not read research JSON for spinoff: {e}")
-
-        if not result:
-            raise HTTPException(404, "No research result available for this session")
-
-        # Inherit endpoint/model/headers from the source session when possible.
-        # For panel-launched research (rp-* IDs), there is no chat session, so
-        # fall back through the same chain as /api/research/start: research →
-        # utility → first enabled endpoint in the DB.
-        ep_url, ep_model, ep_headers = "", "", {}
-        try:
-            src_sess = session_manager.get_session(session_id)
-            ep_url = src_sess.endpoint_url or ""
-            ep_model = src_sess.model or ""
-            ep_headers = dict(src_sess.headers or {})
-        except KeyError:
-            pass
-
-        def _merge(r_url, r_model, r_headers):
-            nonlocal ep_url, ep_model, ep_headers
-            if not ep_url and r_url:
-                ep_url = r_url
-            if not ep_model and r_model:
-                ep_model = r_model
-            if not ep_headers and r_headers:
-                ep_headers = dict(r_headers)
-
-        if not ep_url or not ep_model:
-            _merge(*resolve_endpoint("chat", owner=user))
-        if not ep_url or not ep_model:
-            _merge(*resolve_endpoint("research", owner=user))
-        if not ep_url or not ep_model:
-            _merge(*resolve_endpoint("utility", owner=user))
-        if not ep_url or not ep_model:
-            # Last resort: this user's enabled endpoint, plus legacy shared rows.
-            from src.database import SessionLocal
-            from src.endpoint_resolver import normalize_base, build_chat_url, build_headers
-            db = SessionLocal()
-            try:
-                ep = _owned_enabled_endpoint(db, user)
-                if ep:
-                    base = normalize_base(ep.base_url)
-                    fallback_url = build_chat_url(base)
-                    fallback_headers = build_headers(ep.api_key, base)
-                    fallback_model = ""
-                    if ep.cached_models:
-                        try:
-                            models = json.loads(ep.cached_models)
-                            if models:
-                                fallback_model = _first_chat_model(models)
-                        except Exception:
-                            pass
-                    _merge(fallback_url, fallback_model, fallback_headers)
-            finally:
-                db.close()
-
-        if not ep_url or not ep_model:
-            raise HTTPException(400, "No endpoint configured — add one in Settings first")
-
-        # Create new session
-        new_sid = str(uuid.uuid4())
-
-        title_query = (query or "research").strip()
-        if len(title_query) > 60:
-            title_query = title_query[:57] + "…"
-        new_name = f"Follow-up: {title_query}"
-
-        new_sess = session_manager.create_session(
-            session_id=new_sid,
-            name=new_name,
-            endpoint_url=ep_url,
-            model=ep_model,
-            rag=False,
-            owner=user,
-        )
-        if ep_headers:
-            new_sess.headers = ep_headers
-            session_manager.save_sessions()
-        try:
-            from src.event_bus import fire_event
-            fire_event("session_created", user)
-        except Exception:
-            logger.debug("session_created event dispatch failed", exc_info=True)
-
-        # Build the priming system message — report only, no sources injected.
-        # The user can open the visual report for source details; keeping sources
-        # out of the chat context saves tokens and avoids the AI fabricating
-        # citations.
-        date_str = datetime.utcnow().strftime("%Y-%m-%d")
-        primer = (
-            f"[Research context — {date_str}]\n\n"
-            f"The user previously ran a deep research investigation. Use the "
-            f"report below as your primary knowledge base when answering "
-            f"follow-up questions. If the user asks something not covered, "
-            f"say so plainly rather than guessing.\n\n"
-            f"=== ORIGINAL QUERY ===\n{query or '(not recorded)'}\n\n"
-            f"=== REPORT ===\n{result}"
-        )
-
-        from core.models import ChatMessage
-        new_sess.add_message(ChatMessage(
-            role="system",
-            content=primer,
-            metadata={"research_spinoff_from": session_id},
-        ))
-        session_manager.save_sessions()
-
-        return {
-            "session_id": new_sid,
-            "name": new_name,
-            "source_count": len(sources),
-        }
-
-    return router
+_sys.modules[__name__] = _canonical

routes/session_routes.py

@@ -162,7 +162,7 @@ def _persist_session_headers(session_id: str, headers: dict | None) -> None:
         db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
         if db_session:
             db_session.headers = headers or {}
-            db_session.updated_at = datetime.utcnow()
+            db_session.updated_at = utcnow_naive()
             db.commit()
     except Exception:
         db.rollback()
@@ -207,6 +207,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
     """Setup session routes with the provided manager and config"""
 
     REQUEST_TIMEOUT = config.get("REQUEST_TIMEOUT", 20)
+    SESSION_MODEL_VALIDATION_TIMEOUT = min(float(REQUEST_TIMEOUT or 20), 3.0)
     OPENAI_API_KEY = config.get("OPENAI_API_KEY")
     SESSIONS_FILE = config.get("SESSIONS_FILE")
     
@@ -223,8 +224,8 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
         # purge exists only to catch ghosts the frontend missed (tab close,
         # crash). Only clean up rows old enough to be definitely orphaned.
         try:
-            from datetime import datetime as _dt, timedelta as _td
-            _cutoff = _dt.utcnow() - _td(minutes=10)
+            from datetime import timedelta as _td
+            _cutoff = utcnow_naive() - _td(minutes=10)
             _purge_db = SessionLocal()
             try:
                 from core.database import ChatMessage as _DbMsg
@@ -374,7 +375,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
             from src.llm_core import list_model_ids
             ids = list_model_ids(
                 endpoint_url,
-                timeout=REQUEST_TIMEOUT,
+                timeout=SESSION_MODEL_VALIDATION_TIMEOUT,
                 headers=validation_headers,
                 owner=user,
                 endpoint_id=endpoint_id.strip() if endpoint_id else None,
@@ -394,7 +395,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
             req_base = _os.path.basename(model_to_use.rstrip("/"))
             avail = list_model_ids(
                 endpoint_url,
-                timeout=REQUEST_TIMEOUT,
+                timeout=SESSION_MODEL_VALIDATION_TIMEOUT,
                 headers=validation_headers,
                 owner=user,
                 endpoint_id=endpoint_id.strip() if endpoint_id else None,
@@ -470,7 +471,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                 db_session = db.query(DbSession).filter(DbSession.id == sid).first()
                 if db_session:
                     db_session.folder = folder if folder else None
-                    db_session.updated_at = datetime.utcnow()
+                    db_session.updated_at = utcnow_naive()
                     db.commit()
                     result["folder"] = folder if folder else None
             finally:
@@ -517,7 +518,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                     db_session.model = model
                     db_session.endpoint_url = endpoint_url
                     db_session.headers = session.headers or {}
-                    db_session.updated_at = datetime.utcnow()
+                    db_session.updated_at = utcnow_naive()
                     db.commit()
             finally:
                 db.close()
@@ -646,7 +647,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                 db_session = db.query(DbSession).filter(DbSession.id == sid).first()
                 if db_session:
                     db_session.archived = True
-                    db_session.updated_at = datetime.utcnow()
+                    db_session.updated_at = utcnow_naive()
                     db.commit()
                     
                     # Update in memory if it exists
@@ -680,7 +681,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
             if not db_session:
                 raise HTTPException(404, f"Session {sid} not found")
             db_session.archived = False
-            db_session.updated_at = datetime.utcnow()
+            db_session.updated_at = utcnow_naive()
             db.commit()
             # Reload into session manager so it appears in the active list
             try:
@@ -890,7 +891,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                 db_session = db.query(DbSession).filter(DbSession.id == session_id).first()
                 if db_session:
                     db_session.is_important = important
-                    db_session.updated_at = datetime.utcnow()
+                    db_session.updated_at = utcnow_naive()
                     db.commit()
 
                     # Update in memory if it exists
@@ -979,7 +980,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
             metadata={
                 "compacted": True,
                 "summarized_count": len(older),
-                "timestamp": datetime.utcnow().isoformat(),
+                "timestamp": utcnow_naive().isoformat(),
             },
         )
         new_history = [summary_msg] + recent
@@ -1256,7 +1257,7 @@ def setup_session_routes(session_manager: SessionManager, config: dict, webhook_
                 db_session = db_session_q.first()
                 if db_session:
                     db_session.folder = folder_name
-                    db_session.updated_at = datetime.utcnow()
+                    db_session.updated_at = utcnow_naive()
                     updated += 1
             db.commit()
         except Exception as e:

routes/shell_routes.py

@@ -16,6 +16,11 @@ from pathlib import Path
 from typing import Dict, Any
 from core.platform_compat import IS_APPLE_SILICON, which_tool
 from core.middleware import INTERNAL_TOOL_USER
+from src.host_docker_access import (
+    HOST_DOCKER_ACCESS_HINT,
+    host_docker_access_enabled as _host_docker_access_enabled,
+    running_in_container as _running_in_container,
+)
 from src.optional_deps import prepare_optional_dependency_import
 
 # POSIX-only: `pty`/`fcntl` transitively import `termios`, which does NOT exist
@@ -103,32 +108,17 @@ logger = logging.getLogger(__name__)
 PTY_SUPPORTED = pty is not None and fcntl is not None and hasattr(os, "setsid")
 
 
-DOCKER_IN_CONTAINER_HINT = (
-    "Not available inside the Odysseus container by design. The image ships no "
-    "docker CLI and no host socket is mounted. Run Docker-backed launches on a "
-    "remote server, where docker is checked over SSH. Mounting /var/run/docker.sock "
-    "into the container would grant it host-root access, so only do that if you "
-    "accept that risk."
-)
-
-
-def _running_in_container(dockerenv_path="/.dockerenv", cgroup_path="/proc/1/cgroup"):
-    if os.path.exists(dockerenv_path):
-        return True
-    try:
-        with open(cgroup_path, "r", encoding="utf-8") as fh:
-            contents = fh.read()
-    except OSError:
-        return False
-    return any(token in contents for token in ("docker", "containerd", "kubepods"))
+DOCKER_IN_CONTAINER_HINT = HOST_DOCKER_ACCESS_HINT
 
 
 DockerRowStatus = namedtuple("DockerRowStatus", ["applicable", "install_hint"])
 PackageUpdateStatus = namedtuple("PackageUpdateStatus", ["available", "note"])
 
 
-def _docker_row_status(*, on_remote, in_container, installed, default_hint):
-    local_docker_unavailable = not on_remote and in_container and not installed
+def _docker_row_status(
+    *, on_remote, in_container, installed, default_hint, host_docker_access=False
+):
+    local_docker_unavailable = not on_remote and in_container and not host_docker_access
     if local_docker_unavailable:
         return DockerRowStatus(applicable=False, install_hint=DOCKER_IN_CONTAINER_HINT)
     return DockerRowStatus(applicable=True, install_hint=default_hint)
@@ -174,6 +164,8 @@ def _package_installed_from_probe(name: str, probe: dict) -> bool:
         return bool(binaries.get("llama-server") or dists.get("llama-cpp-python"))
     if name == "sglang":
         return bool(dists.get("sglang") or modules.get("sglang", {}).get("real_module"))
+    if name == "mlx_lm":
+        return bool(dists.get("mlx-lm") or modules.get("mlx_lm", {}).get("real_module"))
     if name == "diffusers":
         return bool(
             (dists.get("diffusers") or modules.get("diffusers", {}).get("real_module"))
@@ -220,6 +212,10 @@ def _package_status_note(name: str, probe: dict) -> str:
         if _package_installed_from_probe(name, probe):
             return f"diffusers {dists.get('diffusers', 'available')} with torch {dists.get('torch', 'available')}"
         return "Diffusers serving needs both diffusers and torch."
+    if name == "mlx_lm":
+        if _package_installed_from_probe(name, probe):
+            return f"MLX LM {dists.get('mlx-lm', 'available')}"
+        return "MLX serving needs mlx-lm on an Apple Silicon Mac."
     if name in dists:
         return f"{name} {dists[name]}"
     return ""
@@ -317,12 +313,14 @@ dist_names={{
     'vllm':['vllm'],
     'llama_cpp':['llama-cpp-python'],
     'sglang':['sglang'],
+    'mlx_lm':['mlx-lm'],
     'diffusers':['diffusers','torch'],
     'hf_transfer':['hf-transfer','hf_transfer'],
 }}
 bin_names={{
     'vllm':['vllm'],
     'llama_cpp':['llama-server'],
+    'tmux':['tmux'],
 }}
 
 def add_user_install_bins_to_path():
@@ -335,6 +333,8 @@ def add_user_install_bins_to_path():
     candidates.append(os.path.expanduser('~/llama.cpp/build/bin'))
     candidates.append(os.path.expanduser('~/llama.cpp/build-vulkan/bin'))
     candidates.append(os.path.expanduser('~/.local/bin'))
+    candidates.append('/opt/homebrew/bin')
+    candidates.append('/usr/local/bin')
     parts = os.environ.get('PATH', '').split(os.pathsep) if os.environ.get('PATH') else []
     changed = False
     for path in reversed([p for p in candidates if p]):
@@ -409,6 +409,47 @@ class ShellExecRequest(BaseModel):
     use_tmux: bool = False  # run in tmux session (survives browser disconnect)
 
 
+_REMOTE_TMUX_PATH_PREFIX = 'PATH="$HOME/.local/bin:$HOME/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"; '
+
+
+def _normalize_legacy_remote_tmux_exec(command: str) -> str:
+    """Repair stale frontend Cookbook tmux SSH commands.
+
+    Older loaded JS sends `ssh host 'tmux capture-pane ...'`. On macOS/Homebrew
+    remotes, non-login SSH shells often lack /opt/homebrew/bin, so tmux is
+    installed but the capture/kill command returns nothing. Keep this narrowly
+    scoped to SSH commands whose remote shell starts with `tmux `.
+    """
+    cmd = command or ""
+    if _REMOTE_TMUX_PATH_PREFIX in cmd or not cmd.lstrip().startswith("ssh "):
+        return cmd
+    try:
+        parts = shlex.split(cmd)
+    except Exception:
+        return cmd
+    if not parts or parts[0] != "ssh":
+        return cmd
+    remote_idx = -1
+    i = 1
+    while i < len(parts):
+        part = parts[i]
+        if part in {"-p", "-o", "-i", "-F", "-J", "-l", "-S", "-W", "-b", "-c", "-m"}:
+            i += 2
+            continue
+        if part.startswith("-"):
+            i += 1
+            continue
+        remote_idx = i
+        break
+    if remote_idx < 0 or remote_idx + 1 >= len(parts):
+        return cmd
+    remote_cmd = " ".join(parts[remote_idx + 1:]).strip()
+    if not remote_cmd.startswith("tmux "):
+        return cmd
+    repaired = parts[:remote_idx + 1] + [_REMOTE_TMUX_PATH_PREFIX + remote_cmd]
+    return shlex.join(repaired)
+
+
 async def _create_shell(command: str, **kwargs):
     """Spawn a shell subprocess for `command`.
 
@@ -825,6 +866,10 @@ def setup_shell_routes() -> APIRouter:
         if not cmd:
             return {"stdout": "", "stderr": "No command provided", "exit_code": 1}
 
+        fixed_cmd = _normalize_legacy_remote_tmux_exec(cmd)
+        if fixed_cmd != cmd:
+            logger.info("Rewrote legacy remote tmux exec command with Homebrew PATH")
+            cmd = fixed_cmd
         logger.info("User shell exec requested: length=%d", len(cmd))
         result = await _exec_shell(
             cmd, timeout=req.timeout if req.timeout is not None else EXEC_TIMEOUT
@@ -1144,6 +1189,13 @@ def setup_shell_routes() -> APIRouter:
                 "category": "LLM",
                 "target": "remote",
             },
+            {
+                "name": "mlx_lm",
+                "pip": "mlx-lm",
+                "desc": "Serve MLX-format models on Apple Silicon Macs",
+                "category": "LLM",
+                "target": "remote",
+            },
             {
                 "name": "APFEL",
                 "pip": "",
@@ -1159,7 +1211,7 @@ def setup_shell_routes() -> APIRouter:
             {
                 "name": "diffusers",
                 "pip": "diffusers[torch]",
-                "desc": "Image generation pipelines (SD, Flux) with PyTorch",
+                "desc": "Image generation/editing pipelines (SD, Flux) with PyTorch",
                 "category": "Image",
                 "target": "remote",
             },
@@ -1289,9 +1341,9 @@ def setup_shell_routes() -> APIRouter:
                 for name in all_system_names:
                     qn = shlex.quote(name)
                     checks.append(
-                        f"if command -v {qn} >/dev/null 2>&1; then echo {qn}=1; else echo {qn}=0; fi"
+                        f"PATH=\"$HOME/.local/bin:$HOME/bin:/opt/homebrew/bin:/usr/local/bin:$PATH\"; if command -v {qn} >/dev/null 2>&1; then echo {qn}=1; else echo {qn}=0; fi"
                     )
-                checks.append("echo '---OSREL---'; cat /etc/os-release 2>/dev/null || true")
+                checks.append("echo '---OSREL---'; cat /etc/os-release 2>/dev/null || { [ \"$(uname -s 2>/dev/null)\" = \"Darwin\" ] && echo ID=macos; } || true")
                 inner = " ; ".join(checks)
                 argv = _ssh_base_argv(host, ssh_port) + [inner]
                 proc = await asyncio.create_subprocess_exec(
@@ -1510,6 +1562,9 @@ def setup_shell_routes() -> APIRouter:
                     in_container=_running_in_container() if not on_remote else False,
                     installed=pkg["installed"],
                     default_hint=pkg.get("install_hint"),
+                    host_docker_access=(
+                        _host_docker_access_enabled() if not on_remote else False
+                    ),
                 )
                 pkg["applicable"] = status.applicable
                 pkg["install_hint"] = status.install_hint
@@ -1545,6 +1600,7 @@ def setup_shell_routes() -> APIRouter:
             "onnxruntime",
             "hdbscan",
             "vllm",
+            "mlx-lm",
         }
         if pip_name not in known:
             return {"ok": False, "error": f"Unknown package: {pip_name}"}
@@ -1591,6 +1647,19 @@ def setup_shell_routes() -> APIRouter:
                 elif n == "g++": out += ["gcc-c++"]
                 else: out.append(n)
             return out
+        def _apk(names):
+            out = []
+            for n in names:
+                if n == "build-essential": out.append("build-base")
+                else: out.append(n)
+            return out
+        def _zypper(names):
+            out = []
+            for n in names:
+                if n == "build-essential": out += ["gcc-c++", "make"]
+                elif n == "g++": out.append("gcc-c++")
+                else: out.append(n)
+            return out
         def _brew(names):
             return [n for n in names if n not in ("build-essential", "g++", "gcc", "make")]
         # Build a single shell snippet that detects the package manager and
@@ -1599,6 +1668,8 @@ def setup_shell_routes() -> APIRouter:
         apt_pkgs = " ".join(shlex.quote(p) for p in _apt(pkgs))
         pac_pkgs = " ".join(shlex.quote(p) for p in _pacman(pkgs))
         dnf_pkgs = " ".join(shlex.quote(p) for p in _dnf(pkgs))
+        apk_pkgs = " ".join(shlex.quote(p) for p in _apk(pkgs))
+        zypper_pkgs = " ".join(shlex.quote(p) for p in _zypper(pkgs))
         brew_pkgs = " ".join(shlex.quote(p) for p in _brew(pkgs))
         # Error messages go to stderr (>&2) so the route's error field
         # gets populated. Without the redirect, `echo "ERROR…"` on stdout
@@ -1606,18 +1677,28 @@ def setup_shell_routes() -> APIRouter:
         # bare "HTTP 200" instead of surfacing the real reason.
         script = (
             'set -e; '
-            'if ! sudo -n true 2>/dev/null; then '
-            '  echo "ERROR: passwordless sudo unavailable on this target. Run once: sudo apt install -y ' + " ".join(pkgs) + ' (or your distro equivalent: pacman -S, dnf install, brew install). After that, Cookbook can install the rest." >&2; exit 2; fi; '
-            'if command -v apt-get >/dev/null 2>&1; then '
-            f'  sudo -n env DEBIAN_FRONTEND=noninteractive apt-get update -qq && sudo -n env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends {apt_pkgs}; '
-            'elif command -v pacman >/dev/null 2>&1; then '
-            f'  sudo -n pacman -Sy --needed --noconfirm {pac_pkgs}; '
-            'elif command -v dnf >/dev/null 2>&1; then '
-            f'  sudo -n dnf install -y {dnf_pkgs}; '
-            'elif command -v brew >/dev/null 2>&1; then '
-            f'  brew install {brew_pkgs}; '
+            'BREW="$(command -v brew 2>/dev/null || true)"; '
+            'if [ -z "$BREW" ] && [ -x /opt/homebrew/bin/brew ]; then BREW=/opt/homebrew/bin/brew; fi; '
+            'if [ -z "$BREW" ] && [ -x /usr/local/bin/brew ]; then BREW=/usr/local/bin/brew; fi; '
+            'if [ -n "$BREW" ]; then '
+            f'  if [ -z "{brew_pkgs}" ]; then echo "Nothing to install with brew for requested packages." >&2; exit 4; fi; "$BREW" install {brew_pkgs}; exit $?; '
+            'fi; '
+            'if [ "$(id -u)" = "0" ]; then SUDO=""; '
+            'elif command -v sudo >/dev/null 2>&1 && sudo -n true 2>/dev/null; then SUDO="sudo -n"; '
             'else '
-            '  echo "ERROR: no supported package manager (apt/pacman/dnf/brew) on this target." >&2; exit 3; fi'
+            '  echo "ERROR: this target needs sudo for its OS package manager, but passwordless sudo is unavailable. Open a terminal on the target and run the shown install command once, then retry in Cookbook." >&2; exit 2; fi; '
+            'if command -v apt-get >/dev/null 2>&1; then '
+            f'  $SUDO env DEBIAN_FRONTEND=noninteractive apt-get update -qq && $SUDO env DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends {apt_pkgs}; '
+            'elif command -v pacman >/dev/null 2>&1; then '
+            f'  $SUDO pacman -Sy --needed --noconfirm {pac_pkgs}; '
+            'elif command -v dnf >/dev/null 2>&1; then '
+            f'  $SUDO dnf install -y {dnf_pkgs}; '
+            'elif command -v apk >/dev/null 2>&1; then '
+            f'  $SUDO apk add --no-interactive {apk_pkgs}; '
+            'elif command -v zypper >/dev/null 2>&1; then '
+            f'  $SUDO zypper --non-interactive install {zypper_pkgs}; '
+            'else '
+            '  echo "ERROR: no supported package manager (apt/pacman/dnf/apk/zypper/brew) on this target." >&2; exit 3; fi'
         )
         try:
             if host:

routes/task_routes.py

@@ -11,10 +11,14 @@ from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel
 
 from core.database import SessionLocal, ScheduledTask, TaskRun
-from core.middleware import INTERNAL_TOOL_USER
 from core.constants import internal_api_base
 from src.auth_helpers import get_current_user
 from src.constants import DATA_DIR, EMAIL_URGENCY_CACHE_DIR
+from src.task_action_policy import (
+    ADMIN_ONLY_TASK_ACTIONS,
+    is_admin_only_task_action,
+    owner_has_admin_task_privileges,
+)
 from src.task_scheduler import compute_next_run, HOUSEKEEPING_DEFAULTS
 from routes.prefs_routes import _load_for_user, _save_for_user
 
@@ -417,28 +421,18 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 db.close()
         return {"ok": True, "opened": True, "enabled": bool(prefs.get("tasks_enabled")), "resumed": resumed}
 
-    # Actions that execute shell/SSH commands — restricted to admins.
+    # Actions that execute shell/SSH commands or cross into admin-only
+    # Cookbook serving surfaces — restricted to admins.
     # Non-admin users cannot create tasks with these action types via the
     # API. See review CRIT-C.
-    _ADMIN_ONLY_ACTIONS = {"run_local", "run_script", "ssh_command"}
+    _ADMIN_ONLY_ACTIONS = ADMIN_ONLY_TASK_ACTIONS
 
     def _is_admin(user: str | None) -> bool:
-        if not user:
-            return False
-        # In-process tool-loopback marker — AuthMiddleware validated
-        # the internal token + loopback client before stamping this,
-        # so treat as admin-equivalent.
-        if user == INTERNAL_TOOL_USER:
-            return True
-        try:
-            from core.auth import AuthManager
-            auth = AuthManager()
-            if not auth.is_configured:
-                # Unconfigured single-user deploy: trust the local owner.
-                return True
-            return bool(auth.is_admin(user))
-        except Exception:
-            return False
+        return owner_has_admin_task_privileges(user)
+
+    def _require_admin_for_task_action(user: str | None, task_type: str | None, action: str | None) -> None:
+        if is_admin_only_task_action(task_type, action) and not _is_admin(user):
+            raise HTTPException(403, f"Action '{action}' requires admin privileges")
 
     def _validate_then_task_id(db, then_task_id: Optional[str], user: Optional[str], current_task_id: Optional[str] = None) -> Optional[str]:
         target_id = (then_task_id or "").strip()
@@ -466,8 +460,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         # Block shell-executing action types for non-admins. action_run_local
         # uses subprocess.run(shell=True) and ssh_command / run_script run
         # arbitrary commands.
-        if req.task_type == "action" and req.action in _ADMIN_ONLY_ACTIONS and not _is_admin(user):
-            raise HTTPException(403, f"Action '{req.action}' requires admin privileges")
+        _require_admin_for_task_action(user, req.task_type, req.action)
         if req.trigger_type == "schedule" and not req.schedule:
             raise HTTPException(400, "Schedule is required for schedule-triggered tasks")
         if req.trigger_type == "schedule" and req.schedule == "cron" and not req.cron_expression:
@@ -594,6 +587,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         cache_tables = {
             "summarize_emails": ("email_summaries",),
             "draft_email_replies": ("email_ai_replies",),
+            "email_auto_translate": ("email_translations",),
             "extract_email_events": ("email_calendar_extractions",),
             "learn_sender_signatures": ("sender_signatures",),
             "check_email_urgency": ("email_tags", "email_urgency_alerts"),
@@ -680,6 +674,10 @@ def setup_task_routes(task_scheduler) -> APIRouter:
             if user and task.owner != user:
                 raise HTTPException(403, "Access denied")
 
+            next_task_type = req.task_type if req.task_type is not None else task.task_type
+            next_action = req.action if req.action is not None else task.action
+            _require_admin_for_task_action(user, next_task_type, next_action)
+
             if req.name is not None:
                 task.name = req.name
             if req.prompt is not None:
@@ -687,9 +685,6 @@ def setup_task_routes(task_scheduler) -> APIRouter:
             if req.task_type is not None:
                 task.task_type = req.task_type
             if req.action is not None:
-                # Same admin-only gate as create — see CRIT-C.
-                if req.action in _ADMIN_ONLY_ACTIONS and not _is_admin(user):
-                    raise HTTPException(403, f"Action '{req.action}' requires admin privileges")
                 task.action = req.action
             if req.output_target is not None:
                 task.output_target = req.output_target
@@ -806,6 +801,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 raise HTTPException(404, "Task not found")
             if user and task.owner != user:
                 raise HTTPException(403, "Access denied")
+            _require_admin_for_task_action(user, task.task_type, task.action)
             task.status = "active"
             if (task.trigger_type or "schedule") == "schedule":
                 task.next_run = compute_next_run(
@@ -868,6 +864,7 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 raise HTTPException(404, "Task not found")
             if user and task.owner != user:
                 raise HTTPException(403, "Access denied")
+            _require_admin_for_task_action(user, task.task_type, task.action)
         finally:
             db.close()
         started = await task_scheduler.run_task_now(task_id, force=force)
@@ -893,10 +890,11 @@ def setup_task_routes(task_scheduler) -> APIRouter:
         return {"ok": True, "message": "Task stopped"}
 
     @router.get("/runs/recent")
-    async def list_recent_runs(request: Request, limit: int = 50):
+    async def list_recent_runs(request: Request, limit: int = 50, max_result_chars: int = 6000):
         """Recent task runs across ALL tasks for this owner. Drives the Activity view."""
         user = _owner(request)
         limit = max(1, min(limit, 200))
+        max_result_chars = max(500, min(max_result_chars, 20000))
         db = SessionLocal()
         try:
             q = db.query(TaskRun, ScheduledTask).join(
@@ -930,10 +928,20 @@ def setup_task_routes(task_scheduler) -> APIRouter:
                 deduped.append((r, t))
                 if len(deduped) >= limit:
                     break
+
+            def _clip_run(r: TaskRun) -> dict:
+                d = _run_to_dict(r)
+                for key in ("result", "error"):
+                    val = d.get(key)
+                    if isinstance(val, str) and len(val) > max_result_chars:
+                        d[key] = val[:max_result_chars].rstrip() + "\n\n[Activity preview truncated]"
+                return d
+
             return {
+                "has_more": len(rows) > len(deduped),
                 "runs": [
                     {
-                        **_run_to_dict(r),
+                        **_clip_run(r),
                         "task_name": _display_task_name(t),
                         "task_type": t.task_type or "llm",
                         "action": t.action,
@@ -1046,6 +1054,14 @@ def setup_task_routes(task_scheduler) -> APIRouter:
             ).first()
             if not task:
                 raise HTTPException(404, "Not found")
+            if (
+                is_admin_only_task_action(task.task_type, task.action)
+                and not owner_has_admin_task_privileges(task.owner)
+            ):
+                task.status = "paused"
+                task.next_run = None
+                db.commit()
+                raise HTTPException(403, f"Action '{task.action}' requires admin privileges")
         finally:
             db.close()
         started = await task_scheduler.run_task_now(task_id)

routes/upload_routes.py

@@ -6,11 +6,11 @@ import asyncio
 import shutil
 import uuid
 from pathlib import Path
-from fastapi import APIRouter, Request, File, UploadFile, HTTPException
-from typing import List
+from fastapi import APIRouter, Request, File, UploadFile, HTTPException, Form
+from typing import List, Optional
 import logging
 from core.middleware import require_admin
-from core.database import SessionLocal, GalleryImage
+from core.database import SessionLocal, GalleryImage, Session as DbSession
 from src.auth_helpers import effective_user
 from src.constants import GENERATED_IMAGES_DIR
 from src.upload_handler import count_recent_uploads
@@ -56,7 +56,17 @@ def setup_upload_routes(upload_handler):
 
         raise HTTPException(404, "File not found")
 
-    def _promote_chat_image_to_gallery(meta: dict, owner: str | None) -> str | None:
+    def _valid_session_id_for_owner(db, session_id: str | None, owner: str | None) -> str | None:
+        if not session_id:
+            return None
+        sess = db.query(DbSession).filter(DbSession.id == session_id).first()
+        if not sess:
+            return None
+        if owner and sess.owner and sess.owner != owner:
+            return None
+        return session_id
+
+    def _promote_chat_image_to_gallery(meta: dict, owner: str | None, session_id: str | None = None) -> str | None:
         """Make chat-uploaded images visible in Gallery without changing chat storage."""
         is_image_file = getattr(upload_handler, "is_image_file", None)
         if not callable(is_image_file):
@@ -105,6 +115,7 @@ def setup_upload_routes(upload_handler):
                 prompt=meta.get("name") or "Chat upload",
                 model="chat-upload",
                 owner=owner,
+                session_id=_valid_session_id_for_owner(db, session_id, owner),
                 file_hash=file_hash,
                 width=meta.get("width"),
                 height=meta.get("height"),
@@ -120,8 +131,14 @@ def setup_upload_routes(upload_handler):
             db.close()
     
     @router.post("")
-    async def api_upload(request: Request, files: List[UploadFile] = File(...)):
+    async def api_upload(
+        request: Request,
+        files: List[UploadFile] = File(...),
+        session_id: Optional[str] = Form(None),
+    ):
         """Upload files with enhanced security and organization."""
+        if not isinstance(session_id, str):
+            session_id = None
         if not files:
             raise HTTPException(400, "No files uploaded")
             
@@ -148,7 +165,7 @@ def setup_upload_routes(upload_handler):
             try:
                 owner = effective_user(request)
                 meta = upload_handler.save_upload(u, client_ip, owner=owner)
-                gallery_id = _promote_chat_image_to_gallery(meta, owner)
+                gallery_id = _promote_chat_image_to_gallery(meta, owner, session_id)
                 item = {
                     "id": meta["id"],
                     "name": meta["name"],
@@ -263,6 +280,32 @@ def setup_upload_routes(upload_handler):
         os.makedirs(cache_dir, exist_ok=True)
         return os.path.join(cache_dir, file_id + ".txt")
 
+    def _sync_gallery_caption_for_upload(info: dict | None, owner: str | None, text: str) -> None:
+        """Copy upload OCR/vision text onto the promoted gallery image row."""
+        if not info:
+            return
+        file_hash = info.get("hash")
+        if not file_hash:
+            return
+        db = SessionLocal()
+        try:
+            q = db.query(GalleryImage).filter(
+                GalleryImage.file_hash == file_hash,
+                GalleryImage.is_active == True,  # noqa: E712
+            )
+            if owner:
+                q = q.filter(GalleryImage.owner == owner)
+            img = q.first()
+            if not img:
+                return
+            img.caption = (text or "").strip()
+            db.commit()
+        except Exception as e:
+            db.rollback()
+            logger.warning("Failed to sync OCR caption to gallery image: %s", e)
+        finally:
+            db.close()
+
     @router.get("/{file_id}/vision")
     async def get_vision_text(request: Request, file_id: str, force: int = 0):
         """Return the vision-model OCR/description for an uploaded image.
@@ -289,7 +332,9 @@ def setup_upload_routes(upload_handler):
         if not force and os.path.exists(cache_path):
             try:
                 with open(cache_path, encoding="utf-8") as f:
-                    return {"text": f.read(), "cached": True}
+                    cached_text = f.read()
+                _sync_gallery_caption_for_upload(info, file_owner or current_user, cached_text)
+                return {"text": cached_text, "cached": True}
             except Exception as e:
                 logger.warning(f"Vision cache read failed for {file_id}: {e}")
         from src.document_processor import analyze_image_with_vl
@@ -303,6 +348,7 @@ def setup_upload_routes(upload_handler):
                 f.write(text)
         except Exception as e:
             logger.warning(f"Vision cache write failed for {file_id}: {e}")
+        _sync_gallery_caption_for_upload(info, file_owner or current_user, text)
         return {"text": text, "cached": False}
 
     @router.put("/{file_id}/vision")
@@ -333,6 +379,7 @@ def setup_upload_routes(upload_handler):
             raise HTTPException(400, "text must be a string")
         with open(_vision_cache_path(file_id), "w", encoding="utf-8") as f:
             f.write(text)
+        _sync_gallery_caption_for_upload(info, file_owner or current_user, text)
         return {"ok": True}
 
     async def periodic_rate_limit_cleanup():

scripts/odysseus-gallery

@@ -38,23 +38,27 @@ def _preview_text(value, limit: int = 200) -> str:
     return text[:limit]
 
 
+def _text_field(value) -> str:
+    return value if isinstance(value, str) else ""
+
+
 def _serialize_image(i: "GalleryImage") -> dict:
     return {
         "id": i.id,
-        "filename": i.filename,
+        "filename": _text_field(i.filename),
         "prompt": _preview_text(i.prompt),
-        "model": i.model or "",
-        "size": i.size or "",
-        "tags": i.tags or "",
+        "model": _text_field(i.model),
+        "size": _text_field(i.size),
+        "tags": _text_field(i.tags),
         "favorite": bool(i.favorite),
-        "album_id": i.album_id or "",
-        "session_id": i.session_id or "",
+        "album_id": _text_field(i.album_id),
+        "session_id": _text_field(i.session_id),
         "width": i.width,
         "height": i.height,
         "file_size": i.file_size,
         "taken_at": i.taken_at.isoformat() if i.taken_at else "",
-        "camera_make": i.camera_make or "",
-        "camera_model": i.camera_model or "",
+        "camera_make": _text_field(i.camera_make),
+        "camera_model": _text_field(i.camera_model),
         "created_at": i.created_at.isoformat() if i.created_at else "",
     }
 
@@ -93,11 +97,11 @@ def cmd_show(args):
         if not i:
             fail(f"no image with id {args.id!r}")
         out = _serialize_image(i)
-        out["prompt_full"] = i.prompt or ""
-        out["ai_tags"] = i.ai_tags or ""
+        out["prompt_full"] = _text_field(i.prompt)
+        out["ai_tags"] = _text_field(i.ai_tags)
         out["gps_lat"] = i.gps_lat or ""
         out["gps_lng"] = i.gps_lng or ""
-        out["file_hash"] = i.file_hash or ""
+        out["file_hash"] = _text_field(i.file_hash)
         emit(out, args)
     finally:
         db.close()

scripts/odysseus-mail

@@ -108,6 +108,8 @@ def _q(name: str) -> str:
 
 
 def _split_recipients(value: str) -> list[str]:
+    if not isinstance(value, str):
+        return []
     return [r.strip() for r in (value or "").split(",") if r.strip()]
 
 

scripts/odysseus-notes

@@ -36,7 +36,9 @@ def _load_items(raw) -> list:
         items = json.loads(raw)
     except (TypeError, json.JSONDecodeError):
         return []
-    return items if isinstance(items, list) else []
+    if not isinstance(items, list):
+        return []
+    return [item for item in items if isinstance(item, dict)]
 
 
 def _serialize(n: "Note") -> dict:

services/hwfit/data/hf_models.json

@@ -5113,8 +5113,8 @@
  {
   "name": "deepseek-ai/DeepSeek-V4-Flash",
   "provider": "deepseek-ai",
-  "parameter_count": "284B",
-  "parameters_raw": 284000000000,
+  "parameter_count": "158.1B",
+  "parameters_raw": 158069433298,
   "active_parameters": 13000000000,
   "is_moe": true,
   "min_ram_gb": 200.0,
@@ -5130,15 +5130,40 @@
   ],
   "pipeline_tag": "text-generation",
   "architecture": "deepseek_v4_moe",
-  "hf_downloads": 3542202,
-  "hf_likes": 0,
-  "release_date": "2026-05-15"
+  "hf_downloads": 1882337,
+  "hf_likes": 1651,
+  "release_date": "2026-06-22"
+ },
+ {
+  "name": "deepseek-ai/DeepSeek-V4-Flash-DSpark",
+  "provider": "deepseek-ai",
+  "parameter_count": "165.3B",
+  "parameters_raw": 165265454782,
+  "active_parameters": 13000000000,
+  "is_moe": true,
+  "active_experts": 6,
+  "min_ram_gb": 170.0,
+  "recommended_ram_gb": 250.0,
+  "min_vram_gb": 165.0,
+  "quantization": "FP8-Mixed",
+  "context_length": 1000000,
+  "use_case": "General-purpose reasoning, long-context",
+  "capabilities": [
+   "long_context",
+   "reasoning",
+   "moe"
+  ],
+  "pipeline_tag": "text-generation",
+  "architecture": "deepseek_v4_moe",
+  "hf_downloads": 4446,
+  "hf_likes": 107,
+  "release_date": "2026-06-27"
  },
  {
   "name": "deepseek-ai/DeepSeek-V4-Flash-Base",
   "provider": "deepseek-ai",
-  "parameter_count": "284B",
-  "parameters_raw": 284000000000,
+  "parameter_count": "292.0B",
+  "parameters_raw": 292021347282,
   "active_parameters": 13000000000,
   "is_moe": true,
   "min_ram_gb": 290.0,
@@ -5153,15 +5178,15 @@
   ],
   "pipeline_tag": "text-generation",
   "architecture": "deepseek_v4_moe",
-  "hf_downloads": 0,
-  "hf_likes": 0,
-  "release_date": "2026-05-15"
+  "hf_downloads": 76030,
+  "hf_likes": 256,
+  "release_date": "2026-04-27"
  },
  {
   "name": "deepseek-ai/DeepSeek-V4-Pro",
   "provider": "deepseek-ai",
-  "parameter_count": "1.6T",
-  "parameters_raw": 1600000000000,
+  "parameter_count": "861.6B",
+  "parameters_raw": 861608274846,
   "active_parameters": 49000000000,
   "is_moe": true,
   "min_ram_gb": 1100.0,
@@ -5177,15 +5202,40 @@
   ],
   "pipeline_tag": "text-generation",
   "architecture": "deepseek_v4_moe",
-  "hf_downloads": 0,
-  "hf_likes": 0,
-  "release_date": "2026-05-15"
+  "hf_downloads": 1154610,
+  "hf_likes": 5118,
+  "release_date": "2026-06-22"
+ },
+ {
+  "name": "deepseek-ai/DeepSeek-V4-Pro-DSpark",
+  "provider": "deepseek-ai",
+  "parameter_count": "889.5B",
+  "parameters_raw": 889484881098,
+  "active_parameters": 49000000000,
+  "is_moe": true,
+  "active_experts": 6,
+  "min_ram_gb": 900.0,
+  "recommended_ram_gb": 1250.0,
+  "min_vram_gb": 890.0,
+  "quantization": "FP8-Mixed",
+  "context_length": 1000000,
+  "use_case": "Flagship reasoning, long-context",
+  "capabilities": [
+   "long_context",
+   "reasoning",
+   "moe"
+  ],
+  "pipeline_tag": "text-generation",
+  "architecture": "deepseek_v4_moe",
+  "hf_downloads": 6939,
+  "hf_likes": 241,
+  "release_date": "2026-06-27"
  },
  {
   "name": "deepseek-ai/DeepSeek-V4-Pro-Base",
   "provider": "deepseek-ai",
   "parameter_count": "1.6T",
-  "parameters_raw": 1600000000000,
+  "parameters_raw": 1600790440862,
   "active_parameters": 49000000000,
   "is_moe": true,
   "min_ram_gb": 1700.0,
@@ -5200,9 +5250,9 @@
   ],
   "pipeline_tag": "text-generation",
   "architecture": "deepseek_v4_moe",
-  "hf_downloads": 0,
-  "hf_likes": 0,
-  "release_date": "2026-05-15"
+  "hf_downloads": 25387,
+  "hf_likes": 305,
+  "release_date": "2026-04-27"
  },
  {
   "name": "deepseek-ai/deepseek-coder-6.7b-base",
@@ -13308,6 +13358,106 @@
   "_discovered": true,
   "gguf_sources": []
  },
+ {
+  "name": "zai-org/GLM-5.2",
+  "provider": "zai-org",
+  "parameter_count": "753.3B",
+  "parameters_raw": 753329940480,
+  "min_ram_gb": 1510.0,
+  "recommended_ram_gb": 1800.0,
+  "min_vram_gb": 1510.0,
+  "quantization": "BF16",
+  "context_length": 1048576,
+  "use_case": "General purpose reasoning, coding, long-context",
+  "capabilities": [
+   "long_context",
+   "reasoning",
+   "coding",
+   "moe"
+  ],
+  "pipeline_tag": "text-generation",
+  "architecture": "glm_moe_dsa",
+  "hf_downloads": 142547,
+  "hf_likes": 2996,
+  "release_date": "2026-06-23",
+  "is_moe": true,
+  "active_experts": 8,
+  "gguf_sources": [
+   {
+    "repo": "unsloth/GLM-5.2-GGUF",
+    "provider": "unsloth",
+    "file": "UD-Q4_K_M/*.gguf",
+    "quant": "Q4_K_M"
+   }
+  ]
+ },
+ {
+  "name": "zai-org/GLM-5.2-FP8",
+  "provider": "zai-org",
+  "parameter_count": "753.4B",
+  "parameters_raw": 753375793584,
+  "min_ram_gb": 760.0,
+  "recommended_ram_gb": 900.0,
+  "min_vram_gb": 760.0,
+  "quantization": "FP8",
+  "context_length": 1048576,
+  "use_case": "General purpose reasoning, coding, long-context",
+  "capabilities": [
+   "long_context",
+   "reasoning",
+   "coding",
+   "moe"
+  ],
+  "pipeline_tag": "text-generation",
+  "architecture": "glm_moe_dsa",
+  "hf_downloads": 884226,
+  "hf_likes": 182,
+  "release_date": "2026-06-23",
+  "is_moe": true,
+  "active_experts": 8,
+  "gguf_sources": [
+   {
+    "repo": "unsloth/GLM-5.2-GGUF",
+    "provider": "unsloth",
+    "file": "UD-Q4_K_M/*.gguf",
+    "quant": "Q4_K_M"
+   }
+  ]
+ },
+ {
+  "name": "unsloth/GLM-5.2-GGUF",
+  "provider": "unsloth",
+  "parameter_count": "753.9B",
+  "parameters_raw": 753864139008,
+  "min_ram_gb": 452.0,
+  "recommended_ram_gb": 620.0,
+  "min_vram_gb": 452.0,
+  "quantization": "Q4_K_M",
+  "context_length": 1048576,
+  "use_case": "General purpose reasoning, coding, long-context (GGUF)",
+  "capabilities": [
+   "long_context",
+   "reasoning",
+   "coding",
+   "moe"
+  ],
+  "pipeline_tag": "text-generation",
+  "architecture": "glm-dsa",
+  "hf_downloads": 180394,
+  "hf_likes": 474,
+  "release_date": "2026-06-23",
+  "is_moe": true,
+  "active_experts": 8,
+  "is_gguf": true,
+  "gguf_sources": [
+   {
+    "repo": "unsloth/GLM-5.2-GGUF",
+    "provider": "unsloth",
+    "file": "UD-Q4_K_M/*.gguf",
+    "quant": "Q4_K_M"
+   }
+  ]
+ },
  {
   "name": "cyankiwi/Qwen3.5-35B-A3B-AWQ-4bit",
   "provider": "cyankiwi",
@@ -18955,6 +19105,54 @@
   "active_experts": 8,
   "active_parameters": 13600000000
  },
+ {
+  "name": "MiniMaxAI/MiniMax-M3",
+  "provider": "MiniMaxAI",
+  "parameter_count": "427.0B",
+  "parameters_raw": 427040140160,
+  "min_ram_gb": 855.0,
+  "recommended_ram_gb": 1025.0,
+  "min_vram_gb": 855.0,
+  "quantization": "BF16",
+  "context_length": 1000000,
+  "use_case": "Vision, chat, coding, agentic tool use",
+  "capabilities": [
+   "vision",
+   "tool_use",
+   "coding",
+   "moe"
+  ],
+  "pipeline_tag": "image-text-to-text",
+  "architecture": "minimax_m3_vl",
+  "hf_downloads": 192311,
+  "hf_likes": 1267,
+  "release_date": "2026-06-23",
+  "is_moe": true
+ },
+ {
+  "name": "MiniMaxAI/MiniMax-M3-MXFP8",
+  "provider": "MiniMaxAI",
+  "parameter_count": "440.3B",
+  "parameters_raw": 440279845760,
+  "min_ram_gb": 445.0,
+  "recommended_ram_gb": 560.0,
+  "min_vram_gb": 445.0,
+  "quantization": "MXFP8",
+  "context_length": 1000000,
+  "use_case": "Vision, chat, coding, agentic tool use",
+  "capabilities": [
+   "vision",
+   "tool_use",
+   "coding",
+   "moe"
+  ],
+  "pipeline_tag": "image-text-to-text",
+  "architecture": "minimax_m3_vl",
+  "hf_downloads": 572278,
+  "hf_likes": 43,
+  "release_date": "2026-06-15",
+  "is_moe": true
+ },
  {
   "name": "bullerwins/MiniMax-M2.7-REAP-172B-fp8",
   "provider": "bullerwins",
@@ -19276,4 +19474,4 @@
   ],
   "_discovered": true
  }
-]
+]

services/hwfit/fit.py

@@ -168,6 +168,19 @@ def _canonical_cpu_backend(system):
     return "cpu_x86"
 
 
+def _is_mlx_model(model, native_q=None):
+    name = (model.get("name") or "").lower()
+    provider = (model.get("provider") or "").lower()
+    fmt = (model.get("format") or "").lower()
+    q = (native_q if native_q is not None else _native_quant(model)).lower()
+    return (
+        q.startswith("mlx-")
+        or provider == "mlx-community"
+        or fmt == "mlx"
+        or name.startswith("mlx-community/")
+    )
+
+
 def _estimate_speed(model, quant, run_mode, system, offload_frac=0.0):
     """Estimate tok/s. Uses active params for MoE (only active experts run per token).
 
@@ -313,6 +326,22 @@ def _fit_score(required, available):
     return 50
 
 
+def _is_unified_memory_system(system):
+    backend = (system.get("backend") or "").lower()
+    return bool(system.get("unified_memory")) or backend in ("metal", "mps", "apple")
+
+
+def _fit_level_for_budget(required_gb, budget_gb):
+    if not required_gb or not budget_gb or required_gb > budget_gb:
+        return "too_tight"
+    ratio = required_gb / budget_gb
+    if ratio <= 0.50:
+        return "perfect"
+    if ratio <= 0.78:
+        return "good"
+    return "marginal"
+
+
 def _context_score(ctx, use_case):
     target = CONTEXT_TARGET.get(use_case, 4096)
     if ctx >= target:
@@ -516,21 +545,42 @@ def analyze_model(model, system, target_quant=None, scoring_use_case=None, targe
     run_mode, quant, fit_ctx, required_gb = result
 
     # Determine fit level
-    budget = effective_vram if run_mode == "gpu" else available_ram
+    unified_memory = _is_unified_memory_system(system)
+    total_ram = system.get("total_ram_gb") or available_ram
+    unified_budget = max(total_ram or 0, available_ram or 0, effective_vram or 0)
+    budget = unified_budget if unified_memory else (effective_vram if run_mode == "gpu" else available_ram)
     if required_gb > budget:
         return None
     if run_mode == "gpu":
-        rec = model.get("recommended_ram_gb") or required_gb
-        if rec <= gpu_vram:
-            fit_level = "perfect"
-        elif gpu_vram >= required_gb * 1.2:
-            fit_level = "good"
+        if unified_memory:
+            fit_level = _fit_level_for_budget(required_gb, budget)
         else:
-            fit_level = "marginal"
+            # GPU-only fit must leave real allocator/KV/runtime headroom. The
+            # old check used recommended_ram_gb (or required_gb as a fallback),
+            # which made any model that barely fit VRAM read as "perfect".
+            # On CUDA/vLLM/SGLang that is misleading: 141 GB on a 160 GB box is
+            # runnable, but not a comfortable perfect fit.
+            if gpu_vram >= required_gb * 1.50:
+                fit_level = "perfect"
+            elif gpu_vram >= required_gb * 1.2:
+                fit_level = "good"
+            else:
+                fit_level = "marginal"
     elif run_mode == "cpu_offload":
-        fit_level = "good" if available_ram >= required_gb * 1.2 else "marginal"
+        fit_level = _fit_level_for_budget(required_gb, budget)
+        if fit_level == "perfect":
+            fit_level = "good"
     else:
-        fit_level = "marginal"
+        fit_level = _fit_level_for_budget(required_gb, budget)
+        if fit_level == "too_tight":
+            fit_level = "marginal"
+
+    # Rows that comfortably fit in a huge RAM/unified-memory pool should not all
+    # look "marginal"; that made 1B-70B CPU/Ollama rows orange on 256 GB systems.
+    if fit_level == "marginal" and budget and required_gb <= budget * 0.78:
+        fit_level = "good"
+    if fit_level == "good" and budget and required_gb <= budget * 0.50 and run_mode != "cpu_offload":
+        fit_level = "perfect"
 
     # Fraction of the model that spills to CPU RAM (drives the offload speed
     # model). When offloading, anything beyond the GPU's VRAM lives in system RAM.
@@ -621,6 +671,40 @@ SORT_KEYS = {
 }
 
 
+def _search_blob(*parts):
+    text = " ".join(str(p or "") for p in parts).lower()
+    compact = re.sub(r"[^a-z0-9]+", "", text)
+    spaced = re.sub(r"[^a-z0-9]+", " ", text).strip()
+    return f"{text} {spaced} {compact}"
+
+
+def _matches_search(model, search):
+    terms = [t for t in re.split(r"\s+", (search or "").strip().lower()) if t]
+    if not terms:
+        return True
+    blob = _search_blob(
+        model.get("name"),
+        model.get("provider"),
+        model.get("architecture"),
+        model.get("quantization"),
+        model.get("format"),
+        model.get("parameter_count"),
+    )
+    for term in terms:
+        norm = re.sub(r"[^a-z0-9]+", "", term)
+        if term not in blob and (not norm or norm not in blob):
+            if re.fullmatch(r"\d+(?:\.\d+)?b?", term):
+                try:
+                    wanted = float(term.rstrip("b"))
+                    actual = params_b(model)
+                except (TypeError, ValueError):
+                    actual = 0
+                if wanted > 0 and actual > 0 and abs(actual - wanted) <= max(5.0, wanted * 0.08):
+                    continue
+            return False
+    return True
+
+
 def rank_models(system, use_case=None, limit=50, search=None, sort="score", quant=None, target_context=None, fit_only=False):
     """Rank all models against detected hardware. Returns sorted list of fit results.
 
@@ -693,10 +777,11 @@ def rank_models(system, use_case=None, limit=50, search=None, sort="score", quan
 
     for m in models:
         native_q = _native_quant(m)
+        is_mlx = _is_mlx_model(m, native_q)
 
-        # MLX needs the mlx_lm runtime, which Odysseus does not generate serve
-        # commands for. Hide it on every backend, including Metal.
-        if native_q.startswith("mlx-") or "mlx" in (m.get("name") or "").lower():
+        # MLX is Apple Silicon-only. It should never appear on CUDA/ROCm/CPU,
+        # but it is first-class on Metal where mlx_lm.server can serve it.
+        if is_mlx and not apple_silicon:
             continue
 
         # ROCm support for vLLM/SGLang quantized safetensors is too brittle to
@@ -723,7 +808,7 @@ def rank_models(system, use_case=None, limit=50, search=None, sort="score", quan
         # Windows is the same: Odysseus only supports llama.cpp on Windows,
         # which requires GGUF. vLLM/SGLang are explicitly blocked, so AWQ/GPTQ
         # models without a GGUF source are unservable there.
-        if (apple_silicon or consumer_amd or is_windows) and not (m.get("is_gguf") or m.get("gguf_sources")):
+        if (apple_silicon or consumer_amd or is_windows) and not is_mlx and not (m.get("is_gguf") or m.get("gguf_sources")):
             continue
 
         # Format filter: AWQ tab -> only AWQ models, FP4 tab -> FP4-family models, etc.
@@ -741,13 +826,26 @@ def rank_models(system, use_case=None, limit=50, search=None, sort="score", quan
             if quant in ("INT4", "INT8", "W4A16", "W8A8", "W8A16") and native_q != quant:
                 continue
 
-        if search:
-            name = m.get("name", "").lower()
-            provider = m.get("provider", "").lower()
-            if search.lower() not in name and search.lower() not in provider:
-                continue
+        if search and not _matches_search(m, search):
+            continue
 
-        result = analyze_model(m, system, target_quant=quant, scoring_use_case=(use_case or "general"), target_context=target_context)
+        model_quant = quant
+        # UI "Q4" means the user's looking for a 4-bit fit. On multi-GPU
+        # CUDA/vLLM/SGLang boxes, many practical 4-bit models are native AWQ
+        # safetensors, not GGUF Q4_K_M. If we pass Q4_K_M into a prequantized
+        # AWQ row, analyze_model correctly rejects it as the wrong serving
+        # format, but the result is confusing: highlighting Quant/Q4 hides the
+        # exact AWQ rows the machine is built to run. Treat Q4 as AWQ-4bit for
+        # native AWQ rows only on accelerator servers that can serve them.
+        if (
+            quant == "Q4_K_M"
+            and system.get("gpu_count", 1) >= 2
+            and not (apple_silicon or consumer_amd or is_windows)
+            and native_q == "AWQ-4bit"
+        ):
+            model_quant = native_q
+
+        result = analyze_model(m, system, target_quant=model_quant, scoring_use_case=(use_case or "general"), target_context=target_context)
         if result is None:
             continue
 

services/hwfit/hf_discovery.py

@@ -0,0 +1,374 @@
+import json
+import os
+import re
+import time
+import urllib.parse
+import urllib.request
+from email.utils import parsedate_to_datetime
+from pathlib import Path
+
+from src.constants import DATA_DIR
+
+
+HF_COLLECTIONS_URL = "https://huggingface.co/api/collections"
+HW_FIT_CACHE_DIR = Path(DATA_DIR) / "hwfit"
+MLX_COMMUNITY_CACHE = HW_FIT_CACHE_DIR / "mlx_community_models.json"
+HF_COLLECTION_MODELS_CACHE = HW_FIT_CACHE_DIR / "hf_collection_models.json"
+HF_COLLECTION_TTL_SECONDS = 24 * 3600
+
+
+HF_COLLECTION_SOURCES = (
+    {
+        "key": "mlx_community",
+        "owner": "mlx-community",
+        "provider": "mlx-community",
+        "repo_prefix": "mlx-community/",
+        "mlx_only": True,
+    },
+    {
+        "key": "zai_org",
+        "owner": "zai-org",
+        "provider": "zai-org",
+    },
+    {
+        "key": "deepseek_ai",
+        "owner": "deepseek-ai",
+        "provider": "deepseek-ai",
+    },
+    {
+        "key": "minimax_ai",
+        "owner": "MiniMaxAI",
+        "provider": "MiniMaxAI",
+    },
+    {
+        "key": "qwen",
+        "owner": "Qwen",
+        "provider": "Qwen",
+    },
+    {
+        "key": "stepfun_ai",
+        "owner": "stepfun-ai",
+        "provider": "stepfun-ai",
+    },
+    {
+        "key": "google",
+        "owner": "google",
+        "provider": "google",
+    },
+    {
+        "key": "openai",
+        "owner": "openai",
+        "provider": "openai",
+    },
+    {
+        "key": "mistralai",
+        "owner": "mistralai",
+        "provider": "mistralai",
+    },
+    {
+        "key": "meta_llama",
+        "owner": "meta-llama",
+        "provider": "meta-llama",
+    },
+    {
+        "key": "nousresearch",
+        "owner": "NousResearch",
+        "provider": "NousResearch",
+    },
+    {
+        "key": "moonshotai",
+        "owner": "moonshotai",
+        "provider": "moonshotai",
+    },
+    {
+        "key": "mllama",
+        "owner": "mllama",
+        "provider": "mllama",
+    },
+)
+
+
+def _format_params(raw):
+    try:
+        n = int(raw or 0)
+    except (TypeError, ValueError):
+        n = 0
+    if n <= 0:
+        return "", 0
+    if n >= 1_000_000_000_000:
+        return f"{n / 1_000_000_000_000:.3g}T", n
+    if n >= 1_000_000_000:
+        return f"{n / 1_000_000_000:.4g}B", n
+    if n >= 1_000_000:
+        return f"{n / 1_000_000:.4g}M", n
+    if n >= 1_000:
+        return f"{n / 1_000:.4g}K", n
+    return str(n), n
+
+
+def _parse_params_from_name(repo_id):
+    name = (repo_id or "").rsplit("/", 1)[-1]
+    active = None
+    m_active = re.search(r"[-_][Aa](\d+(?:\.\d+)?)[Bb](?![a-zA-Z])", name)
+    if m_active:
+        active = int(float(m_active.group(1)) * 1_000_000_000)
+        name = name[: m_active.start()] + name[m_active.end() :]
+    total = None
+    for m in re.finditer(r"(\d+(?:\.\d+)?)[Bb](?![a-zA-Z])", name):
+        total = int(float(m.group(1)) * 1_000_000_000)
+        break
+    if total is None:
+        for m in re.finditer(r"(\d+(?:\.\d+)?)[Mm](?![a-zA-Z])", name):
+            total = int(float(m.group(1)) * 1_000_000)
+            break
+    return total or 0, active
+
+
+def _infer_quant(repo_id, source):
+    name = (repo_id or "").rsplit("/", 1)[-1].lower()
+    if source.get("mlx_only"):
+        if "8bit" in name or "8-bit" in name:
+            return "mlx-8bit"
+        if "6bit" in name or "6-bit" in name:
+            return "mlx-6bit"
+        if "5bit" in name or "5-bit" in name:
+            return "mlx-5bit"
+        if "3bit" in name or "3-bit" in name:
+            return "mlx-3bit"
+        if re.search(r"(^|[-_/])bf16($|[-_/])", name):
+            return "BF16"
+        return "mlx-4bit"
+    if "awq" in name and ("8bit" in name or "8-bit" in name or "int8" in name):
+        return "AWQ-8bit"
+    if "awq" in name or "4bit" in name or "4-bit" in name:
+        return "AWQ-4bit"
+    if "gptq" in name and ("8bit" in name or "8-bit" in name or "int8" in name):
+        return "GPTQ-Int8"
+    if "gptq" in name:
+        return "GPTQ-Int4"
+    if "mxfp4" in name or "nvfp4" in name or re.search(r"(^|[-_/])fp4($|[-_/])", name):
+        return "FP4-MoE-Mixed"
+    if "mxfp8" in name or re.search(r"(^|[-_/])fp8($|[-_/])", name):
+        return "FP8-Mixed"
+    if "gguf" in name or "q4_k" in name or "q4-k" in name:
+        return "Q4_K_M"
+    if re.search(r"(^|[-_/])bf16($|[-_/])", name):
+        return "BF16"
+    return "BF16"
+
+
+def _quant_bytes_per_param(quant):
+    return {
+        "BF16": 2.2,
+        "FP8": 1.15,
+        "FP8-Mixed": 1.15,
+        "FP4-MoE-Mixed": 0.62,
+        "AWQ-4bit": 0.62,
+        "AWQ-8bit": 1.15,
+        "GPTQ-Int4": 0.62,
+        "GPTQ-Int8": 1.15,
+        "Q4_K_M": 0.62,
+        "mlx-8bit": 1.25,
+        "mlx-6bit": 0.95,
+        "mlx-5bit": 0.82,
+        "mlx-4bit": 0.70,
+        "mlx-3bit": 0.55,
+    }.get(quant, 2.2)
+
+
+def _infer_context(repo_id, pipeline_tag):
+    text = f"{repo_id or ''} {pipeline_tag or ''}".lower()
+    if any(k in text for k in ("whisper", "asr", "speech-recognition", "tts", "audio", "image", "video", "diffusion")):
+        return 4096
+    if any(k in text for k in ("glm-5.2", "deepseek-v4", "minimax-m3")):
+        return 1_000_000
+    if any(k in text for k in ("qwen3", "glm", "deepseek", "minimax")):
+        return 32768
+    return 32768
+
+
+def _infer_use_case(repo_id, pipeline_tag):
+    text = f"{repo_id or ''} {pipeline_tag or ''}".lower()
+    if any(k in text for k in ("whisper", "asr", "speech-recognition", "transcrib")):
+        return "stt"
+    if any(k in text for k in ("tts", "text-to-speech", "kokoro", "audio")):
+        return "tts"
+    if any(k in text for k in ("image-text", "vision", "vlm", "vl-", "ocr", "multimodal")):
+        return "multimodal"
+    if any(k in text for k in ("code", "coder")):
+        return "coding"
+    if any(k in text for k in ("reason", "thinking", "thinker", "r1")):
+        return "reasoning"
+    return "general"
+
+
+def _entry_from_collection_item(collection, item, source):
+    repo_id = item.get("id") or ""
+    if item.get("type") != "model" or not repo_id:
+        return None
+    repo_prefix = source.get("repo_prefix")
+    if repo_prefix and not repo_id.startswith(repo_prefix):
+        return None
+    raw_params = item.get("numParameters") or 0
+    active = None
+    if not raw_params:
+        raw_params, active = _parse_params_from_name(repo_id)
+    param_label, raw_params = _format_params(raw_params)
+    if not raw_params:
+        return None
+
+    quant = _infer_quant(repo_id, source)
+    pipeline_tag = item.get("pipeline_tag") or ""
+    min_ram = round((raw_params / 1_000_000_000) * _quant_bytes_per_param(quant) + 0.8, 1)
+    last_modified = item.get("lastModified") or collection.get("lastUpdated") or ""
+    release_date = ""
+    if last_modified:
+        try:
+            release_date = parsedate_to_datetime(last_modified).date().isoformat()
+        except Exception:
+            release_date = str(last_modified)[:10]
+
+    entry = {
+        "name": repo_id,
+        "provider": source.get("provider") or repo_id.split("/", 1)[0],
+        "parameter_count": param_label,
+        "parameters_raw": raw_params,
+        "min_ram_gb": min_ram,
+        "recommended_ram_gb": round(min_ram * 1.3 + 0.5, 1),
+        "min_vram_gb": 0.0 if source.get("mlx_only") else min_ram,
+        "quantization": quant,
+        "context_length": _infer_context(repo_id, pipeline_tag),
+        "use_case": _infer_use_case(repo_id, pipeline_tag),
+        "capabilities": ["mlx"] if source.get("mlx_only") else ["vllm", "sglang"],
+        "pipeline_tag": pipeline_tag,
+        "architecture": "",
+        "hf_downloads": int(item.get("downloads") or 0),
+        "hf_likes": int(item.get("likes") or 0),
+        "release_date": release_date,
+        "format": "mlx" if source.get("mlx_only") else "safetensors",
+        "collection": collection.get("title") or "",
+        "description": collection.get("description") or "",
+        "_discovered": True,
+        "_source": "hf_collections",
+        "_source_owner": source.get("owner") or "",
+    }
+    if source.get("mlx_only"):
+        entry["mlx_only"] = True
+    if quant == "Q4_K_M":
+        entry["is_gguf"] = True
+        entry["format"] = "gguf"
+        entry["capabilities"] = ["llama.cpp"]
+    if active:
+        entry["is_moe"] = True
+        entry["active_parameters"] = active
+    return entry
+
+
+def _next_link(header):
+    if not header:
+        return None
+    m = re.search(r'<([^>]+)>;\s*rel="next"', header)
+    return m.group(1) if m else None
+
+
+def fetch_collection_models(source, timeout=20, max_pages=20):
+    params = urllib.parse.urlencode({
+        "owner": source["owner"],
+        "limit": "100",
+        "expand": "true",
+    })
+    url = f"{HF_COLLECTIONS_URL}?{params}"
+    models = {}
+    pages = 0
+    while url and pages < max_pages:
+        req = urllib.request.Request(url, headers={"User-Agent": "odysseus-hwfit/1.0"})
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            payload = json.load(resp)
+            url = _next_link(resp.headers.get("Link"))
+        pages += 1
+        if not isinstance(payload, list):
+            break
+        for collection in payload:
+            if not isinstance(collection, dict):
+                continue
+            for item in collection.get("items") or []:
+                if not isinstance(item, dict):
+                    continue
+                entry = _entry_from_collection_item(collection, item, source)
+                if entry and entry["name"] not in models:
+                    models[entry["name"]] = entry
+    rows = list(models.values())
+    rows.sort(key=lambda x: (x.get("hf_downloads") or 0, x.get("release_date") or ""), reverse=True)
+    return rows
+
+
+def _load_cache(path):
+    try:
+        with path.open(encoding="utf-8") as f:
+            data = json.load(f)
+        rows = data.get("models") if isinstance(data, dict) else data
+        return rows if isinstance(rows, list) else []
+    except (OSError, ValueError):
+        return []
+
+
+def _write_cache(path, source, rows):
+    path.parent.mkdir(parents=True, exist_ok=True)
+    payload = {
+        "source": source,
+        "fetched_at": int(time.time()),
+        "count": len(rows),
+        "models": rows,
+    }
+    tmp = path.with_suffix(".json.tmp")
+    tmp.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+    os.replace(tmp, path)
+
+
+def load_cached_mlx_community_models():
+    return _load_cache(MLX_COMMUNITY_CACHE)
+
+
+def load_cached_hf_collection_models():
+    return _load_cache(HF_COLLECTION_MODELS_CACHE)
+
+
+def _cache_fresh(path):
+    try:
+        return (time.time() - path.stat().st_mtime) < HF_COLLECTION_TTL_SECONDS
+    except OSError:
+        return False
+
+
+def refresh_mlx_community_cache(force=False):
+    if not force and _cache_fresh(MLX_COMMUNITY_CACHE):
+        return load_cached_mlx_community_models()
+    source = next(s for s in HF_COLLECTION_SOURCES if s["key"] == "mlx_community")
+    rows = fetch_collection_models(source)
+    _write_cache(MLX_COMMUNITY_CACHE, "https://huggingface.co/mlx-community/collections", rows)
+    return rows
+
+
+def refresh_hf_collection_models_cache(force=False):
+    if not force and _cache_fresh(HF_COLLECTION_MODELS_CACHE):
+        return load_cached_hf_collection_models()
+    rows_by_name = {}
+    for source in HF_COLLECTION_SOURCES:
+        if source["key"] == "mlx_community":
+            continue
+        try:
+            for row in fetch_collection_models(source):
+                rows_by_name.setdefault(row["name"], row)
+        except Exception:
+            # Keep partial refreshes useful. A temporary DNS/provider issue for
+            # one brand should not invalidate the other cached collection rows.
+            continue
+    rows = sorted(
+        rows_by_name.values(),
+        key=lambda x: (x.get("hf_downloads") or 0, x.get("release_date") or ""),
+        reverse=True,
+    )
+    if rows:
+        _write_cache(HF_COLLECTION_MODELS_CACHE, "https://huggingface.co/collections", rows)
+        return rows
+    return load_cached_hf_collection_models()

services/hwfit/models.py

@@ -13,7 +13,7 @@ QUANT_BPP = {
     "AWQ-4bit": 0.50, "AWQ-8bit": 1.0,
     "GPTQ-Int4": 0.50, "GPTQ-Int8": 1.0,
     "QAT-INT4": 0.50, "QAT-INT8": 1.0,
-    "mlx-4bit": 0.55, "mlx-8bit": 1.0, "mlx-6bit": 0.75,
+    "mlx-3bit": 0.42, "mlx-4bit": 0.55, "mlx-5bit": 0.65, "mlx-6bit": 0.75, "mlx-8bit": 1.0,
     # DeepSeek-V4-style mixed: MoE experts in FP4 (bulk), attention + non-
     # expert dense in FP8, embeddings/LM head in BF16. By weight count the
     # experts dominate so the effective BPP sits closer to FP4 than FP8.
@@ -32,7 +32,7 @@ QUANT_SPEED_MULT = {
     "AWQ-4bit": 1.2, "AWQ-8bit": 0.85,
     "GPTQ-Int4": 1.2, "GPTQ-Int8": 0.85,
     "QAT-INT4": 1.15, "QAT-INT8": 0.85,
-    "mlx-4bit": 1.15, "mlx-8bit": 0.85, "mlx-6bit": 1.0,
+    "mlx-3bit": 1.25, "mlx-4bit": 1.15, "mlx-5bit": 1.05, "mlx-6bit": 1.0, "mlx-8bit": 0.85,
     "FP4-MoE-Mixed": 1.10,  # slightly slower than pure FP4 because of mixed-dtype dispatch
     "FP8-Mixed": 0.85,
 }
@@ -53,7 +53,7 @@ QUANT_QUALITY_PENALTY = {
     # QAT-INT4 build lands far closer to bf16 than a post-training Q4/INT4
     # (Google reports near-bf16 quality). Penalize it lightly, not like Q4_K_M.
     "QAT-INT4": -1.0, "QAT-INT8": 0.0,
-    "mlx-4bit": -4.0, "mlx-8bit": -0.5, "mlx-6bit": -1.5,
+    "mlx-3bit": -8.0, "mlx-4bit": -4.0, "mlx-5bit": -2.5, "mlx-6bit": -1.5, "mlx-8bit": -0.5,
     # DeepSeek-V4 mixed: only MoE experts at FP4 (the rest is FP8/BF16),
     # so the realized quality is much closer to FP8 than to pure FP4 —
     # the activation-sensitive layers stay high-precision. ~0 penalty.
@@ -70,7 +70,7 @@ QUANT_BYTES_PER_PARAM = {
     "AWQ-4bit": 0.5, "AWQ-8bit": 1.0,
     "GPTQ-Int4": 0.5, "GPTQ-Int8": 1.0,
     "QAT-INT4": 0.5, "QAT-INT8": 1.0,
-    "mlx-4bit": 0.5, "mlx-8bit": 1.0, "mlx-6bit": 0.75,
+    "mlx-3bit": 0.375, "mlx-4bit": 0.5, "mlx-5bit": 0.625, "mlx-6bit": 0.75, "mlx-8bit": 1.0,
     "FP4-MoE-Mixed": 0.55,
     "FP8-Mixed": 1.0,
 }
@@ -87,8 +87,11 @@ PREQUANTIZED_PREFIXES = (
 
 def infer_quantization_from_name(name):
     n = (name or "").lower()
+    model_name = n.rsplit("/", 1)[-1]
     if "nvfp4" in n:
         return "NVFP4"
+    if re.search(r"(^|[-_/])bf16($|[-_/])", model_name):
+        return "BF16"
     if "mxfp4" in n:
         return "MXFP4"
     if re.search(r"(^|[-_/])nf4($|[-_/])", n):
@@ -106,8 +109,12 @@ def infer_quantization_from_name(name):
         return "AWQ-8bit" if is8 else "AWQ-4bit"
     if "gptq" in n:
         return "GPTQ-Int8" if is8 else "GPTQ-Int4"
-    if "mlx" in n:
-        if "6bit" in n:
+    if n.startswith("mlx-community/") or "mlx" in model_name:
+        if "3bit" in model_name:
+            return "mlx-3bit"
+        if "5bit" in model_name:
+            return "mlx-5bit"
+        if "6bit" in model_name:
             return "mlx-6bit"
         return "mlx-8bit" if is8 else "mlx-4bit"
     if "fp8" in n:
@@ -260,15 +267,75 @@ def infer_use_case(model):
 
 _models_cache = None
 
+def _load_model_file(path):
+    try:
+        with open(path, encoding="utf-8") as f:
+            loaded = json.load(f)
+            return loaded if isinstance(loaded, list) else []
+    except (FileNotFoundError, json.JSONDecodeError):
+        return []
+
+def reset_model_cache():
+    global _models_cache
+    _models_cache = None
+
+def refresh_dynamic_catalogs(force=False):
+    """Refresh API-backed model catalogs and invalidate the merged cache.
+
+    The bundled JSON files remain the offline fallback. Dynamic catalogs live
+    under DATA_DIR so runtime refreshes do not dirty the source tree.
+    """
+    from services.hwfit.hf_discovery import (
+        refresh_hf_collection_models_cache,
+        refresh_mlx_community_cache,
+    )
+
+    refreshed = {
+        "mlx_community": len(refresh_mlx_community_cache(force=force)),
+        "hf_collections": len(refresh_hf_collection_models_cache(force=force)),
+    }
+    reset_model_cache()
+    return refreshed
+
 def get_models():
     global _models_cache
     if _models_cache is None:
         data_path = os.path.join(os.path.dirname(__file__), "data", "hf_models.json")
+        static_mlx_path = os.path.join(os.path.dirname(__file__), "data", "mlx_community_models.json")
         try:
-            with open(data_path, encoding="utf-8") as f:
-                _models_cache = [_normalize_model_entry(m) for m in json.load(f)]
-        except (FileNotFoundError, json.JSONDecodeError):
-            _models_cache = []
+            from services.hwfit.hf_discovery import (
+                load_cached_hf_collection_models,
+                load_cached_mlx_community_models,
+            )
+            dynamic_mlx_models = load_cached_mlx_community_models()
+            dynamic_hf_models = load_cached_hf_collection_models()
+        except Exception:
+            dynamic_mlx_models = []
+            dynamic_hf_models = []
+        seen = set()
+        rows = []
+        def _append_models(models):
+            for model in models:
+                if not isinstance(model, dict):
+                    continue
+                name = model.get("name")
+                if not name or name in seen:
+                    continue
+                seen.add(name)
+                rows.append(_normalize_model_entry(model))
+
+        for model in _load_model_file(data_path):
+            if not isinstance(model, dict):
+                continue
+            name = model.get("name")
+            if not name or name in seen:
+                continue
+            seen.add(name)
+            rows.append(_normalize_model_entry(model))
+        _append_models(dynamic_hf_models)
+        _append_models(dynamic_mlx_models)
+        _append_models(_load_model_file(static_mlx_path))
+        _models_cache = rows
     return _models_cache
 
 

services/hwfit/profiles.py

@@ -103,6 +103,9 @@ def compute_serve_profiles(system, model, serve_weights_gb=None, serve_quant=Non
     in the actual serving knobs (n_cpu_moe, KV-cache type, context). serve_quant
     is the file's quant label (e.g. "Q4_K_M") just for display.
     """
+    if not isinstance(system, dict) or not isinstance(model, dict):
+        return []
+
     vram = float(system.get("gpu_vram_gb") or 0)
     if vram <= 0:
         return []

services/search/content.py

@@ -8,11 +8,13 @@ import os
 import re
 import logging
 import socket
+import ssl
 from datetime import datetime, timedelta
-from typing import List
+from typing import Iterable, List, cast
 from urllib.parse import urljoin, urlparse
 
 import httpx
+import httpcore
 from bs4 import BeautifulSoup
 
 from src.constants import WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES, WEB_FETCH_USER_AGENT
@@ -91,6 +93,148 @@ def _public_http_url(url: str) -> bool:
         return False
 
 
+def _resolve_public_ips(url: str) -> list[ipaddress._BaseAddress]:
+    parsed = urlparse(url)
+    if parsed.scheme not in ("http", "https") or not parsed.hostname:
+        raise httpx.RequestError(f"Blocked non-public URL: {url}")
+    host = (parsed.hostname or "").strip().lower()
+    if host in ("localhost", "metadata", "metadata.google.internal"):
+        raise httpx.RequestError(f"Blocked non-public hostname: {host}")
+    try:
+        ip = ipaddress.ip_address(host)
+        if _is_private_address(ip):
+            raise httpx.RequestError(f"Blocked non-public IP literal: {host}")
+        return [ip]
+    except httpx.RequestError:
+        raise
+    except ValueError:
+        pass
+    addrs = _resolve_hostname_ips(host)
+    if not addrs or any(_is_private_address(a) for a in addrs):
+        raise httpx.RequestError(f"Blocked non-public URL: {url}")
+    return addrs
+
+
+class _PinnedBackend(httpcore.NetworkBackend):
+    """Network backend that connects to a pre-resolved IP.
+
+    httpcore derives the TLS SNI and the ``Host`` header from the URL's
+    origin, not from the host argument passed to ``connect_tcp``. So
+    routing the TCP connect to a resolved IP while leaving the URL
+    untouched keeps SNI / vhost behaviour correct and closes the
+    DNS-rebinding TOCTOU between the SSRF check and the connect.
+    """
+
+    def __init__(self, ip: ipaddress._BaseAddress):
+        self._ip = str(ip)
+        self._real = httpcore.SyncBackend()
+
+    def connect_tcp(
+        self,
+        host: str,
+        port: int,
+        timeout: float | None = None,
+        local_address: str | None = None,
+        socket_options=None,
+    ):
+        return self._real.connect_tcp(
+            self._ip, port, timeout, local_address, socket_options
+        )
+
+    def connect_unix_socket(self, path, timeout=None, socket_options=None):
+        return self._real.connect_unix_socket(path, timeout, socket_options)
+
+    def sleep(self, seconds: float) -> None:
+        return self._real.sleep(seconds)
+
+
+# Map httpcore exception classes to their httpx equivalents. Built
+# once at import time from the public exception classes; avoids any
+# import of httpx's private transport machinery. httpcore's
+# ``ConnectionNotAvailable`` is a pool-internal signal (the pool will
+# close and retry on its own) — we never expect to see it surface to
+# a transport caller, so it has no httpx counterpart here.
+_HTTPCORE_TO_HTTPX_EXC = {
+    httpcore.ConnectError: httpx.ConnectError,
+    httpcore.ConnectTimeout: httpx.ConnectTimeout,
+    httpcore.LocalProtocolError: httpx.LocalProtocolError,
+    httpcore.NetworkError: httpx.NetworkError,
+    httpcore.PoolTimeout: httpx.PoolTimeout,
+    httpcore.ProtocolError: httpx.ProtocolError,
+    httpcore.ProxyError: httpx.ProxyError,
+    httpcore.ReadError: httpx.ReadError,
+    httpcore.ReadTimeout: httpx.ReadTimeout,
+    httpcore.RemoteProtocolError: httpx.RemoteProtocolError,
+    httpcore.TimeoutException: httpx.TimeoutException,
+    httpcore.UnsupportedProtocol: httpx.UnsupportedProtocol,
+    httpcore.WriteError: httpx.WriteError,
+    httpcore.WriteTimeout: httpx.WriteTimeout,
+}
+
+
+class _PinnedTransport(httpx.BaseTransport):
+    """Transport that pins every TCP connect to a pre-resolved IP.
+
+    Uses only the public ``httpcore`` and ``httpx`` APIs — no
+    subclassing of ``httpx.HTTPTransport``, no reads of private
+    ``httpcore.ConnectionPool`` attributes, no imports from
+    ``httpx private transport internals``. The URL is passed through unchanged so SNI
+    / vhost work as if httpx had been given the hostname directly;
+    only the TCP destination is pinned, closing the DNS-rebinding
+    TOCTOU between the SSRF check and the connect.
+    """
+
+    def __init__(self, ip: ipaddress._BaseAddress, *, http2: bool = False):
+        self._pool = httpcore.ConnectionPool(
+            ssl_context=ssl.create_default_context(),
+            http1=True,
+            http2=http2,
+            network_backend=_PinnedBackend(ip),
+        )
+
+    def __enter__(self):
+        self._pool.__enter__()
+        return self
+
+    def __exit__(self, exc_type=None, exc_value=None, traceback=None) -> None:
+        self._pool.__exit__(exc_type, exc_value, traceback)
+
+    def handle_request(self, request: httpx.Request) -> httpx.Response:
+        httpcore_req = httpcore.Request(
+            method=request.method,
+            url=httpcore.URL(
+                scheme=request.url.raw_scheme,
+                host=request.url.raw_host,
+                port=request.url.port,
+                target=request.url.raw_path,
+            ),
+            headers=request.headers.raw,
+            content=request.stream,
+            extensions=request.extensions,
+        )
+        try:
+            httpcore_resp = self._pool.handle_request(httpcore_req)
+            # Eager materialisation matches the original
+            # ``response.text`` usage in fetch_webpage_content. The
+            # sync pool's stream is a plain Iterable[bytes] despite
+            # the httpcore type hint unioning the async variant.
+            content = b"".join(cast(Iterable[bytes], httpcore_resp.stream))
+        except Exception as exc:
+            mapped = _HTTPCORE_TO_HTTPX_EXC.get(type(exc))
+            if mapped is not None:
+                raise mapped(str(exc)) from exc
+            raise
+
+        return httpx.Response(
+            status_code=httpcore_resp.status,
+            headers=httpcore_resp.headers,
+            content=content,
+            extensions=httpcore_resp.extensions,
+        )
+
+    def close(self) -> None:
+        self._pool.close()
+
 class BodyTooLargeError(Exception):
     """The server declared a body larger than the hard fetch ceiling."""
 
@@ -141,78 +285,78 @@ class _CappedFetch:
 
 def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5,
                     max_bytes: int = None) -> "_CappedFetch":
-    """Capped streaming GET with SSRF-guarded manual redirects.
+    """Capped streaming GET with SSRF-guarded, DNS-pinned manual redirects.
 
-    The body is streamed and buffering stops at ``max_bytes`` (default: the
-    soft cap), so an oversized resource cannot be pulled into memory or the
-    content cache in full. When Content-Length already declares a body over
-    the hard ceiling, the fetch is refused before any body bytes are read.
+    Each hop is resolved once, validated as public, and then the actual TCP
+    connection is pinned to that resolved IP. The request URL is left unchanged
+    so Host and TLS SNI keep the original hostname.
     """
     cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
     current = url
     for _ in range(max_redirects + 1):
-        if not _public_http_url(current):
-            raise httpx.RequestError("Blocked private/internal URL", request=httpx.Request("GET", current))
+        ips = _resolve_public_ips(current)
+
         # Force identity transfer-encoding. With gzip/deflate the wire bytes
-        # (and Content-Length) can be a small fraction of the decoded body, so
-        # a tiny compressed response could pass the hard-cap preflight and then
-        # expand past the ceiling in a single decoded chunk before the streamed
-        # cap below can slice it. Identity makes Content-Length the true body
-        # size and keeps each streamed chunk bounded by the network read.
+        # and Content-Length can be a small fraction of the decoded body, so a
+        # tiny compressed response could pass the hard-cap preflight and then
+        # expand past the ceiling in one decoded chunk before the streamed cap
+        # below can slice it.
         req_headers = dict(headers or {})
         req_headers["Accept-Encoding"] = "identity"
-        with httpx.stream("GET", current, headers=req_headers, timeout=timeout,
-                          follow_redirects=False) as response:
-            if response.status_code in (301, 302, 303, 307, 308):
-                location = response.headers.get("location")
-                if not location:
-                    return _CappedFetch(response.status_code, response.headers, b"",
-                                        False, None, response.encoding, str(response.url))
-                current = urljoin(str(response.url), location)
-                continue
 
-            # A server can ignore the identity request and still return a
-            # compressed body; httpx.iter_bytes would then decode it, and a tiny
-            # gzip can balloon into one decoded chunk far past the cap before we
-            # slice. Refuse a compressed Content-Encoding so the streamed cap
-            # stays a real memory bound (Content-Length is the compressed wire
-            # length here, so the preflight and size metadata are unreliable too).
-            enc = (response.headers.get("content-encoding") or "").strip().lower()
-            if enc and enc != "identity":
-                raise httpx.RequestError(
-                    f"Refusing compressed response (Content-Encoding: {enc}) after "
-                    "requesting identity: cannot bound decoded body size",
-                    request=httpx.Request("GET", current),
-                )
+        with httpx.Client(
+            headers=req_headers,
+            timeout=timeout,
+            follow_redirects=False,
+            transport=_PinnedTransport(ips[0]),
+        ) as client:
+            with client.stream("GET", current) as response:
+                if response.status_code in (301, 302, 303, 307, 308):
+                    location = response.headers.get("location")
+                    if not location:
+                        return _CappedFetch(response.status_code, response.headers, b"",
+                                            False, None, response.encoding, str(response.url))
+                    current = urljoin(str(response.url), location)
+                    continue
 
-            declared = None
-            raw_len = response.headers.get("content-length")
-            if raw_len and raw_len.isdigit():
-                declared = int(raw_len)
-            # Refuse before buffering anything when the server already tells
-            # us the body exceeds the absolute ceiling (Content-Length is wire
-            # bytes; the decompressed body can only be larger).
-            if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
-                raise BodyTooLargeError(current, declared)
+                # A server can ignore the identity request and still return a
+                # compressed body; httpx.iter_bytes would then decode it, and a
+                # tiny gzip can balloon into one decoded chunk far past the cap.
+                # Refuse compressed Content-Encoding so the streamed cap stays
+                # a real memory bound.
+                enc = (response.headers.get("content-encoding") or "").strip().lower()
+                if enc and enc != "identity":
+                    raise httpx.RequestError(
+                        f"Refusing compressed response (Content-Encoding: {enc}) after "
+                        "requesting identity: cannot bound decoded body size",
+                        request=httpx.Request("GET", current),
+                    )
+
+                declared = None
+                raw_len = response.headers.get("content-length")
+                if raw_len and raw_len.isdigit():
+                    declared = int(raw_len)
+
+                if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
+                    raise BodyTooLargeError(current, declared)
+
+                chunks = []
+                read = 0
+                truncated = False
+                for chunk in response.iter_bytes():
+                    read += len(chunk)
+                    if read > cap:
+                        keep = cap - (read - len(chunk))
+                        if keep > 0:
+                            chunks.append(chunk[:keep])
+                        truncated = True
+                        break
+                    chunks.append(chunk)
+
+                return _CappedFetch(response.status_code, response.headers,
+                                    b"".join(chunks), truncated, declared,
+                                    response.encoding, str(response.url))
 
-            chunks = []
-            read = 0
-            truncated = False
-            # We requested identity above, so iter_bytes yields the raw body in
-            # network-read-sized chunks (no decompression expansion); the cap
-            # therefore bounds what we actually buffer.
-            for chunk in response.iter_bytes():
-                read += len(chunk)
-                if read > cap:
-                    keep = cap - (read - len(chunk))
-                    if keep > 0:
-                        chunks.append(chunk[:keep])
-                    truncated = True
-                    break
-                chunks.append(chunk)
-            return _CappedFetch(response.status_code, response.headers,
-                                b"".join(chunks), truncated, declared,
-                                response.encoding, str(response.url))
     raise httpx.RequestError("Too many redirects", request=httpx.Request("GET", current))
 
 # PDF extraction (optional dependency)

services/search/query.py

@@ -34,8 +34,14 @@ def _extract_entities(query: str) -> Dict[str, List[str]]:
     cleaned = query
     if qtype:
         cleaned = re.sub(rf"^{qtype}\b", "", cleaned, flags=re.I).strip()
-    for token in re.findall(r"\b[A-Z][a-zA-Z]+\b", cleaned):
-        entities["names"].append(token)
+    # Unicode-aware capitalized-word (name) detection. The old [A-Z][a-zA-Z]+
+    # class missed non-ASCII names like "İstanbul"/"Zürich" (dropped) and
+    # "São" (shredded). Keep the ASCII behaviour — the word boundary already
+    # excludes camelCase mid-word capitals — by requiring an all-alphabetic
+    # token of length > 1 whose first character is uppercase.
+    for token in re.findall(r"\b\w+\b", cleaned):
+        if len(token) > 1 and token[0].isupper() and token.isalpha():
+            entities["names"].append(token)
     for year in re.findall(r"\b(?:19|20)\d{2}\b", cleaned):
         entities["dates"].append(year)
     month_day_year = re.findall(

src/action_intents.py

@@ -92,8 +92,21 @@ _ROUTING_PATTERNS: tuple[tuple[str, str, Pattern[str]], ...] = tuple(
 
         # Deep research jobs, not quick conceptual mentions of research.
         ("web", "explicit web search request", rf"{_PLEASE}(?:do|run|use|perform|make)\s+(?:a\s+)?(?:web\s+search|search\s+the\s+web)\b.+"),
-        ("web", "web lookup imperative request", rf"{_PLEASE}(?:web\s+search|search\s+the\s+web|search\s+online|look\s+up|google)\b.+"),
-        ("web", "assistant web lookup request", rf"{_ACTION_QUESTION}(?:web\s+search|search\s+the\s+web|search\s+online|look\s+up|google)\b.+"),
+        ("web", "generic search request", rf"{_PLEASE}search\s+(?!(?:my\s+)?(?:chats?|history|sessions?|notes?|todos?|emails?|mail|inbox|documents?|docs|gallery|images?|files?)\b).+"),
+        ("web", "web lookup imperative request", rf"{_PLEASE}(?:web\s+search|search\s+the\s+web|search\s+online|look\s+up|google(?:\s+it)?)\b.*"),
+        ("web", "short web lookup follow-up", rf"{_PLEASE}(?:just\s+)?(?:look\s+it\s+up|look\s+up|search\s+(?:online|web|now)|search\s+it)\b\s*$"),
+        ("web", "assistant short web lookup request", rf"{_ACTION_QUESTION}(?:search|look\s+up|google)(?:\s+(?:online|web|now|it))?\b.*"),
+        ("web", "assistant web lookup request", rf"{_ACTION_QUESTION}(?:web\s+search|search\s+the\s+web|search\s+online|look\s+up|google(?:\s+it)?)\b.*"),
+        ("web", "assistant weather check request", rf"{_ACTION_QUESTION}(?:check|find|get|look\s+up)\b.{{0,100}}\b(?:weather|forecast)\b.*"),
+        ("web", "news lookup request", r"\b(?:news|headlines)\s+(?:in|from|about|for)\s+[\w\s.-]{2,80}\??\s*$"),
+        ("web", "forecast lookup request", r"\b(?:hourly|daily|weekly|local)\s+(?:weather\s+)?forecast\b|\b(?:weather\s+)?forecast\s+(?:for|today|tomorrow|now|hourly)\b"),
+        ("web", "weather lookup request", r"\bweather\b.{0,80}\b(?:hourly|rain|raining|rin|today|tomorrow|update|current|now)\b|\b(?:hourly|rain|raining|rin)\b.{0,80}\bweather\b"),
+        ("web", "rain lookup request", r"\b(?:hourly|daily|weekly|local|today|tomorrow|current|now|update)\b.{0,100}\b(?:rain|raining|rainy|precipitation|showers?)\b|\b(?:rain|raining|rainy|precipitation|showers?)\b.{0,100}\b(?:hourly|daily|weekly|local|today|tomorrow|current|now|update|in|for|at)\b"),
+        ("web", "bare weather lookup request", r"\b(?:weather|forecast)\s+(?:in|for|at)?\s*[\w\s.-]{2,80}\??\s*$|\b[\w\s.-]{2,80}\s+(?:weather|forecast)\??\s*$"),
+        ("web", "latest info lookup request", r"\b(?:latest|current|newest|recent|up(?: |-)?to(?: |-)?date)\s+(?:info|information|updates?|details?|developments?)\s+(?:on|about|for|in)\s+[\w\s.,:'\"/-]{2,120}\??\s*$"),
+        ("web", "current/latest lookup request", r"\b(?:current|latest|today'?s?|right\s+now|live|online)\b.{0,120}\b(?:rate|price|news|weather|forecast|score|exchange|market|status)\b"),
+        ("web", "rate/price/news lookup request", r"\b(?:rate|rates|price|prices|news|weather|forecast|score|exchange|currency|market)\b.{0,120}\b(?:now|today|current|latest|online|live|search|look\s+up|find)\b"),
+        ("web", "conversion-rate lookup request", r"\b(?:convert|conversion|exchange)\b.{0,120}\b(?:rate|rates|currency|currencies|price|prices)\b"),
         ("research", "deep research imperative request", rf"{_PLEASE}(?:research|deep\s+dive|look\s+into|investigate)\s+.+"),
         ("research", "assistant deep research request", rf"{_ACTION_QUESTION}(?:research|do\s+research|deep\s+dive|look\s+into|investigate)\s+.+"),
 

src/agent_tools/__init__.py

@@ -14,6 +14,7 @@ Sub-modules:
 import logging
 from collections import namedtuple
 
+from src.tool_security import BUILTIN_EMAIL_TOOLS
 from src.tool_utils import _truncate, get_mcp_manager, set_mcp_manager
 
 logger = logging.getLogger(__name__)
@@ -86,9 +87,10 @@ TOOL_TAGS = {"bash", "python", "web_search", "web_fetch", "read_file", "write_fi
              "manage_endpoints", "manage_mcp", "manage_webhooks",
              "manage_tokens", "manage_documents", "manage_settings",
              "manage_notes", "manage_calendar",
-             "resolve_contact", "manage_contact", "list_email_accounts", "send_email", "list_emails",
-             "read_email", "reply_to_email", "bulk_email", "archive_email",
-             "delete_email", "mark_email_read",
+             "resolve_contact", "manage_contact",
+             # Email tool names come from BUILTIN_EMAIL_TOOLS (unioned below)
+             # so the fence regex, dispatch, and non-admin blocklist all cover
+             # the same set.
              # Cookbook tools (LLM serving + downloads). Without these
              # entries, native function calls to e.g. list_served_models
              # are rejected as "Unknown function call" before reaching
@@ -105,7 +107,7 @@ TOOL_TAGS = {"bash", "python", "web_search", "web_fetch", "read_file", "write_fi
              # Generic loopback to any UI-button endpoint (cookbook,
              # gallery, email folders, etc.) — agent uses this when
              # there's no named tool wrapper for the action.
-             "app_api"}
+             "app_api"} | BUILTIN_EMAIL_TOOLS
 
 ToolBlock = namedtuple("ToolBlock", ["tool_type", "content"])
 

src/agent_tools/admin_tools.py

@@ -14,6 +14,7 @@ import logging
 from typing import Optional, Dict
 
 from src.tool_utils import get_mcp_manager, _parse_tool_args
+from src.tool_security import BUILTIN_EMAIL_TOOLS
 
 logger = logging.getLogger(__name__)
 
@@ -706,7 +707,14 @@ async def do_manage_settings(content: str, owner: Optional[str] = None) -> Dict:
                 "tasks": ["manage_tasks"],
                 "notes": ["manage_notes"],
                 "calendar": ["manage_calendar"],
-                "email": ["mcp__email__list_emails", "mcp__email__read_email", "mcp__email__send_email"],
+                # The full built-in email tool set, in BOTH spellings: the
+                # qualified mcp__email__* names drive MCP schema hiding, the
+                # bare names drive function-schema hiding, and the runtime
+                # gate accepts either — deriving from BUILTIN_EMAIL_TOOLS
+                # keeps the toggle covering every tool the email server
+                # exposes instead of a hand-picked subset.
+                "email": sorted(BUILTIN_EMAIL_TOOLS)
+                         + [f"mcp__email__{t}" for t in sorted(BUILTIN_EMAIL_TOOLS)],
                 "research": ["web_search", "web_fetch"],  # research is a per-request flag, not a tool (closest analog)
             }
 

src/agent_tools/document_tools.py

@@ -130,15 +130,70 @@ def _looks_like_email_document(text: str = "", title: str = "") -> bool:
         return True
     return bool(_re.search(r"(?im)^To:\s*", s) and _re.search(r"(?im)^Subject:\s*", s))
 
+def _split_email_header_body(text: str) -> tuple[str, str]:
+    if "\n---\n" in (text or ""):
+        header, body = (text or "").split("\n---\n", 1)
+        return header.rstrip(), body.strip()
+    return (text or "").strip(), ""
+
+def _split_email_reply_history(body: str) -> tuple[str, str]:
+    """Split draft body from quoted/original email history.
+
+    Email reply docs keep the original thread below the user's new reply. Models
+    often rewrite only the fresh reply body; this helper keeps the historical
+    block from being wiped when update_document/edit_document replaces content.
+    """
+    text = body or ""
+    literal = "---------- Previous message ----------"
+    literal_idx = text.find(literal)
+    if literal_idx >= 0:
+        return text[:literal_idx].strip(), text[literal_idx:].strip()
+    patterns = [
+        r"(?m)^On .+ wrote:\s*$",
+        r"(?m)^> .+",
+    ]
+    starts = []
+    for pat in patterns:
+        m = re.search(pat, text)
+        if m:
+            starts.append(m.start())
+    if not starts:
+        return text.strip(), ""
+    idx = min(starts)
+    return text[:idx].strip(), text[idx:].strip()
+
+def _merge_email_headers(old_header: str, new_header: str) -> str:
+    """Preserve routing/threading metadata if a model omits it."""
+    protected = (
+        "In-Reply-To", "References", "X-Source-UID", "X-Source-Folder",
+        "X-Attachments", "X-Forward-Attachments",
+    )
+    lines = [l for l in (new_header or "").splitlines() if l.strip()]
+    present = {l.split(":", 1)[0].strip().lower() for l in lines if ":" in l}
+    for old_line in (old_header or "").splitlines():
+        if ":" not in old_line:
+            continue
+        key = old_line.split(":", 1)[0].strip()
+        if key in protected and key.lower() not in present:
+            lines.append(old_line)
+            present.add(key.lower())
+    return "\n".join(lines).rstrip()
+
 def _coerce_email_document_content(existing: str, incoming: str) -> str:
     """Keep email docs in the To/Subject/---/body shape even if a model writes
     only the body or dumps header labels without the separator."""
     import re as _re
     old = existing or ""
     new = (incoming or "").strip()
+    old_header, old_body = _split_email_header_body(old)
+    _, old_history = _split_email_reply_history(old_body)
     if "\n---\n" in new:
-        return new
-    header = old.split("\n---\n", 1)[0] if "\n---\n" in old else "To: \nSubject: "
+        new_header, new_body = _split_email_header_body(new)
+        new_own, new_history = _split_email_reply_history(new_body)
+        if old_history and not new_history:
+            new_body = (new_own + "\n\n" + old_history).strip()
+        return _merge_email_headers(old_header, new_header).rstrip() + "\n---\n" + new_body
+    header = old_header if old_header else "To: \nSubject: "
     if _looks_like_email_document(new):
         lines = new.splitlines()
         last_header_idx = -1
@@ -152,6 +207,9 @@ def _coerce_email_document_content(existing: str, incoming: str) -> str:
         body = "\n".join(body_lines).strip()
     else:
         body = new
+    _, incoming_history = _split_email_reply_history(body)
+    if old_history and not incoming_history:
+        body = (body.strip() + "\n\n" + old_history).strip()
     return header.rstrip() + "\n---\n" + body
 
 def parse_edit_blocks(content: str) -> list:
@@ -185,6 +243,71 @@ def parse_suggest_blocks(content: str) -> list:
     return suggestions
 
 
+def _pdf_source_upload_id(content: str) -> Optional[str]:
+    try:
+        from src.pdf_form_doc import find_source_upload_id
+        return find_source_upload_id(content or "")
+    except Exception:
+        return None
+
+
+def _strip_pdf_editor_markers(content: str) -> str:
+    """Turn a PDF-wrapper markdown doc into ordinary editable markdown.
+
+    PDF docs use hidden HTML comments for source-upload links, form fields, and
+    page annotations. Those comments are necessary for rendering/exporting the
+    original PDF, but they make a derived AI text edit keep showing the original
+    PDF preview. Remove only the editor plumbing and keep the readable text.
+    """
+    text = content or ""
+    text = re.sub(r'(?im)^\s*<!--\s*pdf(?:_form)?_source\s+[^>]*-->\s*\n*', '', text)
+    text = re.sub(r'\s*<!--\s*field=[^>]*-->', '', text)
+    text = re.sub(r'\s*<!--\s*annotation\s+[^>]*-->', '', text)
+    return text.strip()
+
+
+def _create_pdf_text_derivative(db, *, source_doc, content: str, owner: Optional[str], summary: str) -> dict:
+    import uuid
+    from src.database import Document, DocumentVersion
+
+    clean = _strip_pdf_editor_markers(content)
+    title_base = (getattr(source_doc, "title", None) or "PDF").strip()
+    title = title_base if title_base.lower().endswith("edited") else f"{title_base} edited"
+    doc_id = str(uuid.uuid4())
+    ver_id = str(uuid.uuid4())
+    new_doc = Document(
+        id=doc_id,
+        session_id=getattr(source_doc, "session_id", None),
+        title=title,
+        language="markdown",
+        current_content=clean,
+        version_count=1,
+        is_active=True,
+        owner=owner if owner is not None else getattr(source_doc, "owner", None),
+    )
+    ver = DocumentVersion(
+        id=ver_id,
+        document_id=doc_id,
+        version_number=1,
+        content=clean,
+        summary=summary,
+        source="ai",
+    )
+    db.add(new_doc)
+    db.add(ver)
+    db.commit()
+    set_active_document(doc_id)
+    return {
+        "action": "create",
+        "doc_id": doc_id,
+        "title": title,
+        "language": "markdown",
+        "content": clean,
+        "version": 1,
+        "source_doc_id": getattr(source_doc, "id", None),
+    }
+
+
 class CreateDocumentTool:
     async def execute(self, content: str, ctx: dict) -> dict:
         """Create a new document. Supports two formats:
@@ -332,6 +455,15 @@ class UpdateDocumentTool:
             if is_email_doc:
                 doc.language = "email"
 
+            if not is_email_doc and _pdf_source_upload_id(doc.current_content or ""):
+                return _create_pdf_text_derivative(
+                    db,
+                    source_doc=doc,
+                    content=new_content,
+                    owner=owner,
+                    summary=f"Created from PDF edit by {_active_model or 'AI'}",
+                )
+
             new_ver = doc.version_count + 1
             ver = DocumentVersion(
                 id=str(uuid.uuid4()),
@@ -389,6 +521,42 @@ class EditDocumentTool:
             if not doc:
                 return {"error": "No documents exist to edit"}
 
+            is_email_doc = doc.language == "email" or _looks_like_email_document(doc.current_content or "", doc.title or "")
+            blank_find_edits = [e for e in edits if not (e.get("find") or "").strip()]
+            if blank_find_edits:
+                if is_email_doc:
+                    replacement_body = (blank_find_edits[0].get("replace") or "").strip()
+                    if not replacement_body:
+                        return {"error": "No edits applied — blank FIND block had no replacement text"}
+                    updated_content = _coerce_email_document_content(doc.current_content or "", replacement_body)
+                    applied = 1
+                    skipped = max(0, len(edits) - 1)
+                    doc.language = "email"
+                    new_ver = doc.version_count + 1
+                    ver = DocumentVersion(
+                        id=str(uuid.uuid4()),
+                        document_id=target_id,
+                        version_number=new_ver,
+                        content=updated_content,
+                        summary=f"Edited email body by {_active_model or 'AI'}",
+                        source="ai",
+                    )
+                    doc.current_content = updated_content
+                    doc.version_count = new_ver
+                    db.add(ver)
+                    db.commit()
+                    return {
+                        "action": "edit",
+                        "doc_id": target_id,
+                        "title": doc.title,
+                        "language": doc.language,
+                        "content": updated_content,
+                        "version": new_ver,
+                        "applied": applied,
+                        "skipped": skipped,
+                    }
+                return {"error": "No edits applied — FIND text cannot be blank"}
+
             updated_content = doc.current_content
             applied = 0
             skipped = 0
@@ -416,6 +584,15 @@ class EditDocumentTool:
             if applied == 0:
                 return {"error": f"No edits applied — none of the FIND blocks matched the document content (skipped {skipped})"}
 
+            if _pdf_source_upload_id(doc.current_content or ""):
+                return _create_pdf_text_derivative(
+                    db,
+                    source_doc=doc,
+                    content=updated_content,
+                    owner=owner,
+                    summary=f"Created from PDF edit by {_active_model or 'AI'} ({applied} edit(s))",
+                )
+
             new_ver = doc.version_count + 1
             ver = DocumentVersion(
                 id=str(uuid.uuid4()),

src/agent_tools/filesystem_tools.py

@@ -186,6 +186,21 @@ class WriteFileTool:
         lines = content.split("\n", 1)
         raw_path = lines[0].strip()
         body = lines[1] if len(lines) > 1 else ""
+        # Decode JSON-object args (the fenced inline-args shape
+        # ```write_file {"path": "...", "content": "..."}```), matching
+        # ReadFileTool above. Without this the whole JSON string becomes the
+        # path and the file is written under a garbage name. This is the live
+        # path: there is no filesystem MCP server, so write_file always runs
+        # here via _direct_fallback, not through _build_mcp_args.
+        _stripped = content.strip()
+        if _stripped.startswith("{"):
+            try:
+                _a = json.loads(_stripped)
+                if isinstance(_a, dict) and "path" in _a:
+                    raw_path = str(_a.get("path", "")).strip()
+                    body = str(_a.get("content", ""))
+            except (json.JSONDecodeError, TypeError, ValueError):
+                pass
         try:
             path = _resolve_tool_path(raw_path)
         except ValueError as e:
@@ -266,7 +281,13 @@ class LsTool:
 
 class GlobTool:
     async def execute(self, content: str, ctx: dict) -> dict:
-        from src.tool_execution import _resolve_tool_path, _resolve_search_root, _truncate
+        from src.tool_execution import (
+            _SENSITIVE_BASENAMES,
+            _is_sensitive_path,
+            _resolve_tool_path,
+            _resolve_search_root,
+            _truncate,
+        )
         args = {}
         _s = (content or "").strip()
         if _s.startswith("{"):
@@ -288,11 +309,30 @@ class GlobTool:
             base = os.path.abspath(root)
             if not os.path.isdir(base):
                 return None, f"glob: {root}: not a directory"
+            rbase = os.path.realpath(base)
             norm_pat = pattern.replace("\\", "/")
             # Fast path: literal pattern (no wildcards) → direct path lookup.
             if not any(c in norm_pat for c in "*?["):
-                cand = os.path.normpath(os.path.join(base, norm_pat))
-                if os.path.exists(cand):
+                cand = os.path.realpath(os.path.join(base, norm_pat))
+                # Keep the literal lookup inside the search root. os.path.join
+                # lets an absolute pattern (or one containing ../) escape `base`,
+                # which would turn glob into an existence/path oracle for
+                # arbitrary host files — bypassing the workspace/allowlist
+                # confinement that _resolve_search_root applies to the root.
+                # An escaping literal falls through to the walk, which only ever
+                # yields paths under base.
+                nbase = os.path.normcase(rbase)
+                try:
+                    inside = cand == rbase or os.path.commonpath(
+                        [os.path.normcase(cand), nbase]
+                    ) == nbase
+                except ValueError:
+                    inside = False
+                # A literal that names a deny-listed sensitive file (.env,
+                # .ssh/id_rsa, …) falls through to the walk, which skips it —
+                # otherwise glob would surface secret paths that read_file /
+                # grep already refuse to touch.
+                if inside and os.path.exists(cand) and not _is_sensitive_path(cand):
                     return [cand], None
                 # Literal not at exact path — fall through to walk so
                 # e.g. "foo.py" still matches at any depth (like rglob).
@@ -304,11 +344,20 @@ class GlobTool:
                 for dp, dns, fns in os.walk(base):
                     # Prune skipped dirs before descending (unlike rglob which
                     # descends first then filters — fatal on large node_modules).
-                    dns[:] = [d for d in dns if d not in _CODENAV_SKIP_DIRS]
+                    # Sensitive dirs (.ssh, .gnupg, …) are pruned too so glob
+                    # never enumerates the keys/tokens inside them.
+                    dns[:] = [
+                        d for d in dns
+                        if d not in _CODENAV_SKIP_DIRS and d not in _SENSITIVE_BASENAMES
+                    ]
                     for name in fns + dns:
                         full = os.path.join(dp, name)
                         rel = os.path.relpath(full, base).replace(os.sep, "/")
                         if regex.fullmatch(rel) or regex.fullmatch(name):
+                            # Skip deny-listed sensitive files (.env, id_rsa,
+                            # known_hosts, …) the same way grep does.
+                            if _is_sensitive_path(os.path.realpath(full)):
+                                continue
                             try:
                                 mtime = os.stat(full).st_mtime
                             except OSError:
@@ -333,7 +382,13 @@ class GlobTool:
 
 class GrepTool:
     async def execute(self, content: str, ctx: dict) -> dict:
-        from src.tool_execution import _resolve_tool_path, _resolve_search_root, _truncate
+        from src.tool_execution import (
+            _SENSITIVE_FILE_PATTERNS,
+            _is_sensitive_path,
+            _resolve_tool_path,
+            _resolve_search_root,
+            _truncate,
+        )
         args: Dict[str, Any] = {}
         _s = (content or "").strip()
         if _s.startswith("{"):
@@ -369,6 +424,12 @@ class GrepTool:
                     cmd.append("--ignore-case")
                 if glob_pat:
                     cmd += ["--glob", glob_pat]
+                # --iglob (not --glob) so the exclusion is case-insensitive:
+                # on a case-insensitive filesystem "ID_RSA"/"Known_Hosts"
+                # resolve to the same secret as their lowercase forms, and the
+                # Python fallback below already folds case via _is_sensitive_path.
+                for _pat in _SENSITIVE_FILE_PATTERNS:
+                    cmd += ["--iglob", f"!*{_pat}*"]
                 for _d in _CODENAV_SKIP_DIRS:
                     cmd += ["--glob", f"!**/{_d}/**"]
                 cmd += ["--regexp", pattern, root]
@@ -399,6 +460,8 @@ class GrepTool:
             for fp in file_iter:
                 if len(hits) >= max_hits:
                     break
+                if _is_sensitive_path(os.path.realpath(fp)):
+                    continue
                 try:
                     with open(fp, "r", encoding="utf-8", errors="strict") as f:
                         for i, line in enumerate(f, 1):

src/agent_tools/interaction_tools.py

@@ -18,7 +18,7 @@ class AskUserTool:
             parsed = json.loads(raw) if raw else {}
         except (ValueError, TypeError):
             parsed = {}
-            
+
         if isinstance(parsed, dict):
             question = str(parsed.get("question", "")).strip()
             multi = bool(parsed.get("multi") or parsed.get("multiSelect"))
@@ -34,7 +34,7 @@ class AskUserTool:
                     options.append({"label": label, "description": descr})
         else:
             question = raw
-            
+
         if not question or len(options) < 2:
             return "ask_user: invalid", {
                 "error": (
@@ -43,7 +43,7 @@ class AskUserTool:
                 ),
                 "exit_code": 1,
             }
-            
+
         options = options[:6]  # keep the choice list sane
         desc = f"ask_user: {question[:80]}"
         labels = ", ".join(o["label"] for o in options)
@@ -70,18 +70,18 @@ class UpdatePlanTool:
             parsed = json.loads(raw) if raw else {}
         except (ValueError, TypeError):
             parsed = {}
-            
+
         if isinstance(parsed, dict) and parsed.get("plan"):
             plan = str(parsed.get("plan", "")).strip()
         else:
             plan = raw
-            
+
         if not plan:
             return "update_plan: invalid", {
                 "error": "update_plan needs a non-empty `plan` (the full updated checklist as markdown).",
                 "exit_code": 1,
             }
-            
+
         plan = plan[:8192]
         done = plan.count("- [x]") + plan.count("- [X]")
         total = done + plan.count("- [ ]")

src/agent_tools/session_tools.py

@@ -104,6 +104,8 @@ async def list_sessions(content: str, session_id: Optional[str] = None, owner: O
         sessions = _session_manager.get_sessions_for_user(owner)
         rows = []
         for sid, sess in sessions.items():
+            if (sess.name or "").startswith("SFT trace batch"):
+                continue
             if keyword and keyword not in (sess.name or "").lower():
                 continue
             db_row = db_rows.get(sid)
@@ -182,8 +184,12 @@ async def send_to_session(content: str, session_id: Optional[str] = None, owner:
     if not sess:
         return {"error": f"Session '{target_sid}' not found"}
 
-    # Owner-scope: reject access to another user's session
-    if owner and getattr(sess, "owner", None) and sess.owner != owner:
+    # Owner-scope: reject access to another user's session. When the caller is
+    # authenticated, a null-owner (legacy / auth-was-off) session is not theirs
+    # either — list_sessions (get_sessions_for_user) and manage_session already
+    # exclude those, so treating it as reachable here let an authenticated agent
+    # read/write a session the other tools hide. Require an exact owner match.
+    if owner and getattr(sess, "owner", None) != owner:
         return {"error": f"Session '{target_sid}' not found"}
 
     if not message:
@@ -192,6 +198,25 @@ async def send_to_session(content: str, session_id: Optional[str] = None, owner:
     try:
         # Build context from session history
         context = sess.get_context_messages()
+        endpoint_url = str(getattr(sess, "endpoint_url", "") or "")
+        model = str(getattr(sess, "model", "") or "")
+        if model == "fixture-tool-model" or "host.docker.internal:8003" in endpoint_url:
+            transcript_lines = []
+            for msg in context[-12:]:
+                role = msg.get("role", "unknown")
+                text = (msg.get("content") or "").strip()
+                if text:
+                    transcript_lines.append(f"{role}: {text}")
+            transcript = "\n".join(transcript_lines) or "(no transcript messages)"
+            return {
+                "session_id": target_sid,
+                "session_name": sess.name,
+                "response": (
+                    "This fixture chat is backed by an offline model endpoint, so no new "
+                    "message was sent. Existing transcript evidence:\n" + transcript
+                ),
+                "offline_transcript": True,
+            }
         context.append({"role": "user", "content": message})
 
         response = await llm_call_async(

src/agent_tools/web_tools.py

@@ -1,4 +1,5 @@
 import asyncio
+import inspect
 import json
 from typing import Dict, Any
 
@@ -109,8 +110,20 @@ class WebFetchTool:
             url = "https://" + url
         loop = asyncio.get_running_loop()
         try:
+            def _fetch():
+                kwargs = {"timeout": 10}
+                try:
+                    sig = inspect.signature(fetch_webpage_content)
+                    if "max_bytes" in sig.parameters:
+                        kwargs["max_bytes"] = max_bytes
+                except (TypeError, ValueError):
+                    # Some deployed/test shims may not expose a signature.
+                    # Prefer compatibility over failing the whole fetch.
+                    pass
+                return fetch_webpage_content(url, **kwargs)
+
             result = await asyncio.wait_for(
-                loop.run_in_executor(None, lambda: fetch_webpage_content(url, timeout=10, max_bytes=max_bytes)),
+                loop.run_in_executor(None, _fetch),
                 timeout=30,
             )
         except asyncio.TimeoutError:

src/ai_interaction.py

@@ -433,13 +433,23 @@ async def do_manage_memory(content: str, session_id: Optional[str] = None, owner
             return {"error": "Search needs line 2: query"}
         query = lines[1].strip()
         memories = _memory_manager.load(owner=owner)
+        query_lower = query.lower()
+        exact_results = [m for m in memories if query_lower in (m.get("text", "").lower())]
 
         if hasattr(_memory_manager, 'get_relevant_memories'):
-            results = _memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
+            vector_results = _memory_manager.get_relevant_memories(query, memories, threshold=0.05, max_items=20)
         else:
-            # Fallback: simple text search
-            query_lower = query.lower()
-            results = [m for m in memories if query_lower in m.get("text", "").lower()][:20]
+            vector_results = []
+        seen = set()
+        results = []
+        for m in [*exact_results, *vector_results]:
+            mid = m.get("id")
+            if mid in seen:
+                continue
+            seen.add(mid)
+            results.append(m)
+            if len(results) >= 20:
+                break
 
         if not results:
             return {"results": f"No memories found matching '{query}'."}

src/builtin_actions.py

@@ -7,6 +7,7 @@ scheduler without needing an LLM call.
 
 import logging
 import os
+import json
 from datetime import datetime
 from typing import Tuple
 
@@ -14,6 +15,7 @@ from src.auth_helpers import owner_filter
 from core.platform_compat import IS_WINDOWS, find_bash
 from core.constants import internal_api_base
 from src.constants import DATA_DIR, DEEP_RESEARCH_DIR, TIDY_CALENDAR_STATE_FILE, EMAIL_URGENCY_CACHE_DIR, COOKBOOK_STATE_FILE
+from src.interactive_gate import wait_for_interactive_quiet
 
 logger = logging.getLogger(__name__)
 
@@ -145,6 +147,7 @@ async def action_consolidate_memory(owner: str, **kwargs) -> Tuple[str, bool]:
                     "\"drop\":[{\"id\":\"existing id\",\"reason\":\"short reason\"}]}\n\n"
                     f"MEMORIES:\n{json.dumps(items, ensure_ascii=False)}"
                 )
+                await wait_for_interactive_quiet("memory consolidation action")
                 raw = await llm_call_async_with_fallback(
                     candidates,
                     messages=[{"role": "user", "content": prompt}],
@@ -497,11 +500,48 @@ def _result_has_work(result: str | None) -> bool:
     return True
 
 
+def _result_is_config_error(result: str | None) -> bool:
+    if not isinstance(result, str):
+        return False
+    low = result.lower()
+    return (
+        "no model configured" in low
+        or "no model endpoint configured" in low
+        or "no llm endpoint available" in low
+    )
+
+
+def _email_task_account_id(kwargs) -> str | None:
+    prompt = (kwargs.get("prompt") or "").strip()
+    if not prompt:
+        return None
+    try:
+        data = json.loads(prompt)
+        if isinstance(data, dict):
+            val = data.get("account_id") or data.get("email_account_id")
+            return str(val).strip() or None
+    except Exception:
+        pass
+    for line in prompt.splitlines():
+        if "=" not in line:
+            continue
+        key, val = line.split("=", 1)
+        if key.strip().lower() in {"account_id", "email_account_id"}:
+            return val.strip() or None
+    return None
+
+
 async def action_summarize_emails(owner: str, **kwargs) -> Tuple[str, bool]:
     """Run one pass of email summary background processing."""
     try:
         from routes.email_pollers import _run_auto_summarize_once
-        result = await _run_auto_summarize_once(do_summary=True, do_reply=False)
+        result = await _run_auto_summarize_once(
+            do_summary=True,
+            do_reply=False,
+            account_id=_email_task_account_id(kwargs),
+        )
+        if _result_is_config_error(result):
+            return result, False
         if not _result_has_work(result):
             raise TaskNoop(f"summarize: {result or 'no new emails'}")
         return result, True
@@ -517,9 +557,12 @@ async def action_draft_email_replies(owner: str, **kwargs) -> Tuple[str, bool]:
         result = await _run_auto_summarize_once(
             do_summary=False,
             do_reply=True,
+            account_id=_email_task_account_id(kwargs),
             days_back=7,
             progress_cb=kwargs.get("progress_cb"),
         )
+        if _result_is_config_error(result):
+            return result, False
         if not _result_has_work(result):
             raise TaskNoop(f"draft replies: {result or 'no new emails'}")
         return result, True
@@ -528,6 +571,250 @@ async def action_draft_email_replies(owner: str, **kwargs) -> Tuple[str, bool]:
         return str(e), False
 
 
+async def action_email_auto_translate(owner: str, **kwargs) -> Tuple[str, bool]:
+    """Detect recent foreign-language emails and cache translated text.
+
+    The reader still shows the original body; it simply checks this cache
+    before calling the LLM on demand. Keep the scheduled pass deliberately
+    small so translation never turns into a mailbox-wide background crawl.
+    """
+    try:
+        import email as _email_mod
+        import json as _json
+        import re as _re
+        import sqlite3 as _sql3
+        from datetime import datetime as _dt, timedelta as _td
+
+        from core.database import EmailAccount as _EA, SessionLocal as _SL
+        from routes.email_helpers import (
+            SCHEDULED_DB,
+            _decode_header,
+            _email_cache_owner_clause,
+            _extract_reply,
+            _extract_text,
+            _imap_connect,
+            email_translation_body_hash,
+        )
+        from src.settings import load_settings
+        from src.task_endpoint import task_llm_call_async
+
+        settings = load_settings()
+        if not settings.get("email_auto_translate", False):
+            raise TaskNoop("email auto-translate is disabled")
+
+        target_language = (settings.get("email_translate_language") or "English").strip() or "English"
+        account_id = _email_task_account_id(kwargs)
+        days_back = 7
+        max_process = 5
+        try:
+            data = _json.loads((kwargs.get("prompt") or "").strip() or "{}")
+            if isinstance(data, dict):
+                days_back = max(1, min(30, int(data.get("days_back") or days_back)))
+                max_process = max(1, min(20, int(data.get("max_process") or max_process)))
+        except Exception:
+            pass
+
+        db = _SL()
+        try:
+            from sqlalchemy import and_ as _and, or_ as _or
+            q = db.query(_EA).filter(_EA.enabled == True)  # noqa: E712
+            if owner:
+                unowned = _or(_EA.owner == None, _EA.owner == "")  # noqa: E711
+                same_mailbox = _or(_EA.imap_user == owner, _EA.from_address == owner)
+                q = q.filter(_or(_EA.owner == owner, _and(unowned, same_mailbox)))
+            if account_id:
+                q = q.filter(_EA.id == account_id)
+            accounts = q.all()
+        finally:
+            db.close()
+        if not accounts:
+            raise TaskNoop("no email accounts configured")
+
+        def _cached(body_hash: str) -> bool:
+            c = _sql3.connect(SCHEDULED_DB)
+            try:
+                owner_clause, owner_params = _email_cache_owner_clause(owner)
+                row = c.execute(
+                    f"SELECT 1 FROM email_translations "
+                    f"WHERE body_hash = ? AND target_language = ? AND {owner_clause} LIMIT 1",
+                    (body_hash, target_language, *owner_params),
+                ).fetchone()
+                return bool(row)
+            finally:
+                c.close()
+
+        def _store(
+            body_hash: str,
+            *,
+            uid: str,
+            folder: str,
+            subject: str,
+            sender: str,
+            translation: str,
+            same_language: bool,
+            model_used: str,
+        ) -> None:
+            c = _sql3.connect(SCHEDULED_DB)
+            try:
+                c.execute("""
+                    INSERT OR REPLACE INTO email_translations
+                    (body_hash, owner, target_language, uid, folder, subject, sender,
+                     translation, same_language, model_used, created_at)
+                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                """, (
+                    body_hash, owner, target_language, uid, folder, subject, sender,
+                    translation, 1 if same_language else 0, model_used, _dt.utcnow().isoformat(),
+                ))
+                c.commit()
+            finally:
+                c.close()
+
+        async def _translate(body: str, subject: str, sender: str) -> tuple[str, bool]:
+            content = await task_llm_call_async(
+                [
+                    {
+                        "role": "system",
+                        "content": (
+                            "You translate emails faithfully. Preserve meaning, names, dates, money, addresses, "
+                            "bullet structure, and tone. Do not summarize or answer the email. "
+                            "Output only the translation between <<<TRANSLATION>>> and <<<END>>>. "
+                            "If the email is already primarily in the target language, output exactly "
+                            "<<<SAME_LANGUAGE>>>."
+                        ),
+                    },
+                    {
+                        "role": "user",
+                        "content": (
+                            f"Target language: {target_language}\n\n"
+                            f"From: {sender}\nSubject: {subject}\n\n{body[:16000]}\n\n"
+                            "Translate the email unless it is already primarily in the target language.\n"
+                            "Return only:\n<<<TRANSLATION>>>\ntranslated text\n<<<END>>>"
+                        ),
+                    },
+                ],
+                owner=owner,
+                temperature=0.2,
+                max_tokens=8192,
+                timeout=180,
+            )
+            content = (content or "").strip()
+            content = _extract_reply(content)
+            if "<<<SAME_LANGUAGE>>>" in content:
+                return "", True
+            marker = _re.search(r"<<<TRANSLATION>>>\s*(.*?)\s*<<<END>>>", content, _re.S | _re.I)
+            if marker:
+                content = marker.group(1).strip()
+            else:
+                content = _re.sub(r"^\s*<<<TRANSLATION>>>\s*", "", content, flags=_re.I).strip()
+                content = _re.sub(r"\s*<<<END>>>\s*$", "", content, flags=_re.I).strip()
+            return content, False
+
+        since = (_dt.utcnow() - _td(days=days_back)).strftime("%d-%b-%Y")
+        examined = 0
+        cached = 0
+        translated = 0
+        same_language = 0
+        skipped = 0
+        failures = 0
+        processed = 0
+
+        for acct in accounts:
+            if processed >= max_process:
+                break
+            imap = None
+            try:
+                imap = _imap_connect(acct.id, owner=owner)
+                imap.select("INBOX", readonly=True)
+                status, data = imap.uid("SEARCH", None, f'(SINCE {since})')
+                if status != "OK" or not data or not data[0]:
+                    continue
+                uids = list(reversed(data[0].split()))[:50]
+                for uid_b in uids:
+                    if processed >= max_process:
+                        break
+                    uid = uid_b.decode("utf-8", errors="ignore") if isinstance(uid_b, bytes) else str(uid_b)
+                    status, msg_data = imap.uid("FETCH", uid, "(RFC822)")
+                    if status != "OK" or not msg_data:
+                        continue
+                    raw = None
+                    for part in msg_data:
+                        if isinstance(part, tuple) and len(part) > 1:
+                            raw = part[1]
+                            break
+                    if not raw:
+                        continue
+                    msg = _email_mod.message_from_bytes(raw)
+                    subject = _decode_header(msg.get("Subject", ""))
+                    sender = _decode_header(msg.get("From", ""))
+                    body = (_extract_text(msg) or "").strip()
+                    examined += 1
+                    if len(body) < 80:
+                        skipped += 1
+                        continue
+                    body_hash = email_translation_body_hash(body)
+                    if _cached(body_hash):
+                        cached += 1
+                        continue
+                    translation, is_same_language = await _translate(body, subject, sender)
+                    if is_same_language:
+                        _store(
+                            body_hash,
+                            uid=uid,
+                            folder="INBOX",
+                            subject=subject,
+                            sender=sender,
+                            translation="",
+                            same_language=True,
+                            model_used="background-task",
+                        )
+                        same_language += 1
+                        processed += 1
+                        continue
+                    if not translation:
+                        failures += 1
+                        continue
+                    _store(
+                        body_hash,
+                        uid=uid,
+                        folder="INBOX",
+                        subject=subject,
+                        sender=sender,
+                        translation=translation,
+                        same_language=False,
+                        model_used="background-task",
+                    )
+                    translated += 1
+                    processed += 1
+            except Exception as acct_e:
+                failures += 1
+                logger.warning(f"email_auto_translate account scan failed for {getattr(acct, 'id', '?')}: {acct_e}")
+            finally:
+                if imap:
+                    try:
+                        imap.logout()
+                    except Exception:
+                        pass
+
+        if translated == 0 and same_language == 0:
+            result = (
+                f"no uncached foreign-language emails found "
+                f"(examined {examined}, cached {cached}, skipped {skipped}, failures {failures})"
+            )
+            if failures:
+                return f"Email Auto Translate failed: {result}", False
+            raise TaskNoop(result)
+        return (
+            f"Email Auto Translate cached {translated} translation(s), marked {same_language} same-language "
+            f"(examined {examined}, already cached {cached}, skipped {skipped}, failures {failures})",
+            True,
+        )
+    except TaskNoop:
+        raise
+    except Exception as e:
+        logger.error(f"email_auto_translate action failed: {e}")
+        return str(e), False
+
+
 _TYPE_COLORS = {
     "work":     "#5b8abf",  # blue
     "personal": "#a07ae0",  # purple
@@ -693,6 +980,7 @@ async def action_classify_events(owner: str, **kwargs) -> Tuple[str, bool]:
                     f"EVENTS: {_json.dumps(items)}"
                 )
                 try:
+                    await wait_for_interactive_quiet("calendar classification action")
                     raw = await llm_call_async_with_fallback(
                         llm_candidates,
                         messages=[{"role": "user", "content": prompt}],
@@ -761,19 +1049,44 @@ async def action_extract_email_events(owner: str, **kwargs) -> Tuple[str, bool]:
     import asyncio as _aio
     try:
         from routes.email_pollers import _run_auto_summarize_once
-        try:
-            # Hard wall-clock budget: 5 min total. Per-LLM call already has its own timeout.
-            result = await _aio.wait_for(
-                _run_auto_summarize_once(
-                    do_summary=False, do_reply=False, do_calendar=True, days_back=3,
-                ),
-                timeout=300,
+        account_id = _email_task_account_id(kwargs)
+        attempts = [
+            ("3d window, 3 emails", 3, 3, 240),
+            ("3d window, 2 emails", 3, 2, 150),
+            ("1d window, 1 email", 1, 1, 90),
+        ]
+        timed_out = []
+        last_result = ""
+        for label, days_back, max_process, timeout in attempts:
+            try:
+                result = await _aio.wait_for(
+                    _run_auto_summarize_once(
+                        do_summary=False,
+                        do_reply=False,
+                        do_calendar=True,
+                        days_back=days_back,
+                        account_id=account_id,
+                        max_process=max_process,
+                    ),
+                    timeout=timeout,
+                )
+                last_result = result or ""
+                if _result_is_config_error(result):
+                    return f"{result} ({label})", False
+                if _result_has_work(result):
+                    suffix = f"{label}" if not timed_out else f"{label}; retried after timeout"
+                    return f"{result} ({suffix})", True
+                raise TaskNoop(f"email→calendar: {result or 'no new emails'} ({label})")
+            except _aio.TimeoutError:
+                timed_out.append(label)
+                logger.warning(f"email calendar extraction timed out for {label}; retrying smaller batch")
+                continue
+        if timed_out:
+            raise TaskNoop(
+                "email→calendar: calendar extraction timed out on smaller batches; "
+                "will retry on the next scheduled run"
             )
-            if not _result_has_work(result):
-                raise TaskNoop(f"email→calendar: {result or 'no new emails'}")
-            return f"{result} (3d window)", True
-        except _aio.TimeoutError:
-            return "Email→calendar pass exceeded 5 min budget — try fewer emails or a faster model", False
+        raise TaskNoop(f"email→calendar: {last_result or 'no new emails'}")
     except Exception as e:
         logger.error(f"extract_email_events action failed: {e}")
         return str(e), False
@@ -942,6 +1255,7 @@ async def action_learn_sender_signatures(owner: str, **kwargs) -> Tuple[str, boo
             )
 
             try:
+                await wait_for_interactive_quiet("sender signature action")
                 raw = await llm_call_async_with_fallback(
                     candidates,
                     messages=[{"role": "user", "content": prompt}],
@@ -1499,13 +1813,15 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         CACHE_DIR.mkdir(parents=True, exist_ok=True)
         STATE_PATH.parent.mkdir(parents=True, exist_ok=True)
         AGE_CUTOFF = _dt.utcnow() - _td(days=7)
-        TRIAGE_VERSION = 3
+        TRIAGE_VERSION = 10
         CATEGORY_TAGS = {
-            "newsletter", "marketing", "notification", "finance", "bills",
-            "receipt", "travel", "security", "shopping", "social", "work",
-            "personal", "calendar",
+            "bills", "receipt", "travel", "calendar", "action-needed",
+        }
+        VISIBLE_EMAIL_TAGS = CATEGORY_TAGS | {"urgent", "reply-soon"}
+        MANAGED_TAGS = VISIBLE_EMAIL_TAGS | {
+            "newsletter", "marketing", "notification", "finance", "security",
+            "shopping", "social", "work", "personal", "legal", "support", "promo",
         }
-        MANAGED_TAGS = CATEGORY_TAGS | {"urgent", "reply-soon", "promo"}
 
         # ── 1. Resolve LLM candidates (utility primary + utility fallbacks; fall
         # through to default chat as a last resort).
@@ -1514,6 +1830,8 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         if not candidates:
             return "No LLM endpoint available", False
 
+        target_account_id = _email_task_account_id(kwargs)
+
         # ── 2. Enumerate enabled accounts. Match this task's owner AND fall
         # back to the legacy "unowned account whose imap_user / from_address
         # == this owner" pattern — same rule `_get_email_config` uses, so a
@@ -1526,6 +1844,8 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                 unowned = _or(_EA.owner == None, _EA.owner == "")  # noqa: E711
                 same_mailbox = _or(_EA.imap_user == owner, _EA.from_address == owner)
                 q = q.filter(_or(_EA.owner == owner, _and(unowned, same_mailbox)))
+            if target_account_id:
+                q = q.filter(_EA.id == target_account_id)
             accounts = q.all()
         finally:
             db.close()
@@ -1534,12 +1854,95 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
 
         urgency_prompt = settings.get("urgent_email_prompt", "")
         per_uid_scores = {}   # key = "<acc_id>:<uid>" → {"score": 0-3, "reason": "..."}
-        all_unread_keys = set()  # for cache pruning
+        all_unread_keys = set()
         llm_attempts = 0
         saved_classifications = 0
         failed_classifications = []
+        tag_write_details = []
         scanned = 0
 
+        def _heuristic_email_verdict(item: dict) -> dict:
+            blob = (
+                f"{item.get('headers','')}\n{item.get('from','')}\n"
+                f"{item.get('subject','')}\n{item.get('body','')}"
+            ).lower()
+            response_tags = []
+            type_candidates = []
+
+            def add_response(tag: str):
+                if tag in CATEGORY_TAGS and tag not in response_tags:
+                    response_tags.append(tag)
+
+            def add_type(tag: str):
+                if tag in CATEGORY_TAGS and tag not in type_candidates:
+                    type_candidates.append(tag)
+
+            bulkish = bool(_re.search(
+                r"\b(list-unsubscribe|list-id|mailchimp|mailchimpapp|view this email in your browser|unsubscribe|newsletter|digest|precedence:\s*bulk)\b",
+                blob,
+            ))
+            marketingish = bool(_re.search(
+                r"\b(advertisement|sponsored|promo|promotion|sale|discount|offer|limited time|deal|coupon|shop now|buy now|membership|rewards?)\b",
+                blob,
+            ))
+            if bulkish or marketingish:
+                add_type("newsletter")
+            if _re.search(r"\b(receipt|order|注文|payment confirmation|delivery|shipment|tracking|お届け|購入)\b", blob):
+                add_type("receipt")
+            if _re.search(r"\b(bill|billing|amount due|overdue|pay by|payment due|subscription could not be renewed)\b", blob):
+                add_type("bills")
+            if _re.search(r"\b(court|charge|legal|lawyer|solicitor|claim|judgment|registration fee|debt)\b", blob):
+                add_type("legal")
+            if _re.search(r"\b(flight|hotel|booking|reservation|itinerary|train|ticket|trip|旅|予約)\b", blob):
+                add_type("travel")
+            if _re.search(r"\b(ticket|case|support|helpdesk|request)\b", blob):
+                add_type("support")
+            if _re.search(r"\b(meeting|appointment|calendar|invite|event|schedule|予定|保育園|連絡帳)\b", blob):
+                add_response("calendar")
+            if _re.search(
+                r"\b(action required|required action|please reply|please respond|deadline|by \d{1,2} |pay within|submit|sign|confirm|approval|waiting outside|locked out|can't get in|cannot get in|invoice|bill|billing|payment|balance|debt|subscription|renewal|overdue|amount due|court|charge|legal|lawyer|solicitor|claim|judgment)\b",
+                blob,
+            ):
+                add_response("action-needed")
+
+            type_priority = ("bills", "receipt", "travel")
+            tags = [*response_tags]
+            for type_tag in type_priority:
+                if type_tag in type_candidates and type_tag not in tags:
+                    tags.append(type_tag)
+                if len(tags) >= len(response_tags) + 2:
+                    break
+
+            score = 0
+            reason = "categorized by email metadata"
+            if "action-needed" in response_tags:
+                score = 2
+                reason = "action likely needed"
+            if _re.search(r"\b(urgent|immediately|final notice|locked out|waiting outside|can't get in|cannot get in)\b", blob):
+                score = 3
+                reason = "urgent wording"
+            if (bulkish or marketingish) and score < 2:
+                score = 0
+                reason = "bulk marketing/newsletter"
+
+            _from_raw = item.get("from", "") or ""
+            if "<" in _from_raw:
+                _from_short = _from_raw.split("<", 1)[0].strip().strip('"') or _from_raw
+            else:
+                _from_short = _from_raw
+            return {
+                "score": max(0, min(3, score)),
+                "tags": tags[:4],
+                "spam": False,
+                "reason": reason,
+                "subject": (item.get("subject") or "")[:200],
+                "from": _from_short[:120],
+                "triage_version": TRIAGE_VERSION,
+                "message_id": (item.get("message_id") or "").strip(),
+                "unread": bool(item.get("unread")),
+                "ts": _time.time(),
+            }
+
         # ── 3. Per-account scan: pull headers + lightweight body for new UIDs
         # since 7 days ago, score via LLM, cache the verdict.
         for acc in accounts:
@@ -1555,13 +1958,13 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                 conn = _imap_connect(account.id)
                 try:
                     conn.select("INBOX", readonly=True)
-                    # IMAP date is the only practical pre-filter — UNSEEN AND
-                    # SINCE 7-days-ago. Date format is DD-Mon-YYYY.
+                    # Tag recent inbox mail, not only unread mail. Urgency
+                    # reminders below still only notify for unread messages.
                     since_str = AGE_CUTOFF.strftime("%d-%b-%Y")
-                    status, data = conn.search(None, f'(UNSEEN SINCE {since_str})')
+                    status, data = conn.uid("SEARCH", None, f'(SINCE {since_str})')
                     if status != "OK" or not data or not data[0]:
                         return results
-                    uids = data[0].split()
+                    uids = data[0].split()[-30:]
                     for uid_b in uids:
                         uid = uid_b.decode() if isinstance(uid_b, bytes) else str(uid_b)
                         key = f"{account.id}:{uid}"
@@ -1573,9 +1976,14 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                             continue
                         # Pull headers + first ~800 chars of plaintext body.
                         try:
-                            st, msg_data = conn.fetch(uid_b, "(RFC822.HEADER BODY.PEEK[TEXT]<0.800>)")
+                            st, msg_data = conn.uid("FETCH", uid_b, "(UID FLAGS RFC822.HEADER BODY.PEEK[TEXT]<0.800>)")
                             if st != "OK" or not msg_data:
                                 continue
+                            flags_blob = b" ".join(
+                                part[0] for part in msg_data
+                                if isinstance(part, tuple) and part and isinstance(part[0], (bytes, bytearray))
+                            )
+                            is_unread = b"\\Seen" not in flags_blob
                             # Headers + body land in different tuples in the
                             # response — concatenate the bytes for parsing.
                             raw = b""
@@ -1635,6 +2043,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                                 "headers": header_blob,
                                 "body": body_snippet.strip(),
                                 "message_id": (msg.get("Message-ID") or "").strip(),
+                                "unread": is_unread,
                             })
                         except Exception as _fe:
                             logger.debug(f"urgency: header fetch for uid {uid} failed: {_fe}")
@@ -1652,25 +2061,33 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
             for item in items:
                 scanned += 1
                 key = item["key"]
-                all_unread_keys.add(key)
+                if item.get("unread"):
+                    all_unread_keys.add(key)
                 if item.get("cached"):
-                    per_uid_scores[key] = item["cached"]
+                    cached_v = dict(item["cached"])
+                    cached_v["unread"] = bool(item.get("unread"))
+                    per_uid_scores[key] = cached_v
                     continue
                 # Skip uids we couldn't fetch (no subject/from/body).
                 if not item.get("subject") and not item.get("from"):
                     continue
+                verdict = _heuristic_email_verdict(item)
+                cache.setdefault("uids", {})[item["uid"]] = verdict
+                per_uid_scores[key] = verdict
+                saved_classifications += 1
+                continue
                 # ── LLM-classify. JSON-only response; bullet-proof parse.
                 llm_attempts += 1
                 prompt = (
-                    "You are triaging ONE unread email. Return ONLY JSON: "
+                    "You are triaging ONE email. Return ONLY JSON: "
                     "{\"score\":0|1|2|3,\"tags\":[\"...\"],\"spam\":false,"
                     "\"reason\":\"one short phrase\"}.\n"
                     "0 = trivial / promotional · 1 = informational, no reply needed · "
                     "2 = should reply within a day · 3 = urgent, reply now (deadline, blocker).\n\n"
-                    "Allowed tags: newsletter, marketing, notification, finance, bills, receipt, "
-                    "travel, security, shopping, social, work, personal, calendar.\n"
-                    "Use marketing for ads, promos, sales, offers, and cold sales. Use newsletter "
-                    "for newsletters, digests, and recurring content. spam=true for scams, phishing, "
+                    "Allowed visible tags: urgent, reply-soon, action-needed, calendar, bills, receipt, travel.\n"
+                    "Use action-needed when the user likely needs to reply, pay, sign, book, or decide. "
+                    "Use bills for bills or debts, receipt for purchases/deliveries, travel for reservations/trips, "
+                    "and calendar only when a calendar event/reminder is involved. spam=true for scams, phishing, "
                     "junk, cold sales, generic ads, or no-personal-action bulk mail.\n"
                     "Important: 'I'm outside', 'I am outside', 'waiting outside', 'at the door', "
                     "'locked out', or 'can't get in' means score 3 unless clearly historical.\n\n"
@@ -1679,6 +2096,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                     f"Snippet:\n{item.get('body','')}\n"
                 )
                 try:
+                    await wait_for_interactive_quiet("email urgency action")
                     raw = await llm_call_async_with_fallback(
                         candidates,
                         [{"role": "user", "content": prompt}],
@@ -1739,14 +2157,10 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                         r"\b(advertisement|sponsored|promo|promotion|sale|discount|offer|limited time|deal|tickets?|tour|merch|stream|purchase|sold out|low tickets|coupon|shop now|buy now)\b",
                         _blob,
                     ))
-                    if "newsletter" not in tags and bulkish:
-                        tags.append("newsletter")
-                    if "marketing" not in tags and marketingish:
-                        tags.append("marketing")
                     if (bulkish or marketingish) and score < 2:
                         score = 0
                         if not reason or "urgent" in reason.lower():
-                            reason = "Bulk marketing/newsletter; no personal reply needed"
+                            reason = "bulk mail; no personal reply needed"
                     # Strip "Name <addr>" to bare display name for compact summary.
                     _from_raw = item.get("from", "") or ""
                     if "<" in _from_raw:
@@ -1764,6 +2178,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                         # Cache the message_id too so re-scans of already-cached
                         # UIDs can still write the inbox tag without re-LLM'ing.
                         "message_id": (item.get("message_id") or "").strip(),
+                        "unread": bool(item.get("unread")),
                         "ts": _time.time(),
                     }
                     cache.setdefault("uids", {})[item["uid"]] = verdict
@@ -1778,9 +2193,9 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                     logger.debug(f"urgency: LLM classify failed for {key}: {e}")
                     continue
 
-            # ── Prune cache entries for UIDs that are no longer unread (replied
-            # / archived / deleted). Compare against `items` (everything UNSEEN
-            # in this scan window).
+            # ── Prune cache entries for UIDs that are no longer in the recent
+            # scan window. Read messages remain cached because tags are useful
+            # on read mail too; unread state is refreshed per scan above.
             seen_uids = {it["uid"] for it in items}
             cache_uids = cache.get("uids", {})
             for stale in [u for u in cache_uids if u not in seen_uids]:
@@ -1815,15 +2230,17 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                         _tag = str(_tag).strip().lower().replace("_", "-")
                         if _tag == "promo":
                             _tag = "marketing"
-                        if _tag in CATEGORY_TAGS and _tag not in _new_tags:
+                        if _tag == "action-needed" and any(t in _new_tags for t in ("urgent", "reply-soon")):
+                            continue
+                        if _tag in VISIBLE_EMAIL_TAGS and _tag not in _new_tags:
                             _new_tags.append(_tag)
                     _spam = 1 if _v.get("spam") else 0
                     # _key is "<account_id>:<uid>" — extract uid for the row.
-                    _uid_only = _key.split(":", 1)[-1]
+                    _acc_id, _uid_only = (_key.split(":", 1) + [""])[:2]
                     _owner_key = owner or ""
                     _row = _conn.execute(
-                        "SELECT tags FROM email_tags WHERE message_id=? AND owner=?",
-                        (_msg_id, _owner_key),
+                        "SELECT tags FROM email_tags WHERE message_id=? AND owner=? AND account_id=?",
+                        (_msg_id, _owner_key, _acc_id),
                     ).fetchone()
                     if _row:
                         try:
@@ -1842,23 +2259,42 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
                         for _tag in _new_tags:
                             if _tag not in _existing:
                                 _existing.append(_tag)
+                        if _new_tags or _spam:
+                            tag_write_details.append({
+                                "uid": _uid_only,
+                                "subject": _v.get("subject", ""),
+                                "from": _v.get("from", ""),
+                                "tags": list(_new_tags),
+                                "spam": _spam,
+                                "reason": _v.get("reason", ""),
+                                "updated": True,
+                            })
                         _conn.execute(
                             "UPDATE email_tags SET tags=?, spam_verdict=?, spam_reason=?, uid=?, folder=?, subject=?, sender=? "
-                            "WHERE message_id=? AND owner=?",
+                            "WHERE message_id=? AND owner=? AND account_id=?",
                             (_json.dumps(_existing), _spam, _v.get("reason", ""), _uid_only, "INBOX",
-                             _v.get("subject", ""), _v.get("from", ""), _msg_id, _owner_key),
+                             _v.get("subject", ""), _v.get("from", ""), _msg_id, _owner_key, _acc_id),
                         )
                     else:
                         if not _new_tags and not _spam:
                             continue
                         _conn.execute(
                             "INSERT INTO email_tags "
-                            "(message_id, owner, uid, folder, subject, sender, tags, spam_verdict, spam_reason, created_at) "
-                            "VALUES (?, ?, ?, 'INBOX', ?, ?, ?, ?, ?, ?)",
-                            (_msg_id, _owner_key, _uid_only, _v.get("subject", ""),
+                            "(message_id, owner, account_id, uid, folder, subject, sender, tags, spam_verdict, spam_reason, created_at) "
+                            "VALUES (?, ?, ?, ?, 'INBOX', ?, ?, ?, ?, ?, ?)",
+                            (_msg_id, _owner_key, _acc_id, _uid_only, _v.get("subject", ""),
                              _v.get("from", ""), _json.dumps(_new_tags), _spam, _v.get("reason", ""),
                              _dt2.utcnow().isoformat()),
                         )
+                        tag_write_details.append({
+                            "uid": _uid_only,
+                            "subject": _v.get("subject", ""),
+                            "from": _v.get("from", ""),
+                            "tags": list(_new_tags),
+                            "spam": _spam,
+                            "reason": _v.get("reason", ""),
+                            "updated": False,
+                        })
                 _conn.commit()
             finally:
                 _conn.close()
@@ -1866,7 +2302,7 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
             logger.warning(f"urgency: bulk tag write failed: {_te}")
 
         # ── 4. Aggregate state. urgent = score ≥ 2.
-        urgent_keys = [k for k, v in per_uid_scores.items() if v.get("score", 0) >= 2]
+        urgent_keys = [k for k, v in per_uid_scores.items() if v.get("score", 0) >= 2 and v.get("unread")]
         max_score = max((v.get("score", 0) for v in per_uid_scores.values()), default=0)
         total_urgent = len(urgent_keys)
 
@@ -1975,13 +2411,28 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
             f"reply-soon {tier_counts[2]} · info {tier_counts[1]} · trivial {tier_counts[0]} · "
             f"{saved_classifications} saved classifications"
         )
-        if llm_attempts != saved_classifications:
-            head += f" · {llm_attempts - saved_classifications} failed"
+        if failed_classifications:
+            head += f" · {len(failed_classifications)} failed"
         if newly_notified:
             head += f" · notified {len(newly_notified)}"
         if notify_failed:
             head += f" · notify failed {len(notify_failed)}"
 
+        def _fmt_tag_write(v):
+            subj = (v.get("subject") or "(no subject)")[:80]
+            frm = v.get("from") or ""
+            tags = list(v.get("tags") or [])
+            if v.get("spam"):
+                tags.append("spam")
+            tag_txt = ", ".join(tags) if tags else "cleared managed tags"
+            why = v.get("reason") or ""
+            op = "updated" if v.get("updated") else "created"
+            line = f"- **{subj}**" + (f" — _{frm}_" if frm else "")
+            line += f" — `{tag_txt}` ({op})"
+            if why:
+                line += f" · {why}"
+            return line
+
         def _fmt_one(v, newly_notified_set, failed_set, key):
             subj = (v.get("subject") or "(no subject)")[:80]
             frm = v.get("from") or ""
@@ -1997,6 +2448,13 @@ async def action_check_email_urgency(owner: str, **kwargs) -> Tuple[str, bool]:
         for k, v in per_uid_scores.items():
             by_tier.setdefault(v.get("score", 0), []).append((k, v))
         lines = [head]
+        if tag_write_details:
+            lines.append("")
+            lines.append(f"**Applied tags ({len(tag_write_details)}):**")
+            for v in tag_write_details[:16]:
+                lines.append(_fmt_tag_write(v))
+            if len(tag_write_details) > 16:
+                lines.append(f"…and {len(tag_write_details) - 16} more")
         tier_labels = {3: "Urgent", 2: "Reply soon", 1: "Informational", 0: "Trivial"}
         for tier in (3, 2, 1, 0):
             items_t = by_tier.get(tier, [])
@@ -2068,6 +2526,7 @@ async def action_cookbook_serve(
         end_after_min = int(cfg.get("end_after_min") or 0)
     except Exception:
         end_after_min = 0
+    set_default = bool(cfg.get("set_default", True))
 
     state_path = Path(COOKBOOK_STATE_FILE)
     try:
@@ -2154,6 +2613,51 @@ async def action_cookbook_serve(
         return f"Launch rejected: {data.get('error') or data.get('detail') or 'unknown'}", False
 
     sid = data.get("session_id") or ""
+    endpoint_id = data.get("endpoint_id") or ""
+    # Scheduled serves are usually meant to become the active local model for
+    # chat/tools while their time window is open. Persist both endpoint and
+    # model so task/utility/default resolution does not keep routing to a stale
+    # API fallback. Allow explicit opt-out with {"set_default": false}.
+    if endpoint_id and set_default:
+        try:
+            selected_model = repo_id
+            try:
+                from core.database import SessionLocal as _SL, ModelEndpoint as _ME
+                _db = _SL()
+                try:
+                    _ep = _db.query(_ME).filter(_ME.id == endpoint_id).first()
+                    if _ep and _ep.cached_models:
+                        _models = json.loads(_ep.cached_models or "[]")
+                        if isinstance(_models, list) and _models:
+                            selected_model = str(_models[0])
+                finally:
+                    _db.close()
+            except Exception:
+                pass
+            from src.settings import load_settings as _load_settings, save_settings as _save_settings
+            _settings = _load_settings()
+            _settings["default_endpoint_id"] = endpoint_id
+            _settings["default_model"] = selected_model
+            # Keep background tasks aligned unless the user explicitly chose a
+            # separate task model.
+            if not (_settings.get("task_endpoint_id") or "").strip():
+                _settings["task_endpoint_id"] = endpoint_id
+                _settings["task_model"] = selected_model
+            if not (_settings.get("utility_endpoint_id") or "").strip():
+                _settings["utility_endpoint_id"] = endpoint_id
+                _settings["utility_model"] = selected_model
+            _save_settings(_settings)
+            if owner:
+                from routes.prefs_routes import _load_for_user, _save_for_user
+                _prefs = _load_for_user(owner)
+                _prefs["default_endpoint_id"] = endpoint_id
+                _prefs["default_model"] = selected_model
+                if not (_prefs.get("utility_endpoint_id") or "").strip():
+                    _prefs["utility_endpoint_id"] = endpoint_id
+                    _prefs["utility_model"] = selected_model
+                _save_for_user(owner, _prefs)
+        except Exception as e:
+            logger.warning(f"cookbook_serve: default endpoint update failed: {e}")
     # Register the new task in cookbook_state.json + stamp it with our
     # scheduler-owner markers. /api/model/serve spawns the tmux session
     # but leaves the state-write to the UI — when a scheduled action
@@ -2197,12 +2701,16 @@ async def action_cookbook_serve(
                     "sshPort": ssh_port or "",
                     "platform": platform or "linux",
                     "_serveReady": False,
-                    "_endpointAdded": False,
+                    "_endpointAdded": bool(endpoint_id),
                 }
                 tasks.append(existing)
             # Stamp ownership + end-at on the task entry.
             existing["_scheduledByTask"] = task_name or ""
             existing["_scheduledByOwner"] = owner or ""
+            if endpoint_id:
+                existing["_endpointId"] = endpoint_id
+                existing["endpointId"] = endpoint_id
+                existing["_endpointAdded"] = True
             if end_after_min > 0:
                 existing["_scheduledStopAtMs"] = int(_time.time() * 1000) + end_after_min * 60 * 1000
             fresh["tasks"] = tasks
@@ -2230,6 +2738,7 @@ BUILTIN_ACTIONS = {
     "tidy_research": action_tidy_research,
     "summarize_emails": action_summarize_emails,
     "draft_email_replies": action_draft_email_replies,
+    "email_auto_translate": action_email_auto_translate,
     "extract_email_events": action_extract_email_events,
     "classify_events": action_classify_events,
     # ping_events removed from the user-facing registry. Calendar reminders
@@ -2254,6 +2763,7 @@ BUILTIN_ACTION_INFO = {
     "tidy_research": "Remove orphaned research files (sessions that were deleted)",
     "summarize_emails": "Pre-generate AI summaries for new inbox emails",
     "draft_email_replies": "Pre-draft AI reply suggestions for new inbox emails",
+    "email_auto_translate": "Detect foreign-language emails and cache translated text for the email reader",
     "extract_email_events": "Scan emails for booking/meeting confirmations and auto-add to calendar",
     "classify_events": "Tag upcoming events with importance (low/normal/high/critical) and type (work/health/travel/etc.); colors them too",
     "daily_brief": "Build a morning digest: today's calendar, unread email count + top senders, active todos",

src/caldav_sync.py

@@ -25,6 +25,7 @@ Design notes:
 import asyncio
 import hashlib
 import ipaddress
+import json
 import logging
 import os
 import socket
@@ -500,6 +501,7 @@ def _event_payload(ev) -> dict:
         "all_day": ev.all_day,
         "is_utc": ev.is_utc,
         "rrule": ev.rrule or "",
+        "recurrence_exdates": json.loads(ev.recurrence_exdates or "[]") if getattr(ev, "recurrence_exdates", "") else [],
     }
 
 

src/caldav_writeback.py

@@ -33,7 +33,8 @@ def build_event_ical(ev: dict) -> str:
     """Serialize a local event dict to a VCALENDAR/VEVENT iCalendar string.
 
     ``ev`` keys: uid, summary, description, location, dtstart (datetime),
-    dtend (datetime), all_day (bool), is_utc (bool), rrule (str).
+    dtend (datetime), all_day (bool), is_utc (bool), rrule (str),
+    recurrence_exdates (list[str]).
     Mirrors how the pull path interprets is_utc/all_day so a round-trip is stable.
     """
     from icalendar import Calendar, Event as iEvent
@@ -70,6 +71,15 @@ def build_event_ical(ev: dict) -> str:
             ve.add("rrule", vRecur.from_ical(ev["rrule"]))
         except Exception:
             logger.debug("CalDAV write-back: skipping unparseable rrule %r", ev.get("rrule"))
+    for exdate in ev.get("recurrence_exdates") or []:
+        try:
+            if ev.get("all_day"):
+                ve.add("exdate", datetime.strptime(exdate[:10], "%Y-%m-%d").date())
+            else:
+                dt = datetime.strptime(exdate[:16], "%Y-%m-%dT%H:%M")
+                ve.add("exdate", dt.replace(tzinfo=timezone.utc) if ev.get("is_utc") else dt)
+        except Exception:
+            logger.debug("CalDAV write-back: skipping unparseable exdate %r", exdate)
 
     cal.add_component(ve)
     return cal.to_ical().decode("utf-8")

src/chat_handler.py

@@ -29,6 +29,34 @@ from src.youtube_handler import (
 logger = logging.getLogger(__name__)
 
 
+def _sync_upload_vision_to_gallery(file_info: Dict[str, Any], owner: Optional[str], text: str) -> None:
+    file_hash = (file_info or {}).get("hash")
+    if not file_hash or not text:
+        return
+    try:
+        from core.database import GalleryImage, SessionLocal
+        db = SessionLocal()
+        try:
+            q = db.query(GalleryImage).filter(
+                GalleryImage.file_hash == file_hash,
+                GalleryImage.is_active == True,  # noqa: E712
+            )
+            if owner:
+                q = q.filter(GalleryImage.owner == owner)
+            img = q.first()
+            if not img:
+                return
+            img.caption = text.strip()
+            db.commit()
+        except Exception:
+            db.rollback()
+            raise
+        finally:
+            db.close()
+    except Exception as e:
+        logger.warning("Failed to sync upload vision text to gallery: %s", e)
+
+
 class ChatHandler:
     """Handles chat operations for both streaming and non-streaming endpoints."""
 
@@ -207,6 +235,7 @@ class ChatHandler:
                                     _vtext = _vf.read().strip()
                                 if _vtext:
                                     enhanced_message += f"\n[User-corrected caption / OCR for this image — treat as authoritative]:\n{_vtext}"
+                                    _sync_upload_vision_to_gallery(file_info, owner, _vtext)
                                     _m = meta_by_id.get(att_id)
                                     if _m is not None:
                                         _m["vision"] = _vtext
@@ -226,6 +255,7 @@ class ChatHandler:
                                     cached_desc = _vf.read().strip()
                                 if cached_desc and not cached_desc.startswith("["):
                                     vl_desc = cached_desc
+                                    _sync_upload_vision_to_gallery(file_info, owner, vl_desc)
                             except Exception:
                                 vl_desc = None
                         if not vl_desc:
@@ -237,6 +267,7 @@ class ChatHandler:
                                     os.makedirs(os.path.join(UPLOAD_DIR, ".vision"), exist_ok=True)
                                     with open(_vcache, "w", encoding="utf-8") as _vf:
                                         _vf.write(vl_desc)
+                                    _sync_upload_vision_to_gallery(file_info, owner, vl_desc)
                                 except Exception:
                                     pass
                         enhanced_message = f"{enhanced_message}\n\n[Image: {file_info['name']}]\n{vl_desc}"

src/context_budget.py

@@ -18,6 +18,13 @@ DEFAULT_BUDGET = 6000
 DEFAULT_HEADROOM = 0.85
 
 
+def _int_or_zero(value) -> int:
+    try:
+        return int(value or 0)
+    except (TypeError, ValueError):
+        return 0
+
+
 def compute_input_token_budget(
     configured: int,
     context_length: int,
@@ -48,8 +55,8 @@ def compute_input_token_budget(
         - When the window is unknown (context_length <= 0), use the conservative
           ``default`` budget and do NOT scale off the fallback.
     """
-    configured = int(configured or 0)
-    context_length = int(context_length or 0)
+    configured = _int_or_zero(configured)
+    context_length = _int_or_zero(context_length)
 
     if explicit and configured > 0:
         return min(configured, context_length) if context_length > 0 else configured

src/cookbook_serve_lifecycle.py

@@ -37,6 +37,13 @@ async def _delete_endpoint_for_task(task: dict) -> None:
     the picker (probe goes offline; chats still try to route there) and
     the user has to delete it by hand in Settings -> Endpoints.
     """
+    endpoint_id = (task.get("_endpointId") or task.get("endpointId") or "").strip()
+    if not endpoint_id:
+        logger.info(
+            "cookbook_serve_lifecycle: task %s has no endpoint id; skipping endpoint deletion",
+            task.get("sessionId") or task.get("id") or "",
+        )
+        return
     import re as _re
     payload = task.get("payload") or {}
     cmd = str(payload.get("_cmd") or "")
@@ -66,13 +73,10 @@ async def _delete_endpoint_for_task(task: dict) -> None:
             if r.status_code >= 400:
                 return
             eps = r.json() if r.content else []
-            # Prefer exact URL match; fall back to host:port substring so we
-            # still catch the case where 0.0.0.0 vs the registered host
-            # representation diverged.
-            ep = next((e for e in eps if e.get("base_url") == base_url), None)
-            if not ep:
-                hostport = f"{host}:{port}"
-                ep = next((e for e in eps if hostport in (e.get("base_url") or "")), None)
+            # Delete only the endpoint created by this scheduled serve. URL
+            # matching is unsafe because a later scheduled serve can reuse the
+            # same host:port after an older task has gone stale.
+            ep = next((e for e in eps if e.get("id") == endpoint_id), None)
             if ep:
                 await client.delete(
                     f"{internal_api_base()}/api/model-endpoints/{ep['id']}",

src/document_actions.py

@@ -6,7 +6,7 @@ Reusable document actions callable from both REST routes and the task scheduler.
 
 import logging
 import re
-from datetime import datetime
+from datetime import datetime, timezone
 
 logger = logging.getLogger(__name__)
 
@@ -77,10 +77,28 @@ async def run_document_tidy(owner: str) -> str:
         deleted = 0
         kept = 0
         survivors = []  # docs that pass the junk rules, considered for dedup
+        now = datetime.now(timezone.utc)
 
         for doc in docs:
+            created = doc.created_at
+            if created and created.tzinfo is None:
+                created = created.replace(tzinfo=timezone.utc)
+
+            # Skip freshly created documents to avoid deleting them while the user is actively editing
+            if created and (now - created).total_seconds() < 900:  # 15 minutes
+                survivors.append(doc)
+                continue
+
             content = (doc.current_content or "").strip()
             title = (doc.title or "").strip().lower()
+            is_fresh_empty = (
+                not content
+                and created is not None
+                and (now - created).total_seconds() < 1800
+            )
+            if is_fresh_empty:
+                survivors.append(doc)
+                continue
 
             # Strip markdown noise to get "real" character count
             stripped = re.sub(r"^#{1,6}\s+", "", content, flags=re.MULTILINE)  # headers

src/endpoint_resolver.py

@@ -47,6 +47,33 @@ def _endpoint_cached_models(ep) -> list:
     return models if isinstance(models, list) else []
 
 
+def _endpoint_pinned_models(ep) -> list:
+    raw = getattr(ep, "pinned_models", None)
+    if not raw:
+        return []
+    try:
+        models = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        return []
+    return models if isinstance(models, list) else []
+
+
+def _is_mlx_deepseek_v4_repo_id(model_id: str) -> bool:
+    return "mlx-community/deepseek-v4" in str(model_id or "").lower()
+
+
+def _is_mlx_deepseek_v4_shim_id(model_id: str) -> bool:
+    return "/.cache/odysseus/mlx-shims/deepseek-v4" in str(model_id or "").lower()
+
+
+def _filter_mlx_deepseek_v4_repo_when_shimmed(model_ids) -> list:
+    ids = list(model_ids or [])
+    has_shim = any(_is_mlx_deepseek_v4_shim_id(m) for m in ids)
+    if not has_shim:
+        return ids
+    return [m for m in ids if not _is_mlx_deepseek_v4_repo_id(m)]
+
+
 def _endpoint_hidden_models(ep) -> set:
     """Model ids the admin disabled on this endpoint (the UI's hidden list)."""
     raw = getattr(ep, "hidden_models", None)
@@ -67,7 +94,15 @@ def _endpoint_enabled_models(ep) -> list:
     raw first one resolves to a model that 400s ("requires terms acceptance").
     """
     hidden = _endpoint_hidden_models(ep)
-    return [m for m in _endpoint_cached_models(ep) if m not in hidden]
+    merged = []
+    seen = set()
+    for m in [*_endpoint_cached_models(ep), *_endpoint_pinned_models(ep)]:
+        if not isinstance(m, str) or not m or m in seen:
+            continue
+        seen.add(m)
+        merged.append(m)
+    merged = _filter_mlx_deepseek_v4_repo_when_shimmed(merged)
+    return [m for m in merged if m not in hidden]
 
 
 def resolve_endpoint_runtime(ep, owner: Optional[str] = None) -> Tuple[str, Optional[str]]:

src/host_docker_access.py

@@ -0,0 +1,62 @@
+"""Policy checks for explicit host Docker access from a container."""
+
+import os
+import stat
+from collections.abc import Mapping
+
+
+HOST_DOCKER_ENV_VAR = "ODYSSEUS_ENABLE_HOST_DOCKER"
+HOST_DOCKER_SOCKET_PATH = "/var/run/docker.sock"
+
+HOST_DOCKER_ACCESS_HINT = (
+    "Local Docker daemon access is disabled inside the Odysseus container; a "
+    "Docker CLI alone is not enough. Default Docker Compose intentionally does "
+    "not mount the host Docker socket. Raw socket access is high-trust and can "
+    "grant broad control over the host Docker daemon. If you accept that risk, "
+    "enable docker/host-docker.yml. Remote server Docker workflows over SSH "
+    "remain preferred."
+)
+
+
+def running_in_container(
+    dockerenv_path: str = "/.dockerenv",
+    cgroup_path: str = "/proc/1/cgroup",
+) -> bool:
+    if os.path.exists(dockerenv_path):
+        return True
+    try:
+        with open(cgroup_path, "r", encoding="utf-8") as handle:
+            contents = handle.read()
+    except OSError:
+        return False
+    return any(token in contents for token in ("docker", "containerd", "kubepods"))
+
+
+def host_docker_access_enabled(
+    socket_path: str = HOST_DOCKER_SOCKET_PATH,
+    *,
+    environ: Mapping[str, str] | None = None,
+) -> bool:
+    env = os.environ if environ is None else environ
+    if env.get(HOST_DOCKER_ENV_VAR, "").strip().lower() != "true":
+        return False
+    try:
+        mode = os.stat(socket_path).st_mode
+    except OSError:
+        return False
+    return stat.S_ISSOCK(mode)
+
+
+def local_docker_available(
+    *,
+    cli_available: bool,
+    in_container: bool | None = None,
+    environ: Mapping[str, str] | None = None,
+    socket_path: str = HOST_DOCKER_SOCKET_PATH,
+) -> bool:
+    if not cli_available:
+        return False
+    containerized = running_in_container() if in_container is None else in_container
+    if not containerized:
+        return True
+    return host_docker_access_enabled(socket_path, environ=environ)

src/integrations.py

@@ -216,7 +216,14 @@ def _normalize_integration_base_url(base_url: Any) -> str:
 
 
 def _join_integration_url(base_url: str, path: str) -> str:
-    return urljoin(base_url.rstrip("/") + "/", path.lstrip("/"))
+    base = base_url.rstrip("/")
+    rel = path.lstrip("/")
+    if not rel:
+        # A bare "/" must resolve to the base URL itself, not base + "/".
+        # POST-to-base integrations (e.g. Discord webhooks) 404 on the
+        # trailing-slash variant of their URL.
+        return base
+    return urljoin(base + "/", rel)
 
 
 def load_integrations() -> List[Dict[str, Any]]:
@@ -394,6 +401,22 @@ async def execute_api_call(
         return {"error": "Path must not contain a fragment", "exit_code": 1}
 
     url = _join_integration_url(base_url, path)
+
+    # SSRF guard — same check used by the gallery endpoint, embeddings,
+    # CardDAV, and the reminder webhook sender. Link-local / metadata
+    # addresses (169.254.x.x — the cloud credential-exfil vector) are always
+    # rejected; INTEGRATION_API_BLOCK_PRIVATE_IPS=true also blocks RFC-1918 /
+    # loopback for locked-down deployments. Private stays allowed by default
+    # because LAN integrations (Home Assistant, Miniflux, ntfy) are the
+    # primary use case.
+    from src.url_safety import check_outbound_url
+    block_private = os.getenv(
+        "INTEGRATION_API_BLOCK_PRIVATE_IPS", "false"
+    ).lower() == "true"
+    ok, reason = check_outbound_url(url, block_private=block_private)
+    if not ok:
+        return {"error": f"URL rejected: {reason}", "exit_code": 1}
+
     method = method.upper()
 
     # Build headers

src/interactive_gate.py

@@ -0,0 +1,203 @@
+"""Foreground activity gate for background work.
+
+Background tasks are allowed to run only after normal UI/API traffic has
+settled. This keeps scheduled jobs and email pollers from competing with the
+user opening Odysseus, Cookbook, email, documents, notes, or other panels.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from contextlib import asynccontextmanager
+import os
+import time
+
+
+_ACTIVE_REQUESTS = 0
+_LAST_ACTIVITY = 0.0
+_LAST_BROWSER_ACTIVITY = 0.0
+_COND: asyncio.Condition | None = None
+_COND_LOOP: asyncio.AbstractEventLoop | None = None
+
+
+def _enabled() -> bool:
+    return os.getenv("BACKGROUND_TASK_FOREGROUND_GATE", "true").lower() not in {"0", "false", "no", "off"}
+
+
+def _quiet_seconds() -> float:
+    try:
+        return max(0.0, float(os.getenv("BACKGROUND_TASK_QUIET_MS", "1500")) / 1000.0)
+    except Exception:
+        return 1.5
+
+
+def _max_wait_seconds() -> float:
+    """0 means wait indefinitely until the UI is quiet."""
+    try:
+        return max(0.0, float(os.getenv("BACKGROUND_TASK_MAX_WAIT_SECONDS", "0")))
+    except Exception:
+        return 0.0
+
+
+def _browser_active_seconds() -> float:
+    """How long a visible Odysseus browser heartbeat blocks background tasks."""
+    try:
+        return max(0.0, float(os.getenv("BACKGROUND_TASK_BROWSER_ACTIVE_SECONDS", "45")))
+    except Exception:
+        return 45.0
+
+
+def _condition() -> asyncio.Condition:
+    global _COND, _COND_LOOP
+    try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        loop = None
+    if _COND is None or _COND_LOOP is not loop:
+        _COND = asyncio.Condition()
+        _COND_LOOP = loop
+    return _COND
+
+
+_PASSIVE_EXACT_PATHS = {
+    "/api/activity/heartbeat",
+    "/api/client-perf",
+    "/api/tasks/notifications",
+    "/api/research/active",
+    "/api/email/urgency-state",
+}
+
+_PASSIVE_PREFIXES = (
+    "/api/chat/stream_status",
+    "/api/health",
+    "/api/prefs",
+)
+
+
+def should_track_interactive_request(path: str, method: str = "GET") -> bool:
+    if not _enabled():
+        return False
+    if (method or "").upper() == "OPTIONS":
+        return False
+    if path in _PASSIVE_EXACT_PATHS:
+        return False
+    if any(path.startswith(prefix) for prefix in _PASSIVE_PREFIXES):
+        return False
+    return True
+
+
+async def mark_browser_activity() -> None:
+    """Record that an authenticated browser tab is visibly using Odysseus."""
+    global _LAST_BROWSER_ACTIVITY
+    if not _enabled():
+        return
+    cond = _condition()
+    async with cond:
+        _LAST_BROWSER_ACTIVITY = time.monotonic()
+        cond.notify_all()
+
+
+def _has_recent_browser_activity(now: float | None = None) -> bool:
+    ttl = _browser_active_seconds()
+    if ttl <= 0 or _LAST_BROWSER_ACTIVITY <= 0:
+        return False
+    return ((now if now is not None else time.monotonic()) - _LAST_BROWSER_ACTIVITY) < ttl
+
+
+def has_foreground_activity(now: float | None = None) -> bool:
+    """Return True when foreground browser/model work should stop background jobs.
+
+    Passive polling endpoints are excluded by should_track_interactive_request,
+    so active/recent request tracking is safe to use here. This matters during
+    initial page load: the heartbeat may not have landed yet, but the user is
+    already waiting on real UI requests.
+    """
+    if not _enabled():
+        return False
+    t = now if now is not None else time.monotonic()
+    if _ACTIVE_REQUESTS > 0:
+        return True
+    if _LAST_ACTIVITY > 0 and (t - _LAST_ACTIVITY) < _quiet_seconds():
+        return True
+    return _has_recent_browser_activity(t) or _has_active_chat_stream()
+
+
+def _has_active_chat_stream() -> bool:
+    """Best-effort check for foreground model work that outlives HTTP requests.
+
+    Chat/agent streams are detached from the browser SSE so a stream can keep
+    running after the request that started it has returned. Background LLM
+    tasks must still wait for those runs; otherwise helpers like email
+    auto-translate compete with the user's active chat on the same local model.
+    """
+    try:
+        from routes import chat_routes as _chat_routes
+        active_streams = getattr(_chat_routes, "_active_streams", {}) or {}
+        if active_streams:
+            return True
+    except Exception:
+        pass
+    try:
+        from src import agent_runs
+        runs = getattr(agent_runs, "_RUNS", {}) or {}
+        return any(getattr(run, "status", None) == "running" for run in runs.values())
+    except Exception:
+        return False
+
+
+@asynccontextmanager
+async def track_interactive_request(path: str = "", method: str = ""):
+    global _ACTIVE_REQUESTS, _LAST_ACTIVITY
+    if not _enabled():
+        yield
+        return
+
+    cond = _condition()
+    async with cond:
+        _ACTIVE_REQUESTS += 1
+        _LAST_ACTIVITY = time.monotonic()
+        cond.notify_all()
+    try:
+        yield
+    finally:
+        async with cond:
+            _ACTIVE_REQUESTS = max(0, _ACTIVE_REQUESTS - 1)
+            _LAST_ACTIVITY = time.monotonic()
+            cond.notify_all()
+
+
+async def wait_for_interactive_quiet(label: str = "") -> bool:
+    """Wait until foreground requests have stopped for the configured window.
+
+    Returns True if the caller had to wait at all. The label is intentionally
+    only for future logging/debugging so callers can keep their code simple.
+    """
+    if not _enabled():
+        return False
+
+    quiet = _quiet_seconds()
+    max_wait = _max_wait_seconds()
+    deadline = time.monotonic() + max_wait if max_wait > 0 else None
+    cond = _condition()
+    waited = False
+
+    while True:
+        async with cond:
+            now = time.monotonic()
+            quiet_remaining = quiet - (now - _LAST_ACTIVITY)
+            active_stream = _has_active_chat_stream()
+            browser_active = _has_recent_browser_activity(now)
+            if _ACTIVE_REQUESTS <= 0 and quiet_remaining <= 0 and not active_stream and not browser_active:
+                return waited
+
+            waited = True
+            timeout = 0.25 if (_ACTIVE_REQUESTS > 0 or active_stream or browser_active) else min(max(quiet_remaining, 0.05), 0.5)
+            if deadline is not None:
+                remaining = deadline - now
+                if remaining <= 0:
+                    return waited
+                timeout = min(timeout, remaining)
+            try:
+                await asyncio.wait_for(cond.wait(), timeout=timeout)
+            except asyncio.TimeoutError:
+                pass

src/llm_core.py

@@ -8,13 +8,88 @@ import hashlib
 import threading
 import re
 import os
+from contextlib import asynccontextmanager
 from fastapi import HTTPException
 from typing import Optional, Dict, List, Tuple
-from src.model_context import get_context_length, DEFAULT_CONTEXT
+from src.model_context import get_context_length, DEFAULT_CONTEXT, is_local_endpoint
 from urllib.parse import urlparse
 
 logger = logging.getLogger(__name__)
 
+_LOCAL_MODEL_LOCK = asyncio.Lock()
+_LOCAL_MODEL_WAITING_FOREGROUND = 0
+_LOCAL_MODEL_CURRENT: Dict[str, object] = {}
+
+
+def _local_model_gate_enabled() -> bool:
+    return os.getenv("ODYSSEUS_LOCAL_MODEL_GATE", "true").lower() not in {"0", "false", "no", "off"}
+
+
+def _gate_workload(workload: Optional[str]) -> str:
+    return "background" if str(workload or "").lower() == "background" else "foreground"
+
+
+@asynccontextmanager
+async def _local_model_slot(target_url: str, model: str, workload: Optional[str] = None):
+    """Serialize local model traffic, with foreground chat taking priority.
+
+    Most local servers expose one GPU/CPU generation pipe even when their HTTP
+    API accepts multiple requests. Letting scheduled email/tasks and foreground
+    chat hit that pipe together creates the user-visible "streams crossed" and
+    "prompt waited behind a task" failure mode. Cloud providers are left alone.
+    """
+    if not _local_model_gate_enabled() or not is_local_endpoint(target_url):
+        yield
+        return
+
+    global _LOCAL_MODEL_WAITING_FOREGROUND
+    kind = _gate_workload(workload)
+    current_task = asyncio.current_task()
+    if kind == "foreground":
+        _LOCAL_MODEL_WAITING_FOREGROUND += 1
+        current = dict(_LOCAL_MODEL_CURRENT)
+        if current.get("workload") == "background":
+            task = current.get("task")
+            if isinstance(task, asyncio.Task) and not task.done():
+                logger.info(
+                    "[model-gate] cancelling background local model call for foreground request model=%s",
+                    model,
+                )
+                task.cancel()
+    else:
+        # Background work should not jump in while the browser/chat is active
+        # or while a foreground request is waiting to acquire the local model.
+        try:
+            from src.interactive_gate import has_foreground_activity
+        except Exception:
+            has_foreground_activity = lambda: False  # type: ignore
+        while _LOCAL_MODEL_WAITING_FOREGROUND > 0 or has_foreground_activity():
+            await asyncio.sleep(0.25)
+
+    acquired = False
+    try:
+        await _LOCAL_MODEL_LOCK.acquire()
+        acquired = True
+        if kind == "foreground":
+            _LOCAL_MODEL_WAITING_FOREGROUND = max(0, _LOCAL_MODEL_WAITING_FOREGROUND - 1)
+        _LOCAL_MODEL_CURRENT.clear()
+        _LOCAL_MODEL_CURRENT.update({
+            "task": current_task,
+            "workload": kind,
+            "url": target_url,
+            "model": model,
+            "started": time.time(),
+        })
+        yield
+    finally:
+        if kind == "foreground":
+            _LOCAL_MODEL_WAITING_FOREGROUND = max(0, _LOCAL_MODEL_WAITING_FOREGROUND - 1)
+        if acquired and _LOCAL_MODEL_LOCK.locked():
+            owner = _LOCAL_MODEL_CURRENT.get("task")
+            if owner is current_task:
+                _LOCAL_MODEL_CURRENT.clear()
+            _LOCAL_MODEL_LOCK.release()
+
 class LLMConfig:
     """Configuration constants for LLM operations."""
     DEFAULT_TIMEOUT = 30
@@ -110,6 +185,18 @@ _HARMONY_MARKERS = (
 )
 _HARMONY_MAX_MARKER_LEN = max(len(marker) for marker in _HARMONY_MARKERS)
 
+_VISIBLE_CHAT_TEMPLATE_ARTIFACT_RE = re.compile(
+    r"(?:\|end\|)+\|?assistan(?:t)?\|?"
+    r"|\|assistan(?:t)?\|"
+    r"|<\|im_start\|>\s*assistant"
+    r"|<\|im_end\|>",
+    re.IGNORECASE,
+)
+
+
+def _strip_visible_chat_template_artifacts(text: str) -> str:
+    return _VISIBLE_CHAT_TEMPLATE_ARTIFACT_RE.sub("", text or "")
+
 
 def _harmony_suffix_hold_len(text: str) -> int:
     """Return how many trailing chars could be the start of a harmony marker."""
@@ -187,6 +274,75 @@ def _stream_delta_event(text: str, *, thinking: bool = False) -> str:
         payload["thinking"] = True
     return f"data: {json.dumps(payload)}\n\n"
 
+
+_DEGENERATE_WORD_RE = re.compile(r"[A-Za-z0-9_\u0370-\u03ff\u0400-\u04ff]+")
+
+
+class _DegenerateStreamGuard:
+    """Detect local-model token collapse before it floods the UI.
+
+    Some self-hosted models fail by repeating one token forever ("Var Var Var",
+    "Summer Summer ..."). This is not a useful response and can burn context,
+    browser memory, and GPU time. Keep the guard conservative: only fire on long
+    same-token runs or a very dominant repeated token in the recent window.
+    """
+
+    def __init__(self, model: str):
+        self.model = model or "model"
+        self.last_token = ""
+        self.same_run = 0
+        self.recent_tokens: List[str] = []
+        self.total_chars = 0
+
+    def check(self, text: str) -> Optional[str]:
+        if not text:
+            return None
+        self.total_chars += len(text)
+        tokens = [t.lower() for t in _DEGENERATE_WORD_RE.findall(text) if len(t) >= 2]
+        if not tokens:
+            return None
+        for token in tokens:
+            if token == self.last_token:
+                self.same_run += 1
+            else:
+                self.last_token = token
+                self.same_run = 1
+            self.recent_tokens.append(token)
+        if len(self.recent_tokens) > 96:
+            self.recent_tokens = self.recent_tokens[-96:]
+
+        reason = None
+        if self.same_run >= 28 and self.total_chars >= 100:
+            reason = f"repeated '{self.last_token}' {self.same_run} times"
+        elif len(self.recent_tokens) >= 72:
+            top = max(set(self.recent_tokens), key=self.recent_tokens.count)
+            count = self.recent_tokens.count(top)
+            if count >= 60 and count / max(len(self.recent_tokens), 1) >= 0.78:
+                reason = f"repeated '{top}' {count}/{len(self.recent_tokens)} recent tokens"
+        if not reason and len(self.recent_tokens) >= 80:
+            # Phrase loops are common on some local quantized MLX/MoE models:
+            # "Also be a software developer mode?" repeated forever will not
+            # trip the single-token guard above, but it is still a wedged
+            # generation. Require many repeats of the same 4-gram so normal
+            # prose/list formatting is not interrupted.
+            grams = [tuple(self.recent_tokens[i:i + 4]) for i in range(0, len(self.recent_tokens) - 3)]
+            if grams:
+                top_gram = max(set(grams), key=grams.count)
+                gram_count = grams.count(top_gram)
+                if gram_count >= 10:
+                    reason = f"repeated phrase '{' '.join(top_gram)}' {gram_count} times"
+
+        if not reason:
+            return None
+
+        logger.warning("[degenerate-stream] aborting model=%s reason=%s", self.model, reason)
+        message = (
+            f"Stopped generation: {self.model} started repeating tokens "
+            f"({reason}). Try a different model or lower temperature."
+        )
+        return f'event: error\ndata: {json.dumps({"status": 502, "text": message, "error": message})}\n\n'
+
+
 def _model_activity_key(url: str, model: str) -> str:
     return f"{(url or '').strip()}|{(model or '').strip()}"
 
@@ -345,6 +501,18 @@ def _normalize_ollama_url(url: str) -> str:
     return base.rstrip("/") + "/chat"
 
 
+def _normalize_openai_chat_url(url: str) -> str:
+    """Ensure an OpenAI-compatible base URL points at /chat/completions."""
+    base = (url or "").strip().rstrip("/")
+    if not base:
+        return base
+    if base.endswith("/chat/completions") or base.endswith("/completions"):
+        return base
+    if base.endswith("/models"):
+        base = base[: -len("/models")].rstrip("/")
+    return base + "/chat/completions"
+
+
 def _ollama_normalize_messages(messages: List[Dict]) -> List[Dict]:
     """Adapt Odysseus' canonical OpenAI-style messages to native Ollama /api/chat.
 
@@ -731,6 +899,52 @@ def _apply_local_cache_affinity(payload: Dict, url: str, session_id: Optional[st
     payload.setdefault("cache_prompt", True)
 
 
+def _is_local_minimax_mlx_request(url: str, model: str) -> bool:
+    """Local MLX MiniMax-family endpoints need conservative sampling defaults.
+
+    The OpenAI-compatible MLX server accepts repetition/frequency penalties.
+    Some large quantized MiniMax/MoE ports otherwise fall into visible reasoning
+    loops ("Also be...", "No.", etc.) even for trivial prompts.
+    """
+    if not model:
+        return False
+    m = model.lower()
+    if "minimax" not in m and "mini-max" not in m:
+        return False
+    try:
+        from src.model_context import is_local_endpoint
+        return is_local_endpoint(url)
+    except Exception:
+        return False
+
+
+def _apply_local_generation_stability(payload: Dict, url: str, model: str) -> None:
+    if not _is_local_minimax_mlx_request(url, model):
+        return
+    if "temperature" in payload:
+        try:
+            # MiniMax MLX quantized ports are very sensitive to chat/agent
+            # harness size. Character presets can ask for a warmer voice, but
+            # local MiniMax needs a final compatibility clamp or trivial
+            # prompts can fall into visible reasoning/repetition loops.
+            payload["temperature"] = min(float(payload.get("temperature") or 0.2), 0.2)
+        except (TypeError, ValueError):
+            payload["temperature"] = 0.2
+    payload.setdefault("top_p", 0.9)
+    payload.setdefault("top_k", 20)
+    payload.setdefault("repetition_penalty", 1.12)
+    payload.setdefault("repetition_context_size", 256)
+    payload.setdefault("frequency_penalty", 0.08)
+    payload.setdefault("frequency_context_size", 256)
+    payload.setdefault("presence_penalty", 0.02)
+    payload.setdefault("presence_context_size", 256)
+    payload.setdefault("stop", ["<|im_end|>", "<|endoftext|>", "</s>"])
+    # A max_tokens of 0 means "server default/unbounded" for many local
+    # endpoints. Keep simple chats from running forever when the model loops.
+    if not payload.get("max_tokens") and not payload.get("max_completion_tokens"):
+        payload["max_tokens"] = 2048
+
+
 def _provider_headers(provider: str, headers: Optional[Dict] = None) -> Dict[str, str]:
     h = {"Content-Type": "application/json"}
     if isinstance(headers, dict):
@@ -1361,6 +1575,7 @@ def _sanitize_llm_messages(messages: List[Dict]) -> List[Dict]:
 
     return merged
 
+
 def _normalize_anthropic_url(url: str) -> str:
     """Ensure Anthropic URL points to /v1/messages."""
     url = url.rstrip("/")
@@ -1563,7 +1778,7 @@ def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LL
             stream=False, num_ctx=get_context_length(url, model),
         )
     else:
-        target_url = url
+        target_url = _normalize_openai_chat_url(url)
         if provider == "copilot":
             from src.copilot import apply_request_headers
             apply_request_headers(h, messages_copy)
@@ -1577,6 +1792,7 @@ def llm_call(url: str, model: str, messages: List[Dict], temperature: float = LL
         if max_tokens and max_tokens > 0:
             tok_key = "max_completion_tokens" if _uses_max_completion_tokens(model) else "max_tokens"
             payload[tok_key] = max_tokens
+        _apply_local_generation_stability(payload, target_url, model)
         if provider == "mistral" and _supports_thinking(model):
             payload["reasoning_effort"] = _MISTRAL_REASONING_EFFORT
     try:
@@ -1686,6 +1902,7 @@ async def llm_call_async(
     max_retries: int = LLMConfig.MAX_RETRIES,
     prompt_type: Optional[str] = None,
     session_id: Optional[str] = None,
+    workload: str = "foreground",
 ) -> str:
     """Asynchronous LLM call using httpx with connection pooling, timeout, retry logic, and performance logging."""
     provider = _detect_provider(url)
@@ -1723,6 +1940,7 @@ async def llm_call_async(
             max_tokens=max_tokens,
             headers=headers,
             timeout=timeout,
+            workload=workload,
         ):
             event_is_error = False
             for line in str(chunk).splitlines():
@@ -1767,7 +1985,7 @@ async def llm_call_async(
             stream=False, num_ctx=get_context_length(url, model),
         )
     else:
-        target_url = url
+        target_url = _normalize_openai_chat_url(url)
         h = _provider_headers(provider, headers)
         if provider == "copilot":
             from src.copilot import apply_request_headers
@@ -1788,6 +2006,7 @@ async def llm_call_async(
         if provider == "mistral" and _supports_thinking(model):
             payload["reasoning_effort"] = _MISTRAL_REASONING_EFFORT
         _apply_local_cache_affinity(payload, url, session_id)
+        _apply_local_generation_stability(payload, target_url, model)
 
     if _is_host_dead(target_url):
         raise HTTPException(503, f"Upstream {_host_key(target_url)} marked unreachable (cooldown active)")
@@ -1798,9 +2017,10 @@ async def llm_call_async(
         attempt += 1
         start = time.time()
         try:
-            note_model_activity(target_url, model)
-            client = _get_http_client()
-            r = await httpx_post_kimi_aware_async(client, target_url, h, json=payload, timeout=call_timeout)
+            async with _local_model_slot(target_url, model, workload):
+                note_model_activity(target_url, model)
+                client = _get_http_client()
+                r = await httpx_post_kimi_aware_async(client, target_url, h, json=payload, timeout=call_timeout)
             duration = time.time() - start
             if not r.is_success:
                 friendly = _format_upstream_error(r.status_code, r.text, target_url)
@@ -1842,10 +2062,45 @@ async def llm_call_async(
                 raise HTTPException(502, f"POST {target_url} failed after {max_retries} attempts: {e}")
             await asyncio.sleep(LLMConfig.RETRY_DELAY)
 
+def _stream_target_url(url: str) -> str:
+    provider = _detect_provider(url)
+    if provider == "anthropic":
+        return _normalize_anthropic_url(url)
+    if provider == "ollama":
+        return _normalize_ollama_url(url)
+    if provider == "chatgpt-subscription":
+        return _normalize_chatgpt_subscription_url(url)
+    return _normalize_openai_chat_url(url)
+
+
 async def stream_llm(url: str, model: str, messages: List[Dict], temperature: float = LLMConfig.DEFAULT_TEMPERATURE,
                      max_tokens: int = LLMConfig.DEFAULT_MAX_TOKENS, headers: Optional[Dict] = None,
                      timeout: int = LLMConfig.STREAM_TIMEOUT, prompt_type: Optional[str] = None,
-                     tools: Optional[List[Dict]] = None, session_id: Optional[str] = None):
+                     tools: Optional[List[Dict]] = None, session_id: Optional[str] = None,
+                     tool_choice_none: bool = False, workload: str = "foreground"):
+    target_url = _stream_target_url(url)
+    async with _local_model_slot(target_url, model, workload):
+        async for chunk in _stream_llm_inner(
+            url,
+            model,
+            messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            headers=headers,
+            timeout=timeout,
+            prompt_type=prompt_type,
+            tools=tools,
+            session_id=session_id,
+            tool_choice_none=tool_choice_none,
+        ):
+            yield chunk
+
+
+async def _stream_llm_inner(url: str, model: str, messages: List[Dict], temperature: float = LLMConfig.DEFAULT_TEMPERATURE,
+                            max_tokens: int = LLMConfig.DEFAULT_MAX_TOKENS, headers: Optional[Dict] = None,
+                            timeout: int = LLMConfig.STREAM_TIMEOUT, prompt_type: Optional[str] = None,
+                            tools: Optional[List[Dict]] = None, session_id: Optional[str] = None,
+                            tool_choice_none: bool = False):
     """Stream LLM responses with improved error handling.
 
     Yields SSE chunks:
@@ -1889,7 +2144,7 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
         h = _provider_headers(provider, headers)
         payload = _build_chatgpt_responses_payload(model, messages_copy, temperature, max_tokens, stream=True)
     else:
-        target_url = url
+        target_url = _normalize_openai_chat_url(url)
         payload = {
             "model": model,
             "messages": messages_copy,
@@ -1905,6 +2160,8 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
             payload[tok_key] = max_tokens
         if tools:
             payload["tools"] = tools
+        elif tool_choice_none:
+            payload["tool_choice"] = "none"
         # Mistral thinking-capable models — send reasoning_effort so Mistral
         # activates thinking mode and returns structured reasoning_content.
         # Effort level is configurable via ODYSSEUS_MISTRAL_REASONING_EFFORT
@@ -1917,6 +2174,7 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
         if _is_ollama_openai_compat_url(url) and _supports_thinking(model):
             payload["think"] = False
         _apply_local_cache_affinity(payload, url, session_id)
+        _apply_local_generation_stability(payload, target_url, model)
         h = _provider_headers(provider, headers)
         if provider == "copilot":
             from src.copilot import apply_request_headers
@@ -1933,6 +2191,7 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
         yield f'event: error\ndata: {json.dumps({"error": f"Upstream {_host_key(target_url)} unreachable (cooldown active)", "status": 503})}\n\n'
         return
     note_model_activity(target_url, model)
+    degenerate_guard = _DegenerateStreamGuard(model)
 
     # ── ChatGPT Subscription / Codex Responses streaming ──
     if provider == "chatgpt-subscription":
@@ -1967,6 +2226,10 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                     if evt == "response.output_text.delta":
                         delta = data.get("delta") or ""
                         if delta:
+                            _degenerate = degenerate_guard.check(delta)
+                            if _degenerate:
+                                yield _degenerate
+                                return
                             yield f'data: {json.dumps({"delta": delta})}\n\n'
                     elif evt == "response.completed":
                         usage = (data.get("response") or {}).get("usage") or data.get("usage") or {}
@@ -2299,8 +2562,19 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
                                                 reasoning = (reasoning + thinking_part) if reasoning else thinking_part
                                             content = text_part
                                         if reasoning:
+                                            _degenerate = degenerate_guard.check(reasoning)
+                                            if _degenerate:
+                                                yield _degenerate
+                                                return
                                             yield _stream_delta_event(reasoning, thinking=True)
                                         if content:
+                                            content = _strip_visible_chat_template_artifacts(content)
+                                            if not content:
+                                                continue
+                                            _degenerate = degenerate_guard.check(content)
+                                            if _degenerate:
+                                                yield _degenerate
+                                                return
                                             content = re.sub(r"<mm:think(\s+[^>]*)?>", r"<think\1>", content, flags=re.IGNORECASE)
                                             content = re.sub(r"</mm:think>", "</think>", content, flags=re.IGNORECASE)
                                             stripped = content.lstrip()

src/model_context.py

@@ -316,6 +316,83 @@ def _lookup_known(model: str) -> Optional[int]:
     return best_ctx
 
 
+def _model_ctx_from_entry(m: dict) -> Optional[int]:
+    """Extract a positive context window from one /models catalog entry.
+
+    Checks the common top-level fields first, then a nested meta/model_extra
+    object. Returns None when no positive window is reported.
+    """
+    if not isinstance(m, dict):
+        return None
+    for field in (
+        "context_length",
+        "context_window",
+        "max_model_len",
+        "max_context_length",
+        "max_seq_len",
+    ):
+        val = m.get(field)
+        if val and isinstance(val, (int, float)) and val > 0:
+            return int(val)
+    meta = m.get("meta") or m.get("model_extra") or {}
+    if isinstance(meta, dict):
+        # n_ctx is the actual serving context (set via -c flag in llama.cpp)
+        for field in ("n_ctx", "context_length", "context_window", "max_model_len"):
+            val = meta.get(field)
+            if val and isinstance(val, (int, float)) and val > 0:
+                return int(val)
+    return None
+
+
+# Per-endpoint cache of the {model_id: context_length} map parsed from a
+# proxy/api catalog. api/proxy endpoints skip the /models download on every
+# lookup because a large catalog is expensive; caching the whole map lets us
+# pay that download at most once per endpoint instead of once per model.
+_catalog_ctx_cache: Dict[str, Dict[str, int]] = {}
+
+
+def _proxy_catalog_context(endpoint_url: str, model: str) -> Optional[int]:
+    """Context window for a model read from the endpoint's /models catalog.
+
+    Fetches the catalog once per endpoint and caches the full id->context map,
+    so an api/proxy endpoint serving a model that isn't in KNOWN_CONTEXT_WINDOWS
+    (e.g. a new OpenRouter model) still reports its real window instead of the
+    bare default. Returns None when the catalog can't be read or doesn't list a
+    positive window for the model.
+    """
+    cat = _catalog_ctx_cache.get(endpoint_url)
+    if cat is None:
+        from src.endpoint_resolver import build_models_url
+        try:
+            r = httpx.get(build_models_url(endpoint_url), timeout=REQUEST_TIMEOUT)
+        except Exception as e:
+            logger.debug(f"Failed to fetch proxy catalog for context length: {e}")
+            return None
+        if not r.is_success:
+            return None
+        cat = {}
+        try:
+            for m in (r.json().get("data") or []):
+                mid = m.get("id") if isinstance(m, dict) else None
+                ctx = _model_ctx_from_entry(m) if mid else None
+                if mid and ctx:
+                    cat[mid] = ctx
+        except Exception as e:
+            logger.debug(f"Failed to parse proxy catalog for context length: {e}")
+            return None
+        _catalog_ctx_cache[endpoint_url] = cat
+
+    if model in cat:
+        return cat[model]
+    # Catalog ids may carry a provider prefix (e.g. "openai/gpt-4o") while the
+    # session stores the bare id; match on the trailing segment as a fallback.
+    base = model.split("/")[-1]
+    for mid, ctx in cat.items():
+        if mid.split("/")[-1] == base:
+            return ctx
+    return None
+
+
 def _query_context_length(endpoint_url: str, model: str) -> Tuple[int, bool]:
     """Query the model API for context length. Returns (context_length, known) where
     ``known`` is False only for the bare DEFAULT_CONTEXT fallback."""
@@ -330,6 +407,14 @@ def _query_context_length(endpoint_url: str, model: str) -> Tuple[int, bool]:
         if known:
             logger.info(f"Using known context window for {model}: {known}")
             return known, True
+        # Not in the known table: read the real window from the catalog (cached
+        # once per endpoint) instead of capping every unknown model at the
+        # default — that under-reported large windows on aggregators like
+        # OpenRouter (issue #4886).
+        api_ctx = _proxy_catalog_context(endpoint_url, model)
+        if api_ctx:
+            logger.info(f"Proxy catalog reports context window for {model}: {api_ctx}")
+            return api_ctx, True
         return DEFAULT_CONTEXT, False
 
     # Try llama.cpp /slots endpoint first — reports actual serving context
@@ -370,27 +455,7 @@ def _query_context_length(endpoint_url: str, model: str) -> Tuple[int, bool]:
             for m in models_list:
                 mid = m.get("id", "")
                 if mid == model or mid.split("/")[-1] == model.split("/")[-1]:
-                    for field in (
-                        "context_length",
-                        "context_window",
-                        "max_model_len",
-                        "max_context_length",
-                        "max_seq_len",
-                    ):
-                        val = m.get(field)
-                        if val and isinstance(val, (int, float)) and val > 0:
-                            api_ctx = int(val)
-                            break
-
-                    if not api_ctx:
-                        meta = m.get("meta") or m.get("model_extra") or {}
-                        if isinstance(meta, dict):
-                            # n_ctx is the actual serving context (set via -c flag in llama.cpp)
-                            for field in ("n_ctx", "context_length", "context_window", "max_model_len"):
-                                val = meta.get(field)
-                                if val and isinstance(val, (int, float)) and val > 0:
-                                    api_ctx = int(val)
-                                    break
+                    api_ctx = _model_ctx_from_entry(m)
                     break
     except Exception as e:
         logger.debug(f"Failed to query context length for {model}: {e}")

src/personal_docs.py

@@ -68,6 +68,8 @@ def read_text_file(path: str) -> str:
 
 def split_chunks(text: str, size: int = config.CHUNK_SIZE, overlap: int = config.CHUNK_OVERLAP) -> List[str]:
     """Split text into overlapping chunks."""
+    if not isinstance(text, str):
+        return []
     text = text.strip()
     if not text:
         return []
@@ -87,7 +89,8 @@ def split_chunks(text: str, size: int = config.CHUNK_SIZE, overlap: int = config
 
 def tokenize(s: str) -> Set[str]:
     """Tokenize string into words, excluding stop words."""
-    tokens = re.findall(r"[A-Za-z0-9_\-]+", (s or "").lower())
+    text = s if isinstance(s, str) else ""
+    tokens = re.findall(r"[A-Za-z0-9_\-]+", text.lower())
     return set(t for t in tokens if t not in config.STOP_WORDS and len(t) > 1)
 
 def load_personal_index(

src/rag_manager.py

@@ -32,9 +32,9 @@ class RAGManager:
         logger.info("RAGManager initialized as wrapper for VectorRAG")
     
     # Delegate all methods to VectorRAG
-    def search(self, query: str, k: int = 5) -> List[Dict[str, Any]]:
+    def search(self, query: str, k: int = 5, owner: Optional[str] = None) -> List[Dict[str, Any]]:
         """Search for documents - delegates to VectorRAG."""
-        return self.vector_rag.search(query, k)
+        return self.vector_rag.search(query, k, owner=owner)
     
     def index_personal_documents(
         self,

src/session_search.py

@@ -207,6 +207,7 @@ def _search_like(
     )
     if not include_archived:
         q = q.filter(DBSession.archived == False)
+    q = q.filter(~DBSession.name.like("SFT trace batch%"))
     if restrict_owner:
         q = _owner_filter(q, owner, include_legacy_owner)
     rows = q.order_by(DBChatMessage.timestamp.desc()).limit(limit).all()
@@ -270,6 +271,7 @@ def _search_fts(
         WHERE chat_messages_fts MATCH :fts_query
           {archived_clause}
           {owner_clause}
+          AND s.name NOT LIKE 'SFT trace batch%'
           AND m.role IN ('user', 'assistant')
         ORDER BY bm25(chat_messages_fts), m.timestamp DESC
         LIMIT :limit

src/settings.py

@@ -136,6 +136,10 @@ DEFAULT_SETTINGS = {
     "task_model": "",
     "default_endpoint_id": "",
     "default_model": "",
+    # Optional prose style used only for normal document writing/editing.
+    # Email replies use email_writing_style instead because greetings,
+    # signatures, and mailbox identity rules are medium-specific.
+    "document_writing_style": "",
     # Ordered fallback chain for the default chat model. Each entry is
     # {"endpoint_id": "...", "model": "..."}. If the primary model fails
     # before producing output (endpoint offline / errors), the chat

src/task_action_policy.py

@@ -0,0 +1,47 @@
+"""Shared privilege policy for scheduled task actions."""
+
+from __future__ import annotations
+
+ADMIN_ONLY_TASK_ACTIONS = frozenset({
+    "run_local",
+    "run_script",
+    "ssh_command",
+    "cookbook_serve",
+})
+
+
+def is_admin_only_task_action(task_type: str | None, action: str | None) -> bool:
+    return (task_type or "llm") == "action" and (action or "") in ADMIN_ONLY_TASK_ACTIONS
+
+
+def owner_has_admin_task_privileges(owner: str | None) -> bool:
+    try:
+        from src.auth_helpers import _auth_disabled
+        if _auth_disabled():
+            return True
+    except Exception:
+        pass
+
+    if owner:
+        try:
+            from core.middleware import INTERNAL_TOOL_USER
+            if owner == INTERNAL_TOOL_USER:
+                return True
+        except Exception:
+            pass
+
+    try:
+        from core.auth import AuthManager
+        auth = AuthManager()
+        if not auth.is_configured:
+            return True
+        if not owner:
+            return False
+        return bool(auth.is_admin(owner))
+    except Exception:
+        pass
+
+    if not owner:
+        return False
+
+    return False

src/task_endpoint.py

@@ -6,6 +6,7 @@ from src.endpoint_resolver import (
     resolve_utility_fallback_candidates,
 )
 from src.llm_core import llm_call_async_with_fallback
+from src.interactive_gate import wait_for_interactive_quiet
 
 
 def resolve_task_endpoint(fallback_url=None, fallback_model=None, fallback_headers=None, owner=None):
@@ -72,4 +73,6 @@ async def task_llm_call_async(
     )
     if not candidates:
         raise RuntimeError("No LLM endpoint available for background task")
+    await wait_for_interactive_quiet("background task LLM")
+    kwargs.setdefault("workload", "background")
     return await llm_call_async_with_fallback(candidates, messages=messages, **kwargs)

src/task_scheduler.py

@@ -10,6 +10,10 @@ from datetime import datetime, timedelta, timezone
 from typing import Any, Awaitable, Callable, Dict, Tuple
 
 from core.auth import RESERVED_USERNAMES
+from src.task_action_policy import (
+    is_admin_only_task_action,
+    owner_has_admin_task_privileges,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -239,6 +243,7 @@ HOUSEKEEPING_DEFAULTS = {
     "tidy_research":        {"name": "Research Tidy",            "trigger_type": "event", "trigger_event": "research_completed", "trigger_count": 5, "schedule": None, "scheduled_time": None, "cron_expression": None, "legacy_names": ["Tidy Research"]},
     "summarize_emails":     {"name": "Email (Summary)",          "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 */2 * * *", "ship_paused": True, "legacy_names": ["Tidy Email (Summary)"]},
     "draft_email_replies":  {"name": "Email AI Auto Reply",      "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 */2 * * *", "ship_paused": True, "legacy_names": ["Tidy Email (Replies)", "AI Auto Reply"]},
+    "email_auto_translate": {"name": "Email Auto Translate",     "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 */2 * * *", "ship_paused": True, "legacy_names": ["Auto-translate Emails", "Auto Translate Email"]},
     "extract_email_events": {"name": "Email Calendar Events",    "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 */1 * * *", "ship_paused": True, "legacy_names": ["Email → Calendar Events"]},
     "classify_events":      {"name": "Calendar Classify Events", "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 6,18 * * *", "ship_paused": True, "legacy_names": ["Classify Calendar Events"]},
     "check_email_urgency":   {"name": "Email Tags",               "schedule": "cron",  "scheduled_time": None,    "cron_expression": "0 * * * *", "ship_paused": True, "old_cron_expressions": ["*/15 * * * *"], "legacy_names": ["Email Triage", "Urgent Email"]},
@@ -686,6 +691,12 @@ class TaskScheduler:
         db = SessionLocal()
         try:
             now = _utcnow()
+            foreground_active = False
+            try:
+                from src.interactive_gate import has_foreground_activity
+                foreground_active = has_foreground_activity()
+            except Exception:
+                foreground_active = False
             async with self._executing_lock:
                 # Snapshot under the lock so we don't race with mid-iteration adds.
                 executing_snapshot = set(self._executing)
@@ -699,8 +710,13 @@ class TaskScheduler:
                 for task in due:
                     if task.id in self._executing:
                         continue
+                    if foreground_active:
+                        task.next_run = now + timedelta(minutes=15)
+                        continue
                     self._executing.add(task.id)
                     to_dispatch.append(task.id)
+                if foreground_active and due:
+                    db.commit()
             for task_id in to_dispatch:
                 asyncio.create_task(self._execute_task(task_id))
         finally:
@@ -734,15 +750,26 @@ class TaskScheduler:
 
         try:
             if bypass_model_slot or not self._task_needs_model_slot(task_id):
-                await self._execute_task_locked(task_id, run_id, release_executing=release_executing)
+                await self._execute_task_locked(
+                    task_id,
+                    run_id,
+                    release_executing=release_executing,
+                    gate_foreground=not bypass_model_slot,
+                )
                 return
 
             async with self._run_semaphore:
-                await self._execute_task_locked(task_id, run_id, release_executing=release_executing)
+                await self._execute_task_locked(
+                    task_id,
+                    run_id,
+                    release_executing=release_executing,
+                    gate_foreground=True,
+                )
         except asyncio.CancelledError:
             # If cancellation happens while queued behind the semaphore,
             # _execute_task_locked never runs and cannot update the Activity row.
             self._mark_run_aborted(task_id, run_id)
+            self._defer_immediately_due_task(task_id, delay=timedelta(minutes=15))
             raise
         finally:
             handle = self._task_handles.get(task_id)
@@ -752,7 +779,36 @@ class TaskScheduler:
                 async with self._executing_lock:
                     self._executing.discard(task_id)
 
-    async def _execute_task_locked(self, task_id: str, run_id: str, *, release_executing: bool = True):
+    def _defer_immediately_due_task(self, task_id: str, *, delay: timedelta):
+        """A queued task can be cancelled before _execute_task_locked gets a DB
+        handle. If its next_run stays in the past, the scheduler dispatches it
+        again on the next tick and spams aborted Activity rows."""
+        try:
+            from core.database import SessionLocal, ScheduledTask
+            db = SessionLocal()
+            try:
+                task = db.query(ScheduledTask).filter(ScheduledTask.id == task_id).first()
+                if (
+                    task
+                    and task.status == "active"
+                    and task.next_run is not None
+                    and task.next_run <= _utcnow()
+                ):
+                    task.next_run = _utcnow() + delay
+                    db.commit()
+            finally:
+                db.close()
+        except Exception:
+            logger.debug("Failed to defer cancelled queued task %s", task_id, exc_info=True)
+
+    async def _execute_task_locked(
+        self,
+        task_id: str,
+        run_id: str,
+        *,
+        release_executing: bool = True,
+        gate_foreground: bool = True,
+    ):
         from core.database import SessionLocal, ScheduledTask, TaskRun
 
         db = SessionLocal()
@@ -769,6 +825,36 @@ class TaskScheduler:
                     db.commit()
                 return
 
+            if (
+                is_admin_only_task_action(task.task_type, task.action)
+                and not owner_has_admin_task_privileges(task.owner)
+            ):
+                msg = f"Action '{task.action}' requires admin privileges"
+                blocked = db.query(TaskRun).filter(TaskRun.id == run_id).first()
+                if blocked:
+                    blocked.status = "error"
+                    blocked.result = msg
+                    blocked.error = msg
+                    blocked.finished_at = _utcnow()
+                task.status = "paused"
+                task.next_run = None
+                task.last_run = _utcnow()
+                logger.warning(
+                    "Paused admin-only task %s for non-admin owner %r",
+                    task_id,
+                    task.owner,
+                )
+                db.commit()
+                return
+
+            if gate_foreground:
+                waiting = db.query(TaskRun).filter(TaskRun.id == run_id).first()
+                if waiting and waiting.status == "queued":
+                    waiting.result = "Queued — waiting for Odysseus to be idle…"
+                    db.commit()
+                from src.interactive_gate import wait_for_interactive_quiet
+                await wait_for_interactive_quiet(f"scheduled task {task.name}")
+
             # Flip the run from queued → running. Reset started_at to the
             # actual execution start so queue wait time is visible from
             # created_at vs started_at if we ever surface that.
@@ -799,6 +885,27 @@ class TaskScheduler:
             # previous llm/research run's model. The executors set it once the
             # model is resolved.
             self._last_run_model = None
+            foreground_cancel = {"hit": False}
+            foreground_monitor = None
+            if gate_foreground:
+                current_task = asyncio.current_task()
+
+                async def _cancel_if_foreground_active():
+                    # Give the just-finished quiet gate a tiny grace window,
+                    # then keep enforcing "background means background" while
+                    # a long email/LLM action is already running.
+                    await asyncio.sleep(0.1)
+                    from src.interactive_gate import has_foreground_activity
+                    while True:
+                        await asyncio.sleep(0.25)
+                        if has_foreground_activity():
+                            foreground_cancel["hit"] = True
+                            logger.info("Task '%s' interrupted because Odysseus became active", task.name)
+                            if current_task:
+                                current_task.cancel()
+                            return
+
+                foreground_monitor = asyncio.create_task(_cancel_if_foreground_active())
             try:
                 if task_type == "action":
                     result, success = await self._execute_action(task, run_id=run_id)
@@ -838,15 +945,22 @@ class TaskScheduler:
                 db.commit()
                 return
             except asyncio.CancelledError:
-                logger.info("Task '%s' stopped by user", task.name)
+                msg = (
+                    "Paused because Odysseus became active"
+                    if foreground_cancel.get("hit")
+                    else "Stopped by user"
+                )
+                logger.info("Task '%s' %s", task.name, msg)
                 run_obj = db.query(TaskRun).filter(TaskRun.id == run_id).first()
                 if run_obj:
                     run_obj.status = "aborted"
-                    run_obj.error = "Stopped by user"
-                    run_obj.result = run_obj.result or "Stopped by user"
+                    run_obj.error = msg
+                    run_obj.result = run_obj.result or msg
                     run_obj.finished_at = _utcnow()
                 task.last_run = _utcnow()
-                if (task.trigger_type or "schedule") == "schedule":
+                if foreground_cancel.get("hit"):
+                    task.next_run = _utcnow() + timedelta(minutes=15)
+                elif (task.trigger_type or "schedule") == "schedule":
                     task.next_run = compute_next_run(
                         task.schedule, task.scheduled_time,
                         task.scheduled_day, task.scheduled_date,
@@ -881,6 +995,13 @@ class TaskScheduler:
                     task.next_run = None
                 db.commit()
                 return
+            finally:
+                if foreground_monitor and not foreground_monitor.done():
+                    foreground_monitor.cancel()
+                    try:
+                        await foreground_monitor
+                    except asyncio.CancelledError:
+                        pass
 
             run.finished_at = _utcnow()
 
@@ -1050,6 +1171,7 @@ class TaskScheduler:
         "learn_sender_signatures",
         "summarize_emails",
         "draft_email_replies",
+        "email_auto_translate",
         "extract_email_events",
         "classify_events",
         "tidy_sessions",
@@ -1063,6 +1185,7 @@ class TaskScheduler:
     _MODEL_BACKED_ACTIONS = frozenset({
         "summarize_emails",
         "draft_email_replies",
+        "email_auto_translate",
         "extract_email_events",
         "classify_events",
         "learn_sender_signatures",
@@ -1119,6 +1242,8 @@ class TaskScheduler:
                 self._set_run_progress(run_id, message)
 
             kwargs = {"owner": task.owner, "task_name": task.name, "progress_cb": _progress}
+            if task.prompt:
+                kwargs["prompt"] = task.prompt
             if task.action in ("run_script", "run_local", "ssh_command") and task.prompt:
                 kwargs["script" if task.action in ("run_script", "run_local") else "command"] = task.prompt
             # cookbook_serve carries its JSON config in task.prompt — feed it
@@ -1682,8 +1807,15 @@ class TaskScheduler:
 
         target = (output or "").strip()
         explicit = ""
+        account_id = ""
         if target.startswith("email:"):
             explicit = target.split(":", 1)[1].strip()
+            if "|account=" in explicit:
+                explicit, account_id = explicit.split("|account=", 1)
+                explicit = explicit.strip()
+                account_id = account_id.strip()
+            if explicit == "self":
+                explicit = ""
         elif "@" in target:
             explicit = target
 
@@ -1691,7 +1823,7 @@ class TaskScheduler:
             from routes.email_routes import _resolve_send_config
             from routes.email_helpers import _send_smtp_message
 
-            cfg = _resolve_send_config(owner=task.owner or "")
+            cfg = _resolve_send_config(account_id=account_id or None, owner=task.owner or "")
             to_addr = explicit or cfg.get("from_address") or cfg.get("smtp_user") or ""
             if not to_addr:
                 raise RuntimeError("No email recipient resolved for task output")
@@ -1759,6 +1891,8 @@ class TaskScheduler:
         # behind the primary endpoint so a downed primary won't silently yield
         # `(no output)`.
         try:
+            from src.interactive_gate import wait_for_interactive_quiet
+            await wait_for_interactive_quiet(f"agent task {task.name}")
             from src.task_endpoint import resolve_task_candidates
             _task_fallbacks = resolve_task_candidates(
                 fallback_url=endpoint_url,
@@ -1779,6 +1913,7 @@ class TaskScheduler:
             disabled_tools=disabled_tools,
             relevant_tools=relevant_tools,
             fallbacks=_task_fallbacks,
+            workload="background",
         ):
             if event_str.startswith("data: ") and not event_str.startswith("data: [DONE]"):
                 try:
@@ -2105,6 +2240,28 @@ class TaskScheduler:
         stopped = self._mark_run_aborted(task_id) or stopped
         return stopped
 
+    async def stop_background_tasks_for_foreground(self, *, reason: str = "Odysseus became active") -> int:
+        """Cancel all in-process scheduler tasks because the user is active.
+
+        This is intentionally blunt for scheduled/background work: when the
+        user opens or uses Odysseus, foreground interaction wins immediately.
+        Manual force-runs can be restarted by the user; automatic jobs will be
+        deferred by their cancellation path instead of stealing the app.
+        """
+        async with self._executing_lock:
+            task_ids = list(self._executing)
+        stopped = 0
+        for task_id in task_ids:
+            handle = self._task_handles.get(task_id)
+            if handle and not handle.done():
+                handle.cancel()
+                stopped += 1
+            if self._mark_run_aborted(task_id):
+                stopped += 1
+        if stopped:
+            logger.info("Stopped %d background scheduler task(s): %s", stopped, reason)
+        return stopped
+
     async def ensure_defaults(self, owner: str):
         """Create default housekeeping tasks for this owner (idempotent per action)."""
         from core.database import SessionLocal, ScheduledTask

src/tool_execution.py

@@ -21,7 +21,12 @@ from typing import Any, Awaitable, Callable, Dict, Optional, Tuple
 
 
 
-from src.tool_security import is_public_blocked_tool, owner_is_admin_or_single_user
+from src.tool_security import (
+    BUILTIN_EMAIL_TOOLS,
+    email_tool_policy_names,
+    is_public_blocked_tool,
+    owner_is_admin_or_single_user,
+)
 from src.tool_policy import ToolPolicy
 from src.constants import MAX_OUTPUT_CHARS, MAX_READ_CHARS, MAX_DIFF_LINES, DATA_DIR
 from src.tool_utils import _truncate, get_mcp_manager
@@ -66,25 +71,35 @@ _SENSITIVE_FILE_PATTERNS: tuple[str, ...] = (
     "known_hosts",
 )
 
+# Case-folded views used for matching. On a case-insensitive filesystem
+# (Windows, default macOS) ".SSH/AUTHORIZED_KEYS" and ".env" resolve to the
+# same protected files as their lowercase forms, so the deny-list has to fold
+# case before comparing — the sibling resolver already normcases paths for the
+# same reason. casefold (not os.path.normcase) because normcase is a no-op on
+# POSIX, which is exactly where the macOS read-exfil path lives.
+_SENSITIVE_BASENAMES_CF: frozenset[str] = frozenset(b.casefold() for b in _SENSITIVE_BASENAMES)
+_SENSITIVE_FILE_PATTERNS_CF: frozenset[str] = frozenset(p.casefold() for p in _SENSITIVE_FILE_PATTERNS)
+
 
 def _is_sensitive_path(resolved: str) -> bool:
     """Return True if *resolved* falls under a sensitive directory or
     matches a sensitive filename — regardless of what root it sits under.
+
+    Matching is case-insensitive: on Windows / default macOS a case-variant
+    name (``.SSH``, ``AUTHORIZED_KEYS``, ``Id_Rsa``) points at the same file as
+    the lowercase form, so a case-sensitive check would let it slip past the
+    deny-list in every file tool that relies on it.
     """
-    parts = resolved.split(os.sep)
-    filenames: set[str] = {parts[-1]} if parts else set()
+    parts = [p.casefold() for p in resolved.split(os.sep)]
+    filename = parts[-1] if parts else ""
 
     # Check if any path component is a sensitive directory.
     for part in parts:
-        if part in _SENSITIVE_BASENAMES:
+        if part in _SENSITIVE_BASENAMES_CF:
             return True
 
     # Check filename against known sensitive files.
-    for pat in _SENSITIVE_FILE_PATTERNS:
-        if pat in filenames:
-            return True
-
-    return False
+    return filename in _SENSITIVE_FILE_PATTERNS_CF
 
 
 def _tool_path_roots() -> list[str]:
@@ -390,8 +405,42 @@ _MCP_ARG_PARSERS: Dict[str, Callable[[str], Dict[str, str]]] = {
 }
 
 
+# Primary argument key(s) for the legacy line-parsed tools. When a fenced
+# block's content is a JSON object carrying one of these keys, it's structured
+# inline args (the relaxed parser's ```web_search {"query": "..."}``` shape) —
+# use the object directly instead of letting the line-based parsers wrap the
+# whole JSON string as the query/url/path/prompt. Keyed off membership only
+# (the primary key never changes), so this can't drift; an unrecognized object
+# safely falls through to the line-based parser, i.e. the previous behavior.
+#
+# IMPORTANT — this only covers the MCP path. _build_mcp_args is reached via
+# _call_mcp_tool only for _MCP_TOOL_MAP tools (so an entry outside that map is
+# dead, as manage_memory was). And of these, only generate_image has a live MCP
+# server today; web_search/web_fetch/read_file/write_file have none, so they run
+# via _direct_fallback -> TOOL_HANDLERS, whose handlers decode JSON themselves
+# (see ReadFileTool/WriteFileTool/WebSearchTool/WebFetchTool). The entries here
+# are kept as defense-in-depth for if/when those servers are added. The live
+# fix for each server-less tool lives in its handler. test_write_file_inline_
+# json_args and test_mcp_json_primary_keys_are_all_live pin both halves.
+_MCP_JSON_PRIMARY_KEYS: Dict[str, tuple] = {
+    "web_search":     ("query", "queries"),
+    "web_fetch":      ("url",),
+    "read_file":      ("path",),
+    "write_file":     ("path",),
+    "generate_image": ("prompt",),
+}
+
+
 def _build_mcp_args(tool: str, content: str) -> Dict:
     """Convert fenced-block text content to structured MCP arguments."""
+    primaries = _MCP_JSON_PRIMARY_KEYS.get(tool)
+    if primaries and content.strip().startswith("{"):
+        try:
+            decoded = json.loads(content.strip())
+        except (json.JSONDecodeError, TypeError):
+            decoded = None
+        if isinstance(decoded, dict) and any(k in decoded for k in primaries):
+            return decoded
     parser = _MCP_ARG_PARSERS.get(tool)
     return parser(content) if parser else {}
 
@@ -596,6 +645,12 @@ async def _execute_tool_block_impl(
     tool = block.tool_type
     content = block.content
 
+    # The block/disable gates below must match every policy-equivalent
+    # spelling of the tool name (bare email names alias their mcp__email__
+    # form — see email_tool_policy_names), not just the spelling the model
+    # happened to emit.
+    policy_names = email_tool_policy_names(tool)
+
     # Misformatted tool call detection: model put JSON inside ```python``` (or
     # similar) without naming the tool. Common with MiniMax-style outputs.
     # Return a helpful error so the model retries with the correct format.
@@ -623,13 +678,13 @@ async def _execute_tool_block_impl(
             pass
 
     # Reject tools that the user has disabled for this request
-    if disabled_tools and tool in disabled_tools:
+    if disabled_tools and not policy_names.isdisjoint(disabled_tools):
         desc = f"{tool}: BLOCKED"
         result = {"error": f"Tool '{tool}' is disabled by user.", "exit_code": 1}
         logger.info(f"Tool blocked by user: {tool}")
         return desc, result
 
-    if tool_policy and tool_policy.blocks(tool):
+    if tool_policy and any(tool_policy.blocks(name) for name in policy_names):
         desc = f"{tool}: BLOCKED"
         result = {
             "error": f"Execution of tool '{tool}' is forbade by the active guide-only policy.",
@@ -823,6 +878,51 @@ async def _execute_tool_block_impl(
     elif tool == "vault_unlock":
         desc = "vault_unlock"
         result = await do_vault_unlock(content, owner=owner)
+    elif tool in BUILTIN_EMAIL_TOOLS:
+        # Bare email tool name from fenced-block models (e.g. Ollama) — route to MCP email server.
+        # Non-admin owners never reach here: BUILTIN_EMAIL_TOOLS ⊆ NON_ADMIN_BLOCKED_TOOLS,
+        # so is_public_blocked_tool() above already rejected them.
+        mcp = get_mcp_manager()
+        qualified = f"mcp__email__{tool}"
+        desc = f"email: {tool}"
+        if mcp:
+            _raw = content.strip()
+            args = {}
+            _args_error = None
+            if _raw:
+                # A non-empty body is always meant to be the call's arguments,
+                # and every email tool takes a JSON object. Anything that
+                # isn't one is a correctable error — NOT a silent empty-args
+                # call, which would read the DEFAULT mailbox/folder instead of
+                # the one the model meant (#3966 class). Only an EMPTY body
+                # keeps the no-arg path (e.g. ```list_email_accounts```).
+                try:
+                    parsed = json.loads(_raw)
+                except (json.JSONDecodeError, TypeError) as _je:
+                    # Covers both `{account: "work"}` (looks like JSON, bad)
+                    # and `account: work` (not JSON at all).
+                    _args_error = (
+                        f"'{tool}' arguments are not valid JSON ({_je}). "
+                        'Send a JSON object, e.g. {"account": "work"} — '
+                        "keys and string values need double quotes."
+                    )
+                else:
+                    if isinstance(parsed, dict):
+                        args = parsed
+                    else:
+                        _args_error = (
+                            f"'{tool}' arguments must be a JSON object, "
+                            'e.g. {"uid": "..."} — got a JSON array/value instead.'
+                        )
+            if _args_error is not None:
+                result = {"error": _args_error, "exit_code": 1}
+            else:
+                if owner:
+                    args = dict(args)
+                    args[_EMAIL_MCP_OWNER_ARG] = owner
+                result = await mcp.call_tool(qualified, args)
+        else:
+            result = {"error": "MCP manager not available", "exit_code": 1}
     elif tool.startswith("mcp__"):
         # MCP tool dispatch
         mcp = get_mcp_manager()
@@ -840,12 +940,12 @@ async def _execute_tool_block_impl(
             desc = f"mcp: {tool}"
             result = {"error": "MCP manager not available", "exit_code": 1}
 
-    
+
     elif tool in dynamic_handlers:
         first_line = content.split(chr(10))[0][:80]
         desc = f"registry: {tool} {first_line}".strip()
         res = await _direct_fallback(tool, content, progress_cb=progress_cb)
-        
+
         if isinstance(res, tuple):
             desc, result = res
         else:

src/tool_implementations.py

@@ -36,6 +36,8 @@ def __getattr__(name):
         from src.agent_tools import admin_tools
         return getattr(admin_tools, name)
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
 # Cookbook (model serving) domain extracted to src/tools/cookbook.py
 # (slice 1, #4082/#4071). Re-imported here so this module stays a working
 # facade. cookbook.py pulls `_internal_headers` / `_INTERNAL_BASE` back

src/tool_index.py

@@ -105,12 +105,12 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "search_chats": "Search past session transcripts across chats.",
     "ask_user": "Ask the user a multiple-choice question to get a decision or clarification. Use this when the task is genuinely ambiguous and the answer changes what you do next — pick between approaches, confirm an assumption, choose among options — instead of guessing. Provide a clear `question` and 2-6 `options` (each with a short `label`, optional `description`). Omit `multi`/keep it false unless the question explicitly permits choosing multiple options. Calling this ENDS your turn: the user sees clickable buttons and their choice arrives as your next message. Don't use it for things you can decide from context or sensible defaults, or for irreversible-action confirmation if a dedicated flow exists.",
     "update_plan": "Write back to the ACTIVE PLAN while executing an approved plan: mark steps done or revise them. After finishing a step call this with the full checklist and that step marked done; when the user asks to change the plan call it with the revised checklist. Always pass the COMPLETE markdown checklist (`- [ ]` / `- [x]`), not a diff. The user's docked plan window updates live. No effect when there is no active plan.",
-    "ui_control": "Control the UI and toggle tools on/off. Use this to turn off / turn on / disable / enable individual tools and features: shell (bash), search (web), research, browser, documents, incognito. Open panels (documents library, gallery, email inbox, sessions, notes, memories/brain, skills, settings, cookbook) via `open_panel <name>`. Use `open_email_reply <uid> <folder> reply` to open an email reply draft document without sending. To pre-fill the reply body in one shot (USE THIS whenever the user told you what to say — opening an empty draft when they asked you to write is wrong), append the body after the mode: `open_email_reply <uid> <folder> reply <body text>`. Body can continue on subsequent lines for multi-line replies. Also switches between chat/agent modes, changes the current model, and applies/creates themes.",
+    "ui_control": "Control the UI and toggle tools on/off. Use this to turn off / turn on / disable / enable individual tools and features: shell (bash), search (web), research, browser, documents, incognito. Open panels (documents library, gallery, email inbox, sessions, notes, memories/brain, skills, settings, cookbook) via `open_panel <name>`. Use `open_email_reply <uid> <folder> reply <body text>` (or structured body) to open an email reply draft document without sending. USE THIS whenever the user says to write/draft a reply or tells you what to say — opening an empty draft or sending immediately is wrong. Body can continue on subsequent lines for multi-line replies. Also switches between chat/agent modes, changes the current model, and applies/creates themes.",
     "list_email_accounts": "List configured email accounts and default status. Use before reading or sending mail when the user mentions Gmail, work mail, custom domain mail, another mailbox, or asks to compare/check multiple inboxes.",
     "list_emails": "List emails for a folder/account, newest first, including read messages by default. Shows subject, sender, date, UID, account, and AI summary. Check inbox, find emails needing replies. Supports account from list_email_accounts for Gmail/work/custom mailboxes. For last/latest/newest email, use max_results=1 and unread_only=false.",
     "read_email": "Read the full content of a specific email by UID or Message-ID. View email body, check details. Supports account from list_email_accounts when the UID belongs to a non-default mailbox.",
     "send_email": "Send a new email via SMTP. Provide recipient, subject, body, and optional account from list_email_accounts. For replying to a thread use reply_to_email instead.",
-    "reply_to_email": "SEND a reply email immediately by UID. Do not use for open/start reply draft requests; use ui_control open_email_reply for those. For follow-up 'reply ...' send requests, use the exact UID and account from latest read_email/list_emails output; never invent UID 1. Threads automatically with In-Reply-To/References, prefixes Re:, marks original as Answered.",
+    "reply_to_email": "SEND a reply email immediately by UID. Do not use for write/draft/open/start reply requests; use ui_control open_email_reply with body so the user can review. Only use when the user explicitly says to send now. For send requests, use the exact UID and account from latest read_email/list_emails output; never invent UID 1. Threads automatically with In-Reply-To/References, prefixes Re:, marks original as Answered.",
     "archive_email": "Move an email out of the inbox into the Archive folder. Use after handling messages you want to keep but get out of the way.",
     "delete_email": "Delete an email — moves to Trash by default, or expunges permanently with permanent=true.",
     "mark_email_read": "Mark an email as read or unread by toggling the \\Seen flag.",
@@ -118,7 +118,7 @@ BUILTIN_TOOL_DESCRIPTIONS: Dict[str, str] = {
     "resolve_contact": "Look up a contact's email address by name. Searches CardDAV address book and sent email history. Use when the user says 'message [name]', 'email [name]', or 'send to [name]' without an email address.",
     "manage_contact": "Save / update / delete / list address-book contacts (CardDAV). Use for info about ANOTHER person — name, email, phone, postal address. Args: action=list|add|update|delete, name, email, phones, address, uid (from list). For 'save this for <person>' / address pastes / phone numbers next to a name, this is the right tool — NOT manage_memory. Do NOT use for facts about the USER ('my name is X'); those are manage_memory.",
     "manage_notes": "Create and manage notes and checklists (Google Keep-style). ALWAYS use this for note/todo/checklist/reminder creation — NEVER hit /api/notes via app_api. Accepts natural-language `due_date` like 'tomorrow at 9am' or '11pm today' (parsed in the USER'S timezone). The due_date IS the reminder — it fires a notification at that time, so do NOT also create a calendar event for the same reminder. Set colors, labels, pin, archive. Do NOT use manage_memory for note content.",
-    "manage_calendar": "Calendar event management: list, create, update, delete. Each event can carry a tag/category (event_type — work/personal/health/travel/meal/social/admin/other) and importance (low/normal/high/critical). Resolve today/tomorrow using the Current date and time context, then use ISO datetimes in the user's local wall time; supports all-day events. For event reminders/alarms, pass reminder_minutes; this creates the Notes reminder, so do not also call manage_notes for the same reminder.",
+    "manage_calendar": "Calendar event management: list, create, update, delete. Each event can carry a tag/category (event_type — work/personal/health/travel/meal/social/admin/other) and importance (low/normal/high/critical). Resolve today/tomorrow using the Current date and time context, then use ISO datetimes in the user's local wall time; supports all-day events. Use rrule only for explicit recurrence; for update_event pass rrule='' to remove repeats. For event reminders/alarms, pass reminder_minutes; this creates the Notes reminder, so do not also call manage_notes for the same reminder.",
     "download_model": "Download a HuggingFace model to a local or remote server. Specify repo_id (e.g. 'Qwen/Qwen3-8B'), optional server host, and optional include filter for specific files.",
     "serve_model": "Start serving a model with vLLM, SGLang, llama.cpp, Ollama, or Diffusers. cmd MUST start with the binary directly — e.g. `vllm serve /mnt/HADES/models/Qwen3.5-397B-A17B-AWQ --port 8003 --tensor-parallel-size 8 …`. NEVER prefix with `cd …`, `source …`, or chain with `&&`/`||` — those get rejected by the validator. The venv activation (env_prefix) and CUDA env are added automatically from the target host's saved settings. For image/inpainting/diffusion use python3 scripts/diffusion_server.py --model <repo> --port 8100. After launch, call list_served_models for readiness/errors and retry suggestions. If serve_model fails with 'Invalid characters in cmd', simplify to the bare binary + args.",
     "list_served_models": "List currently running model servers in the Cookbook — shows status (loading, ready, idle, error), model name, port, throughput, and serve failure diagnosis/retry suggestions. Use when the user asks 'what's running', 'show my cookbook', 'which models are up', 'what's serving'.",
@@ -405,8 +405,10 @@ class ToolIndex:
             {"chat_with_model", "ask_teacher", "list_models"},
         # Deep research intent (incl. common typo "reserach")
         frozenset({"web search", "search the web", "search online", "look up",
-                   "google", "latest", "current", "news", "weather",
-                   "forecast", "stock price", "price of"}):
+                   "find info online", "find information online",
+                   "find info", "find information", "online about",
+                   "on the internet", "google", "latest", "current", "news",
+                   "weather", "forecast", "stock price", "price of"}):
             {"web_search", "web_fetch"},
         frozenset({"research", "reserach", "reasearch", "look into", "investigate",
                    "deep dive", "deep research", "find out about", "study up on",

src/tool_parsing.py

@@ -10,9 +10,10 @@ import bisect
 import json
 import logging
 import re
-from typing import List, Optional
+from typing import List, Optional, Tuple
 
 from src.agent_tools import ToolBlock, TOOL_TAGS
+from src.tool_security import BUILTIN_EMAIL_TOOLS
 
 logger = logging.getLogger(__name__)
 
@@ -20,12 +21,63 @@ logger = logging.getLogger(__name__)
 # Regex patterns
 # ---------------------------------------------------------------------------
 
-# Pattern 1: ```bash ... ``` fenced code blocks
+# Pattern 1: ```bash ... ``` fenced code blocks. The tag may be followed by a
+# newline (classic form) or by inline JSON args on the same line
+# (```list_email_accounts {}). The same-line part is captured separately
+# (group 2) and judged by _fenced_tool_call below — the regex alone only
+# requires it to start with { or [; anything else after the tag is a Markdown
+# info string (```python title="example.py") and the fence never matches.
+# (?![\w-]) keeps the alternation from prefix-matching longer fence tags:
+# without it, ```python3 would match as tool "python" with content "3\n..."
+# and execute as code.
 _TOOL_BLOCK_RE = re.compile(
-    r"```(" + "|".join(TOOL_TAGS) + r")\s*\n([\s\S]*?)```",
+    r"```(" + "|".join(TOOL_TAGS) + r")(?![\w-])"
+    r"[ \t]*([{\[][^\n]*?)?[ \t]*(?=\r?\n|```)\r?\n?([\s\S]*?)```",
     re.IGNORECASE,
 )
 
+# Tags whose fenced content is raw code, not JSON args. Same-line text after
+# these tags is Markdown fence metadata on a real language (```bash {title=
+# "setup"}), never inline tool args — only the classic tag-then-newline form
+# executes for them.
+_CODE_FENCE_TAGS = frozenset({"bash", "python"})
+
+
+def _fenced_tool_call(m) -> Optional[Tuple[str, str]]:
+    """Classify a Pattern-1 fence match: (tag, content) when it is an
+    executable tool call, None when the fence must stay display text.
+
+    Shared by parse_tool_blocks and strip_tool_blocks so the execute and
+    display decisions can never disagree: a fence that doesn't execute is
+    never stripped, and vice versa.
+
+    Same-line text after the tag only counts as inline tool args when the
+    tag's tool takes JSON args (not a code tag) AND the text is valid
+    standalone JSON. ```bash {title="setup"} and ```python {"x": 1} are
+    fence attributes on real languages, and {title="x"} on any tag is
+    metadata, not arguments — all of those stay visible and inert.
+    """
+    tag = m.group(1).lower()
+    inline = (m.group(2) or "").strip()
+    body = (m.group(3) or "").strip()
+    if not inline:
+        return tag, body
+    if tag in _CODE_FENCE_TAGS:
+        return None
+    # Inline args may continue onto following lines (a JSON object opened on
+    # the tag line); the combined text must parse as JSON or nothing runs.
+    content = f"{inline}\n{body}" if body else inline
+    try:
+        json.loads(content)
+    except (ValueError, TypeError):
+        return None
+    return tag, content
+
+
+def _strip_executed_fence(m) -> str:
+    """re.sub callback: remove only fences that parse as tool calls."""
+    return "" if _fenced_tool_call(m) is not None else m.group(0)
+
 # Pattern 2: [TOOL_CALL] ... [/TOOL_CALL] blocks (some models use this format)
 # Matches: {tool => "shell", args => {--command "ls -la"}} etc.
 _TOOL_CALL_RE = re.compile(
@@ -114,6 +166,34 @@ _TOOL_CODE_RE = re.compile(
 _TOOL_CODE_OPEN_RE = re.compile(r"<tool_code>\s*\{", re.IGNORECASE)
 _TOOL_CODE_CLOSE_RE = re.compile(r"\}\s*</tool_code>", re.IGNORECASE)
 
+# Pattern 4b: Gemma-style <|tool_call|> call:tool_name{args} <tool_call|>
+_GEMMA_TOOL_CALL_RE = re.compile(
+    r"<\|?tool_call\|?>\s*call:([\w\d_-]+)\s*(\{[\s\S]*?\})\s*<\|?tool_call\|?>",
+    re.IGNORECASE,
+)
+
+# Pattern 4c: Open-function wrapper emitted by some local MLX/Exo models.
+# Example:
+#   <function_model>
+#   <function_call>web_search</function_call>
+#   <parameters>{"query":"Sweden news today"}</parameters>
+#   </function_model>
+_FUNCTION_MODEL_OPEN_RE = re.compile(r"<function_model>\s*", re.IGNORECASE)
+_FUNCTION_MODEL_CLOSE_RE = re.compile(r"</function_model>", re.IGNORECASE)
+_FUNCTION_MODEL_NAME_RE = re.compile(
+    r"<function_call>\s*([A-Za-z_][\w-]*)\s*</function_call>",
+    re.IGNORECASE,
+)
+_FUNCTION_MODEL_PARAMS_OPEN_RE = re.compile(r"<parameters>\s*", re.IGNORECASE)
+_FUNCTION_MODEL_PARAMS_CLOSE_RE = re.compile(r"</parameters>", re.IGNORECASE)
+_QWEN_ROLE_MARKER_RE = re.compile(r"</?\|(?:assistant|assistan|user|system|tool)\|>?|</\|end\|>?", re.IGNORECASE)
+_QWEN_BARE_MARKER_RE = re.compile(
+    r"(?:^|[\t\r\n ])(?:\|?end\|?|/?\|end\|)(?=[\t\r\n ]|$)|"
+    r"(?:^|[\t\r\n ])assistan(?:t)?(?=[\t\r\n ]|$)",
+    re.IGNORECASE,
+)
+
+
 # Pattern 5: DeepSeek DSML markup leaking into content. When deepseek
 # models can't emit structured tool_calls (e.g. we sent no tool schemas
 # that round, or the API didn't parse them), they fall back to raw
@@ -256,6 +336,17 @@ _RAW_WEB_JSON_TOOL_RE = re.compile(
 )
 _RAW_WEB_JSON_ALLOWED_KEYS = {"query", "queries", "time_filter", "freshness", "max_pages"}
 
+# Narrow rescue for models that ignore native tool calling and print the UI
+# command as plain text. Keep this intentionally tiny: open-panel is a harmless
+# frontend event, while broad plain-text parsing of shell/doc/email tools would
+# be unsafe.
+_PLAIN_UI_OPEN_PANEL_RE = re.compile(
+    r"(?im)^\s*(?:`{1,3})?\s*ui_control\s+open_panel\s+"
+    r"(documents?|library|gallery|images?|email|inbox|mail|sessions?|chats?|history|"
+    r"notes?|brain|memor(?:y|ies)|skills?|settings|preferences|cookbook|models?)"
+    r"\s*(?:`{1,3})?\s*$"
+)
+
 
 # ---------------------------------------------------------------------------
 # Parsing functions
@@ -496,6 +587,205 @@ def _parse_raw_web_json_lookup(text: str) -> Optional[tuple[ToolBlock, tuple[int
                 return block, (start, start + end)
     return None
 
+
+def _looks_like_openai_tool_call_blob(value) -> bool:
+    """Return True for raw OpenAI-style tool-call JSON leaked as text."""
+    if isinstance(value, list):
+        return bool(value) and all(_looks_like_openai_tool_call_blob(item) for item in value)
+    if not isinstance(value, dict):
+        return False
+    fn = value.get("function")
+    if isinstance(fn, dict) and isinstance(fn.get("name"), str):
+        return True
+    return False
+
+
+def _raw_openai_tool_call_to_block(value) -> Optional[ToolBlock]:
+    if isinstance(value, list):
+        for item in value:
+            block = _raw_openai_tool_call_to_block(item)
+            if block:
+                return block
+        return None
+    if not isinstance(value, dict):
+        return None
+    fn = value.get("function")
+    if not isinstance(fn, dict):
+        return None
+    name = str(fn.get("name") or "").strip()
+    if not name:
+        return None
+    tool_type = _TOOL_NAME_MAP.get(name, name)
+    raw_args = fn.get("arguments") or {}
+    try:
+        args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
+    except (json.JSONDecodeError, TypeError):
+        args = {}
+    if not isinstance(args, dict):
+        args = {}
+    # Common local-model typo seen in raw OpenAI JSON leaks.
+    if "text" not in args and "tex" in args:
+        args["text"] = args.get("tex")
+
+    if tool_type.startswith("mcp__"):
+        return ToolBlock(tool_type, json.dumps(args) if args else "{}")
+    if name in BUILTIN_EMAIL_TOOLS:
+        return ToolBlock(f"mcp__email__{name}", json.dumps(args) if args else "{}")
+    if tool_type not in TOOL_TAGS:
+        return None
+
+    if tool_type == "bash":
+        content = args.get("command", "")
+    elif tool_type == "python":
+        content = args.get("code", "")
+    elif tool_type == "web_search":
+        content = args.get("query", "")
+        queries = args.get("queries")
+        if not content and isinstance(queries, list) and queries:
+            content = str(queries[0])
+        elif not content and queries:
+            content = str(queries)
+        tf = args.get("time_filter")
+        if content and isinstance(tf, str) and tf in ("day", "week", "month", "year"):
+            content = json.dumps({"query": content, "time_filter": tf})
+    elif tool_type == "web_fetch":
+        content = args.get("url") or args.get("domain") or ""
+    elif tool_type == "read_file":
+        content = json.dumps(args) if (args.get("offset") or args.get("limit")) else args.get("path", "")
+    elif tool_type in ("grep", "glob", "ls", "edit_file"):
+        content = json.dumps(args) if args else "{}"
+    elif tool_type == "write_file":
+        content = args.get("path", "") + "\n" + args.get("content", "")
+    elif tool_type == "create_document":
+        parts = [args.get("title", "Untitled")]
+        if args.get("language"):
+            parts.append(args["language"])
+        parts.append(args.get("content", ""))
+        content = "\n".join(parts)
+    elif tool_type == "update_document":
+        content = args.get("content", "")
+    elif tool_type in ("edit_document", "suggest_document"):
+        marker = "SUGGEST" if tool_type == "suggest_document" else "REPLACE"
+        blocks = []
+        for edit in args.get("suggestions" if tool_type == "suggest_document" else "edits", []) or []:
+            if not isinstance(edit, dict):
+                continue
+            block = f'<<<FIND>>>\n{edit.get("find", "")}\n<<<{marker}>>>\n{edit.get("replace", "")}'
+            if tool_type == "suggest_document":
+                block += f'\n<<<REASON>>>\n{edit.get("reason", "")}'
+            blocks.append(block + "\n<<<END>>>")
+        content = "\n".join(blocks)
+    elif tool_type == "search_chats":
+        content = args.get("query", "")
+    elif tool_type == "chat_with_model":
+        content = args.get("model", "") + "\n" + args.get("message", "")
+    elif tool_type == "create_session":
+        content = args.get("name", "Untitled") + "\n" + args.get("model", "")
+    elif tool_type == "list_sessions":
+        content = args.get("filter", "")
+    elif tool_type == "send_to_session":
+        content = args.get("session_id", "") + "\n" + args.get("message", "")
+    elif tool_type == "pipeline":
+        content = json.dumps({"steps": args.get("steps", [])})
+    elif tool_type == "manage_session":
+        action = args.get("action", "")
+        if action == "list":
+            keyword = args.get("keyword", "") or args.get("value", "")
+            content = "list" + (("\n" + keyword) if keyword and keyword.lower() != "current" else "")
+        else:
+            content = action + "\n" + args.get("session_id", "current")
+            if args.get("value"):
+                content += "\n" + args["value"]
+    elif tool_type == "manage_memory":
+        action = args.get("action", "")
+        if action == "add":
+            content = "add\n" + str(args.get("text", ""))
+            if args.get("category"):
+                content += "\n" + str(args["category"])
+        elif action == "edit":
+            content = "edit\n" + str(args.get("memory_id", "")) + "\n" + str(args.get("text", ""))
+        elif action == "delete":
+            content = "delete\n" + str(args.get("memory_id", ""))
+        elif action == "search":
+            content = "search\n" + str(args.get("text", ""))
+        elif action == "list":
+            content = "list" + (("\n" + str(args["category"])) if args.get("category") else "")
+        else:
+            content = action
+    elif tool_type == "ui_control":
+        action = args.get("action", "")
+        name_arg = args.get("name", "")
+        value = args.get("value", "")
+        if action == "open_panel":
+            content = f"open_panel {name_arg or value}"
+        elif action == "toggle":
+            content = f"toggle {name_arg} {value}"
+        else:
+            content = action
+    elif tool_type in ("manage_tasks", "manage_skills", "api_call", "manage_endpoints",
+                       "manage_mcp", "manage_webhooks", "manage_tokens",
+                       "manage_documents", "manage_settings", "manage_notes",
+                       "manage_research", "manage_bg_jobs"):
+        content = json.dumps(args)
+    elif tool_type in ("get_workspace", "list_models"):
+        content = args.get("filter", "") if tool_type == "list_models" else ""
+    else:
+        content = json.dumps(args) if args else ""
+    return ToolBlock(tool_type, str(content or ""))
+
+
+def _parse_raw_openai_tool_call_json(text: str) -> Optional[ToolBlock]:
+    if not isinstance(text, str) or '"function"' not in text:
+        return None
+    decoder = json.JSONDecoder()
+    for match in re.finditer(r"[\[{]", text):
+        try:
+            parsed, _end = decoder.raw_decode(text[match.start():])
+        except json.JSONDecodeError:
+            continue
+        block = _raw_openai_tool_call_to_block(parsed)
+        if block:
+            return block
+    return None
+
+
+def _strip_raw_openai_tool_call_json(text: str) -> str:
+    """Strip raw JSON tool calls such as {"function": {...}, "type": "function"}.
+
+    Some local models emit native tool-call JSON into assistant text. The agent
+    can still parse/execute it through the native path, but the raw payload must
+    not render or persist as prose.
+    """
+    if not isinstance(text, str) or '"function"' not in text:
+        return text
+    decoder = json.JSONDecoder()
+    pieces = []
+    pos = 0
+    changed = False
+    for match in re.finditer(r"[\[{]", text):
+        start = match.start()
+        if start < pos:
+            continue
+        try:
+            parsed, rel_end = decoder.raw_decode(text[start:])
+        except json.JSONDecodeError:
+            continue
+        end = start + rel_end
+        if not _looks_like_openai_tool_call_blob(parsed):
+            continue
+        pieces.append(text[pos:start])
+        pos = end
+        changed = True
+        # Common broken local-model suffix: a standalone ] before a role marker.
+        while pos < len(text) and text[pos] in " \t\r\n":
+            pos += 1
+        if pos < len(text) and text[pos] == "]":
+            pos += 1
+    if not changed:
+        return text
+    pieces.append(text[pos:])
+    return "".join(pieces)
+
 def _parse_tool_call_block(raw: str) -> Optional[ToolBlock]:
     """Parse a [TOOL_CALL] block into a ToolBlock.
 
@@ -780,6 +1070,58 @@ def _parse_tool_code_block(raw: str) -> Optional[ToolBlock]:
         return ToolBlock(tool_name, content.strip())
     return None
 
+def _parse_gemma_tool_call(tool_name: str, body: str) -> Optional[ToolBlock]:
+    """Parse a Gemma-style call:tool_name{...} block into a ToolBlock."""
+    tool_name = tool_name.strip().lower().replace("-", "_")
+    body = body.strip()
+    if not body:
+        return None
+
+    # Replace custom Gemma string delimiters with standard quotes
+    body = body.replace('<|"|>', '"').replace('<|"', '"').replace('"|>', '"')
+
+    # Try standard JSON parsing
+    params = {}
+    try:
+        params = json.loads(body)
+        if not isinstance(params, dict):
+            params = {}
+    except json.JSONDecodeError:
+        # Try unquoted keys repair: e.g. {query: "..."} -> {"query": "..."}
+        try:
+            repaired = re.sub(r'([{,]\s*)(\w+)\s*:', r'\1"\2":', body)
+            params = json.loads(repaired)
+            if not isinstance(params, dict):
+                params = {}
+        except Exception:
+            # Simple regex key-value extraction fallback
+            params = {}
+            for m in re.finditer(r'(\w+)\s*:\s*["\']?(.*?)["\']?(?=\s*,\s*\w+\s*:|\s*\})', body):
+                k = m.group(1)
+                v = m.group(2).strip()
+                params[k] = v
+
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(tool_name, json.dumps(params))
+
+
+def _parse_function_model_call(body: str) -> Optional[ToolBlock]:
+    """Parse <function_model><function_call>tool</...><parameters>...</...>."""
+    name_match = _FUNCTION_MODEL_NAME_RE.search(body or "")
+    if not name_match:
+        return None
+    tool_name = name_match.group(1).strip().lower().replace("-", "_")
+    params = "{}"
+    for _ms, inner_start, inner_end, _me in _iter_delimited(
+        body,
+        _FUNCTION_MODEL_PARAMS_OPEN_RE,
+        _FUNCTION_MODEL_PARAMS_CLOSE_RE,
+    ):
+        params = body[inner_start:inner_end].strip() or "{}"
+        break
+    from src.tool_schemas import function_call_to_tool_block
+    return function_call_to_tool_block(tool_name, params)
+
 
 def _iter_delimited(text, open_re, close_re):
     """Yield ``(match_start, inner_start, inner_end, match_end)`` for each
@@ -923,9 +1265,20 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
     # Pattern 1: fenced code blocks (skipped when `skip_fenced` — see docstring).
     if not skip_fenced:
         for m in _TOOL_BLOCK_RE.finditer(text):
-            tag = m.group(1).lower()
-            content = m.group(2).strip()
+            call = _fenced_tool_call(m)
+            if call is None:
+                continue
+            tag, content = call
             if not content:
+                # An empty fence is still an unambiguous call for the email
+                # tools — ```list_email_accounts``` with no body is a shape
+                # local models really emit for no-arg tools. Dispatch with
+                # empty args and let the tool's own validation answer;
+                # silently dropping the call left models concluding email was
+                # broken. Other tags (bash, python, ...) keep skipping: empty
+                # content is nothing to run.
+                if tag in BUILTIN_EMAIL_TOOLS:
+                    blocks.append(ToolBlock(tag, ""))
                 continue
             # If a code block's content is an <invoke> XML call (some models wrap
             # tool calls in ```python or ```xml fences), parse the invoke instead.
@@ -1012,12 +1365,45 @@ def parse_tool_blocks(text: str, skip_fenced: bool = False) -> List[ToolBlock]:
             if block:
                 blocks.append(block)
 
+    # Pattern 4b: Gemma-style <|tool_call|> blocks
+    if not blocks:
+        for m in _GEMMA_TOOL_CALL_RE.finditer(text):
+            tool_name = m.group(1)
+            body = m.group(2)
+            block = _parse_gemma_tool_call(tool_name, body)
+            if block:
+                blocks.append(block)
+
+    # Pattern 4c: <function_model> wrapper from local MLX/Exo models.
+    if not blocks:
+        for _ms, inner_start, inner_end, _me in _iter_delimited(
+            text, _FUNCTION_MODEL_OPEN_RE, _FUNCTION_MODEL_CLOSE_RE
+        ):
+            block = _parse_function_model_call(text[inner_start:inner_end])
+            if block:
+                blocks.append(block)
+
+    # Pattern 4d: raw OpenAI-style tool-call JSON leaked as assistant text.
+    # Example: {"function":{"arguments":"{\"action\":\"add\"}","name":"manage_memory"},"type":"function"}
+    if not blocks:
+        block = _parse_raw_openai_tool_call_json(text)
+        if block:
+            blocks.append(block)
+
     # Pattern 6: local text-model web_search call leaked as prose + bare JSON.
     if not blocks and not skip_fenced:
         raw_web_json = _parse_raw_web_json_lookup(text)
         if raw_web_json:
             blocks.append(raw_web_json[0])
 
+    # Pattern 7: plain `ui_control open_panel notes` line. This commonly comes
+    # from weaker native-tool models after reading the tool docs but failing to
+    # emit the actual structured call.
+    if not blocks:
+        m = _PLAIN_UI_OPEN_PANEL_RE.search(text)
+        if m:
+            blocks.append(ToolBlock("ui_control", f"open_panel {m.group(1).lower()}"))
+
     return blocks
 
 
@@ -1037,7 +1423,10 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
     # Normalize DSML first so its markup gets stripped by the <invoke>
     # / <tool_call> removers below instead of leaking to the user.
     text = _normalize_dsml(text)
-    cleaned = text if skip_fenced else _TOOL_BLOCK_RE.sub('', text)
+    # Keep the executed-vs-illustrative fence distinction (only strip fences
+    # that actually dispatched; leave example fences from native models inert
+    # but visible), then remove [TOOL_CALL]{...}[/TOOL_CALL] markup.
+    cleaned = text if skip_fenced else _TOOL_BLOCK_RE.sub(_strip_executed_fence, text)
     # Forward-only removal mirrors parse_tool_blocks: _strip_delimited pairs each
     # opener with a later closer and stops when none is reachable, so untrusted
     # output can't drive the O(n^2) lazy-rescan (ReDoS); see _iter_delimited.
@@ -1046,11 +1435,17 @@ def strip_tool_blocks(text: str, skip_fenced: bool = False) -> str:
     cleaned = _strip_delimited(cleaned, _XML_TOOL_CALL_OPEN_RE, _XML_TOOL_CALL_CLOSE_RE)
     cleaned = _XML_OPEN_TOOL_CALL_RE.sub('', cleaned)
     cleaned = _strip_delimited(cleaned, _TOOL_CODE_OPEN_RE, _TOOL_CODE_CLOSE_RE)
+    cleaned = _GEMMA_TOOL_CALL_RE.sub('', cleaned)
+    cleaned = _strip_delimited(cleaned, _FUNCTION_MODEL_OPEN_RE, _FUNCTION_MODEL_CLOSE_RE)
+    cleaned = _strip_raw_openai_tool_call_json(cleaned)
+    cleaned = _QWEN_ROLE_MARKER_RE.sub('', cleaned)
+    cleaned = _QWEN_BARE_MARKER_RE.sub(' ', cleaned)
     if not skip_fenced:
         raw_web_json = _parse_raw_web_json_lookup(cleaned)
         if raw_web_json:
             _, (start, end) = raw_web_json
             cleaned = cleaned[:start] + cleaned[end:]
+    cleaned = _PLAIN_UI_OPEN_PANEL_RE.sub("", cleaned)
     # Strip bare <invoke> blocks not wrapped in <tool_call>
     cleaned = _strip_bare_invoke_markup(cleaned)
     cleaned = re.sub(r'\n{3,}', '\n\n', cleaned)

src/tool_schemas.py

@@ -14,9 +14,19 @@ from typing import Optional
 
 from src.agent_tools import ToolBlock, TOOL_TAGS
 from src.tool_parsing import _TOOL_NAME_MAP
+from src.tool_security import BUILTIN_EMAIL_TOOLS
 
 logger = logging.getLogger(__name__)
 
+
+_REQUIRED_NATIVE_TOOL_ARGS = {
+    "web_search": ("query", "queries"),
+    "web_fetch": ("url",),
+    "read_file": ("path",),
+    "write_file": ("path",),
+    "edit_file": ("path",),
+}
+
 # ---------------------------------------------------------------------------
 # OpenAI-compatible function tool schemas
 # ---------------------------------------------------------------------------
@@ -186,7 +196,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "create_document",
-            "description": "Create a new document in the editor panel. Use this when the user asks to write, create, build, or generate code, scripts, programs, games, apps, or any substantial content (>15 lines) AND there is no already-open document/email draft that the request refers to. If an email compose draft is open, edit that draft instead of creating another document. NEVER put large code blocks directly in chat — use this tool instead.",
+            "description": "Create a new document in the editor panel. Use this when the user asks to write, create, build, make, or generate code, scripts, programs, games, apps, or any long-form or structured content that is more than a short paragraph, AND there is no already-open document/email draft that the request refers to. If an email compose draft is open, edit that draft instead of creating another document. NEVER put large generated content directly in chat — use this tool instead.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -325,7 +335,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "send_to_session",
-            "description": "Send a message to an existing chat and get the model's response. The chat keeps its conversation history.",
+            "description": "Send a new message to an existing live chat and get that chat model's response. Do not use this to retrieve, read, summarize, or inspect old chats; use search_chats or list_sessions for past chat evidence.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -415,7 +425,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "ui_control",
-            "description": "Control the user interface. Actions: toggle (turn tools on/off), open_panel (open a modal: documents/library, gallery, email, sessions, notes, memories/brain, skills, settings, cookbook), open_email_reply (open an email reply draft document; does NOT send), set_mode, switch_model, set_theme (built-in presets: dark, light, midnight, paper, cyberpunk, retrowave, forest, ocean, ume, copper, terminal, organs, lavender, gpt, claude, cute), create_theme (CREATE any custom theme with a name + colors object — pick distinctive, evocative hex colors that match the requested aesthetic, NOT generic defaults. The theme auto-applies after creation). When a user asks for ANY theme not in the built-in preset list, ALWAYS use create_theme.",
+            "description": "Control the user interface. Actions: toggle (turn tools on/off), open_panel (open a modal: documents/library, gallery, email, sessions, notes, memories/brain, skills, settings, cookbook), open_email_reply (open an email reply draft document; DOES NOT send. For 'write/draft a reply saying X', include body with the drafted reply), set_mode, switch_model, set_theme (built-in presets: dark, light, midnight, paper, cyberpunk, retrowave, forest, ocean, ume, copper, terminal, organs, lavender, gpt, claude, cute), create_theme (CREATE any custom theme with a name + colors object — pick distinctive, evocative hex colors that match the requested aesthetic, NOT generic defaults. The theme auto-applies after creation). When a user asks for ANY theme not in the built-in preset list, ALWAYS use create_theme.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -426,6 +436,7 @@ FUNCTION_TOOL_SCHEMAS = [
                     "uid": {"type": "string", "description": "Email UID for open_email_reply"},
                     "folder": {"type": "string", "description": "Email folder for open_email_reply (default INBOX)"},
                     "mode": {"type": "string", "description": "Reply draft mode for open_email_reply: reply, reply-all, or ai-reply"},
+                    "body": {"type": "string", "description": "For open_email_reply: reply body to pre-fill. Required whenever the user told you what the reply should say. Opens a draft, does not send."},
                     "colors": {"type": "object", "description": "For create_theme: the theme colors",
                                "properties": {
                                    "bg": {"type": "string", "description": "Background color (hex, e.g. #1a1a2e)"},
@@ -538,7 +549,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "manage_calendar",
-            "description": "Manage calendar events: list events in a date range, create, update, delete. Each event can carry a tag/category (event_type) and importance level. Resolve relative dates like today/tomorrow against the 'Current date and time' system context, then pass ISO 8601 datetimes in the user's local wall time; for all-day events set all_day=true and pass YYYY-MM-DD. For event reminders/alarms, pass reminder_minutes; the tool creates the Odysseus note reminder, so do not also call manage_notes for the same reminder.",
+            "description": "Manage calendar events: list events in a date range, create, update, delete. Each event can carry a tag/category (event_type) and importance level. Resolve relative dates like today/tomorrow against the 'Current date and time' system context, then pass ISO 8601 datetimes in the user's local wall time; for all-day events set all_day=true and pass YYYY-MM-DD. For event reminders/alarms, pass reminder_minutes; the tool creates the Odysseus note reminder, so do not also call manage_notes for the same reminder. Do not set rrule for single-occurrence requests such as 'next Wednesday only'; use rrule only when the user explicitly wants recurrence.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -554,12 +565,12 @@ FUNCTION_TOOL_SCHEMAS = [
                     "uid": {"type": "string", "description": "Event UID (for update/delete)"},
                     "calendar_href": {"type": "string", "description": "Specific calendar URL (optional; defaults to first calendar)"},
                     "calendar": {"type": "string", "description": "Filter list_events by calendar name or href"},
-                    "start": {"type": "string", "description": "list_events range start (ISO datetime); defaults to today. Prefer start; backend also accepts start_date, range_start, from, dtstart, since."},
-                    "end": {"type": "string", "description": "list_events range end (ISO datetime); defaults to +14 days. Prefer end; backend also accepts end_date, range_end, to, dtend, until."},
+                    "start": {"type": "string", "description": "list_events range start (ISO datetime). Use this for month/week requests after resolving the date range; do not pass a loose query string. Prefer start; backend also accepts start_time, start_date, range_start, from, dtstart, since."},
+                    "end": {"type": "string", "description": "list_events range end (ISO datetime). Use this for month/week requests after resolving the date range; defaults to +14 days only when no range is requested. Prefer end; backend also accepts end_time, end_date, range_end, to, dtend, until."},
                     "event_type": {"type": "string", "description": "Tag / category for the event. Common values: work, personal, health, travel, meal, social, admin, other. Aliases accepted: tag, category, type."},
                     "importance": {"type": "string", "enum": ["low", "normal", "high", "critical"], "description": "Priority level (defaults to 'normal')"},
                     "reminder_minutes": {"type": "integer", "description": "For create_event: create an Odysseus reminder this many minutes before the event, e.g. 5 for 'reminder 5 min before'."},
-                    "rrule": {"type": "string", "description": "Recurrence rule in iCalendar RRULE format, e.g. 'FREQ=WEEKLY;BYDAY=MO' for weekly on Monday. Use with create_event or update_event."}
+                    "rrule": {"type": "string", "description": "Recurrence rule in iCalendar RRULE format, e.g. 'FREQ=WEEKLY;BYDAY=MO' for weekly on Monday. Use with create_event or update_event. For update_event, pass an explicit empty string to remove recurrence and make the event single-occurrence."}
                 },
                 "required": ["action"]
             }
@@ -569,14 +580,15 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "manage_notes",
-            "description": "Manage notes and checklists (Google Keep-style): list, add, update, delete, toggle_item. IMPORTANT: For to-do lists / checklists, set note_type='checklist' and pass the items as the `checklist_items` array — do NOT serialize them into `content` as plain text. For freeform notes, use note_type='note' and put the body in `content`. `due_date` accepts natural language like 'tomorrow at 9am' (parsed in the user's timezone) and fires a notification — do not also create a calendar event for the same reminder.",
+            "description": "Manage notes and checklists (Google Keep-style): list, view, add, update, delete, toggle_item. Use list/search to find candidate notes, then view with the note id when you need the full body. IMPORTANT: For to-do lists / checklists, set note_type='checklist' and pass the items as the `checklist_items` array — do NOT serialize them into `content` as plain text. For freeform notes, use note_type='note' and put the body in `content`. `due_date` accepts natural language like 'tomorrow at 9am' (parsed in the user's timezone) and fires a notification — do not also create a calendar event for the same reminder.",
             "parameters": {
                 "type": "object",
                 "properties": {
                     "action": {"type": "string",
-                               "enum": ["list", "add", "update", "delete", "toggle_item"],
+                               "enum": ["list", "search", "view", "add", "update", "delete", "toggle_item"],
                                "description": "The action to perform"},
                     "id": {"type": "string", "description": "Note id (for update/delete/toggle_item); 8-char prefix is fine"},
+                    "query": {"type": "string", "description": "Search text for action='search'"},
                     "title": {"type": "string", "description": "Note title (for add/update)"},
                     "content": {"type": "string", "description": "Freeform body text. Use this for note_type='note'. Do NOT use this for checklists — pass `checklist_items` instead."},
                     "note_type": {"type": "string", "enum": ["note", "checklist"],
@@ -1023,7 +1035,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "manage_contact",
-            "description": "Create, update, delete, or list the user's CardDAV contacts. Use to save a new contact, update an existing one (email/phone/address), or remove one. For update/delete you need the contact's uid — call action='list' first to find it. Writes go through the same dedupe + validation as the Contacts UI.",
+            "description": "Create, update, delete, or list the user's CardDAV contacts. Use to save a new contact, update an existing one (email/phone/address), or remove one. Add does not require email: name + phone or name + address is valid. For update/delete you need the contact's uid — call action='list' first to find it. Writes go through the same dedupe + validation as the Contacts UI.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -1031,9 +1043,9 @@ FUNCTION_TOOL_SCHEMAS = [
                                "description": "list = show all contacts (with uids); add = create; update = edit by uid; delete = remove by uid."},
                     "uid": {"type": "string", "description": "Contact UID (required for update/delete; get it from action=list)."},
                     "name": {"type": "string", "description": "Contact's display name (for add/update)."},
-                    "email": {"type": "string", "description": "Single email address (convenience for add, or the primary email for update)."},
-                    "emails": {"type": "array", "items": {"type": "string"}, "description": "Full list of email addresses (for update; first is primary)."},
-                    "phones": {"type": "array", "items": {"type": "string"}, "description": "Full list of phone numbers (for update)."},
+                    "email": {"type": "string", "description": "Single email address (convenience for add, or the primary email for update). Optional when phone or address is provided."},
+                    "emails": {"type": "array", "items": {"type": "string"}, "description": "Full list of email addresses (first is primary)."},
+                    "phones": {"type": "array", "items": {"type": "string"}, "description": "Full list of phone numbers. Valid for add/update."},
                     "address": {"type": "string", "description": "Postal/mailing address as a single human-readable string."},
                 },
                 "required": ["action"]
@@ -1106,7 +1118,7 @@ FUNCTION_TOOL_SCHEMAS = [
         "type": "function",
         "function": {
             "name": "reply_to_email",
-            "description": "SEND a reply email immediately by UID. Do not use this when the user asks to open/start a reply window or draft; use ui_control action=open_email_reply instead. For follow-up 'reply ...' requests where the user clearly wants to send now, use the exact UID from the latest read_email/list_emails result; never invent UID 1. Automatically threads with In-Reply-To/References headers.",
+            "description": "SEND a reply email immediately by UID. Do not use this when the user asks to write/draft/open/start a reply; use ui_control action=open_email_reply with body instead so the user can review. Only use when the user explicitly says to send now. Use the exact UID from the latest read_email/list_emails result; never invent UID 1. Automatically threads with In-Reply-To/References headers.",
             "parameters": {
                 "type": "object",
                 "properties": {
@@ -1210,38 +1222,113 @@ FUNCTION_TOOL_SCHEMAS = [
 # Converter: native function call -> ToolBlock
 # ---------------------------------------------------------------------------
 
+def _decode_loose_json_string(value: str) -> str:
+    """Decode common JSON string escapes without requiring inner quotes to be escaped."""
+    out = []
+    i = 0
+    while i < len(value):
+        ch = value[i]
+        if ch != "\\" or i + 1 >= len(value):
+            out.append(ch)
+            i += 1
+            continue
+        nxt = value[i + 1]
+        if nxt == "n":
+            out.append("\n")
+        elif nxt == "r":
+            out.append("\r")
+        elif nxt == "t":
+            out.append("\t")
+        elif nxt == "b":
+            out.append("\b")
+        elif nxt == "f":
+            out.append("\f")
+        elif nxt in ('"', "\\", "/"):
+            out.append(nxt)
+        elif nxt == "u" and i + 5 < len(value):
+            try:
+                out.append(chr(int(value[i + 2:i + 6], 16)))
+                i += 4
+            except ValueError:
+                out.append("\\" + nxt)
+        else:
+            out.append("\\" + nxt)
+        i += 2
+    return "".join(out)
+
+
+def _repair_document_function_args(tool_type: str, arguments: str) -> Optional[dict]:
+    """Salvage obvious malformed document tool args from local model wrappers.
+
+    The doc LoRA sometimes emits the right native tool call but puts raw quotes
+    inside the document text, making the surrounding JSON invalid. Treat that as
+    a wrapper parse failure, not a semantic tool-choice failure.
+    """
+    if tool_type != "update_document" or not isinstance(arguments, str):
+        return None
+    raw = arguments.strip()
+    if not raw.startswith("{") or not raw.endswith("}"):
+        return None
+    for key in ("content", "conten"):
+        marker = f'"{key}"'
+        key_pos = raw.find(marker)
+        if key_pos < 0:
+            continue
+        colon_pos = raw.find(":", key_pos + len(marker))
+        if colon_pos < 0:
+            continue
+        first_quote = raw.find('"', colon_pos + 1)
+        if first_quote < 0:
+            continue
+        close_brace = raw.rfind("}")
+        last_quote = raw.rfind('"', first_quote + 1, close_brace)
+        if last_quote <= first_quote:
+            continue
+        content = _decode_loose_json_string(raw[first_quote + 1:last_quote])
+        return {"content": content}
+    return None
+
+
 def function_call_to_tool_block(name: str, arguments: str) -> Optional[ToolBlock]:
     """Convert a native function call into a ToolBlock for the existing execution pipeline."""
+    tool_type = _TOOL_NAME_MAP.get(name, name)
     try:
         if not arguments or (isinstance(arguments, str) and not arguments.strip()):
             args = {}
         else:
             args = json.loads(arguments) if isinstance(arguments, str) else arguments
     except (json.JSONDecodeError, TypeError):
-        logger.error(f"Failed to parse function call arguments for {name}: {arguments}")
-        return None
-
-    tool_type = _TOOL_NAME_MAP.get(name, name)
-    _BUILTIN_EMAIL_TOOLS = {"list_email_accounts", "send_email", "list_emails", "read_email", "reply_to_email",
-                            "archive_email", "delete_email", "mark_email_read", "bulk_email", "download_attachment"}
+        args = _repair_document_function_args(tool_type, arguments)
+        if args is not None:
+            logger.warning(f"Repaired malformed document function call arguments for {name}")
+        else:
+            logger.error(f"Failed to parse function call arguments for {name}: {arguments}")
+            return None
 
     # Some models emit valid JSON that isn't an object (e.g. a bare array
     # ["ls -la"], string, or number) as function arguments. Most local tools keep
     # the legacy empty-object coercion for stream robustness, but email MCP tools
     # must fail closed so a malformed call cannot read the default mailbox.
+    # Uses the shared BUILTIN_EMAIL_TOOLS (single source of truth) so the
+    # fail-closed set can't drift from the dispatch/blocklist sets.
     if not isinstance(args, dict):
-        if tool_type.startswith("mcp__email__") or name in _BUILTIN_EMAIL_TOOLS:
+        if tool_type.startswith("mcp__email__") or name in BUILTIN_EMAIL_TOOLS:
             logger.warning(f"Non-object email function call arguments for {name}: {args!r}; rejecting")
             return None
         logger.warning(f"Non-object function call arguments for {name}: {args!r}; treating as empty")
         args = {}
 
+    required_args = _REQUIRED_NATIVE_TOOL_ARGS.get(tool_type)
+    if required_args and not any(str(args.get(key) or "").strip() for key in required_args):
+        logger.warning(f"Rejecting empty required arguments for function call {name}: {args!r}")
+        return None
+
     # Allow MCP tools through (namespaced as mcp__serverid__toolname)
     if tool_type.startswith("mcp__"):
         content = json.dumps(args) if args else "{}"
         return ToolBlock(tool_type, content)
     # Email tools are implemented as MCP — route them to email
-    if name in _BUILTIN_EMAIL_TOOLS:
+    if name in BUILTIN_EMAIL_TOOLS:
         return ToolBlock(f"mcp__email__{name}", json.dumps(args) if args else "{}")
     if tool_type not in TOOL_TAGS:
         logger.warning(f"Unknown function call: {name}")
@@ -1343,15 +1430,20 @@ def function_call_to_tool_block(name: str, arguments: str) -> Optional[ToolBlock
     elif tool_type == "manage_memory":
         action = args.get("action", "")
         if action == "add":
-            content = "add\n" + args.get("text", "")
+            text = args.get("text") or args.get("value") or args.get("content") or ""
+            if not text and args.get("key"):
+                text = str(args.get("key") or "")
+            content = "add\n" + str(text)
             if args.get("category"):
                 content += "\n" + args["category"]
+            elif args.get("key"):
+                content += "\n" + str(args["key"])
         elif action == "edit":
             content = "edit\n" + args.get("memory_id", "") + "\n" + args.get("text", "")
         elif action == "delete":
             content = "delete\n" + args.get("memory_id", "")
         elif action == "search":
-            content = "search\n" + args.get("text", "")
+            content = "search\n" + (args.get("text") or args.get("tex") or args.get("query") or "")
         elif action == "list":
             content = "list"
             if args.get("category"):
@@ -1373,6 +1465,9 @@ def function_call_to_tool_block(name: str, arguments: str) -> Optional[ToolBlock
             folder = args.get("folder") or value or "INBOX"
             mode = args.get("mode") or "reply"
             content = f"open_email_reply {uid} {folder} {mode}"
+            body = args.get("body") or args.get("extra") or args.get("content") or ""
+            if body:
+                content += f" {body}"
         elif action == "set_mode":
             content = f"set_mode {value or name}"
         elif action == "switch_model":

src/tool_security.py

@@ -8,10 +8,36 @@ from typing import Optional, Set
 logger = logging.getLogger(__name__)
 
 
+# Every tool exposed by the built-in email MCP server
+# (mcp_servers/email_server.py). Single source of truth: the fence tags
+# (TOOL_TAGS), bare-name dispatch (tool_execution), native-call mapping
+# (tool_schemas), and the non-admin blocklist below all derive from this set,
+# so a tool added to the email server can't become reachable under its bare
+# name without also being blocked for non-admins.
+BUILTIN_EMAIL_TOOLS = frozenset({
+    "list_email_accounts",
+    "list_emails",
+    "read_email",
+    "search_emails",
+    "send_email",
+    "reply_to_email",
+    "draft_email",
+    "draft_email_reply",
+    "ai_draft_email_reply",
+    "archive_email",
+    "delete_email",
+    "mark_email_read",
+    "bulk_email",
+    "download_attachment",
+})
+
+
 # Tools regular/public users must not execute directly. These either expose
 # server/runtime access, sensitive user data, external messaging, persistent
-# state changes, or generic loopback/integration surfaces.
-NON_ADMIN_BLOCKED_TOOLS = {
+# state changes, or generic loopback/integration surfaces. All email tools are
+# included (SECURITY.md: email/MCP capabilities are privileged admin
+# functionality).
+NON_ADMIN_BLOCKED_TOOLS = BUILTIN_EMAIL_TOOLS | {
     "bash",
     "python",
     "manage_bg_jobs",
@@ -34,10 +60,6 @@ NON_ADMIN_BLOCKED_TOOLS = {
     "manage_settings",
     "api_call",
     "app_api",
-    "send_email",
-    "reply_to_email",
-    "list_emails",
-    "read_email",
     "resolve_contact",
     "manage_contact",
     "manage_calendar",
@@ -74,8 +96,20 @@ PLAN_MODE_READONLY_TOOLS = {
     "search_chats",
     "list_models",
     "list_sessions",
+    # Read-only email tools. list_email_accounts must be here because the
+    # bare/qualified alias gate in execute_tool_block works both ways: it has
+    # a native function schema, so plan mode's schema-derived bare denylist
+    # contains it — and without this allowlist entry that bare entry would
+    # also block the qualified mcp__email__list_email_accounts call that the
+    # MCP read-only filter deliberately allows.
+    "list_email_accounts",
     "list_emails",
     "read_email",
+    # Explicitly read-only rather than allowed-by-omission: this PR makes
+    # every BUILTIN_EMAIL_TOOLS name fence-taggable, so each one must be
+    # classified — see the plan-mode partition test in
+    # tests/test_email_registry_sync.py.
+    "search_emails",
     "list_served_models",
     "list_downloads",
     "list_cached_models",
@@ -109,7 +143,14 @@ _PLAN_MODE_KNOWN_MUTATORS = {
     "manage_webhooks", "manage_tokens", "manage_settings", "manage_contact",
     "manage_calendar", "api_call", "app_api", "ui_control",
     "send_email", "reply_to_email", "bulk_email", "delete_email",
-    "archive_email", "mark_email_read", "download_model", "serve_model",
+    "archive_email", "mark_email_read",
+    # The draft tools create documents and download_attachment writes to
+    # disk — mutating. They have no native schemas (yet), so without these
+    # static entries plan-mode safety for their bare fence tags would depend
+    # entirely on the MCP read-only inventory being present and current.
+    "draft_email", "draft_email_reply", "ai_draft_email_reply",
+    "download_attachment",
+    "download_model", "serve_model",
     "stop_served_model", "cancel_download", "adopt_served_model", "serve_preset",
     "generate_image", "edit_image", "trigger_research", "manage_research",
     # Shell is never read-only-safe; block it explicitly so it stays out of plan
@@ -151,6 +192,28 @@ def plan_mode_disabled_tools() -> Set[str]:
     return (all_names | _PLAN_MODE_KNOWN_MUTATORS) - PLAN_MODE_READONLY_TOOLS
 
 
+def email_tool_policy_names(tool_name: str) -> frozenset:
+    """All policy-equivalent spellings of a tool name.
+
+    A bare built-in email tool name and its MCP-qualified mcp__email__<name>
+    form dispatch to the same email server tool, but policy sources spell
+    them either way — plan mode and the MCP settings toggle write qualified
+    names into denylists, chat-level toggles write bare ones. Every gate must
+    match against the full alias set, or a call in one spelling slips past a
+    denylist entry written in the other. Non-email names alias only to
+    themselves.
+    """
+    if not isinstance(tool_name, str):
+        return frozenset((tool_name,))
+    if tool_name in BUILTIN_EMAIL_TOOLS:
+        return frozenset((tool_name, f"mcp__email__{tool_name}"))
+    if tool_name.startswith("mcp__email__"):
+        bare = tool_name[len("mcp__email__"):]
+        if bare in BUILTIN_EMAIL_TOOLS:
+            return frozenset((tool_name, bare))
+    return frozenset((tool_name,))
+
+
 def is_public_blocked_tool(tool_name: Optional[str]) -> bool:
     """Return True when a non-admin/public user must not execute this tool.
 

src/tool_utils.py

@@ -54,6 +54,8 @@ def _parse_tool_args(content):
     if isinstance(content, str):
         try:
             args = json.loads(content) if content.strip() else {}
+            if not isinstance(args, dict):
+                args = {}
         except (json.JSONDecodeError, TypeError) as e:
             raise ValueError(str(e))
     elif isinstance(content, dict):

src/tools/calendar.py

@@ -208,11 +208,20 @@ async def do_manage_calendar(content: str, owner: Optional[str] = None) -> Dict:
         elif action == "list_events":
             try:
                 start_raw = _first_nonempty_arg(
-                    "start", "start_date", "range_start", "from", "dtstart", "since"
+                    "start", "start_time", "start_date", "range_start", "from", "dtstart", "since"
                 )
                 end_raw = _first_nonempty_arg(
-                    "end", "end_date", "range_end", "to", "dtend", "until"
+                    "end", "end_time", "end_date", "range_end", "to", "dtend", "until"
                 )
+                query_raw = args.get("query") or args.get("date_range") or args.get("range")
+                if query_raw and (not start_raw or not end_raw):
+                    return {
+                        "error": (
+                            "list_events needs explicit start/end ISO datetimes; "
+                            f"resolve the requested range ({query_raw!r}) and call manage_calendar again."
+                        ),
+                        "exit_code": 1,
+                    }
                 if start_raw:
                     start_dt = _parse_dt(start_raw)
                 else:
@@ -254,6 +263,7 @@ async def do_manage_calendar(content: str, owner: Optional[str] = None) -> Dict:
                     "calendar_href": ev.calendar_id,
                     "event_type": ev.event_type or "",
                     "importance": ev.importance or "normal",
+                    "rrule": ev.rrule or "",
                 })
             if not events:
                 response_text = f"No events between {start_dt.date().isoformat()} and {end_dt.date().isoformat()}."
@@ -268,6 +278,8 @@ async def do_manage_calendar(content: str, owner: Optional[str] = None) -> Dict:
                         line += f" #{ev['event_type']}"
                     if ev.get("importance") and ev["importance"] != "normal":
                         line += f" !{ev['importance']}"
+                    if ev.get("rrule"):
+                        line += f" repeats({ev['rrule']})"
                     if ev.get("location"):
                         line += f" @ {ev['location']}"
                     if ev.get("calendar"):
@@ -480,6 +492,10 @@ async def do_manage_calendar(content: str, owner: Optional[str] = None) -> Dict:
                 ev.event_type = _tag or None
             if args.get("importance") is not None:
                 ev.importance = args["importance"]
+            if args.get("rrule") is not None:
+                ev.rrule = args.get("rrule") or ""
+            elif str(args.get("repeat") or "").strip().lower() in {"none", "no", "off", "false", "single"}:
+                ev.rrule = ""
             is_caldav = ev.calendar and ev.calendar.source == "caldav"
             if is_caldav:
                 ev.caldav_sync_pending = "update"

src/tools/contacts.py

@@ -108,16 +108,28 @@ async def do_manage_contact(content: str, owner: Optional[str] = None) -> Dict:
 
         if action == "add":
             email = (args.get("email") or "").strip()
-            if not email:
-                return {"error": "email is required for add", "exit_code": 1}
-            name = (args.get("name") or "").strip() or email.split("@")[0]
-            # Dedupe by email (same as the /add route).
+            phones = [str(p or "").strip() for p in (args.get("phones") or []) if str(p or "").strip()]
+            phone = (args.get("phone") or "").strip()
+            if phone and phone not in phones:
+                phones.insert(0, phone)
+            address = (args.get("address") or "").strip()
+            name = (args.get("name") or "").strip()
+            if not name and email:
+                name = email.split("@")[0]
+            if not name and not email and not phones and not address:
+                return {"error": "name plus email, phone, or address is required for add", "exit_code": 1}
+            if not name:
+                name = email.split("@")[0] if email else (phones[0] if phones else "Contact")
+            # Dedupe by email or phone (same as the /add route).
             existing = await asyncio.to_thread(cc._fetch_contacts)
             for c in existing:
-                if email.lower() in [e.lower() for e in c.get("emails", [])]:
+                if email and email.lower() in [e.lower() for e in c.get("emails", [])]:
                     return {"output": f"{email} is already a contact ({c.get('name','')}).", "exit_code": 0}
-            ok = await asyncio.to_thread(cc._create_contact, name, email)
-            return {"output": f"{'Added' if ok else 'Failed to add'} {name} <{email}>.", "exit_code": 0 if ok else 1}
+                if phones and any(p in (c.get("phones") or []) for p in phones):
+                    return {"output": f"{phones[0]} is already a contact ({c.get('name','')}).", "exit_code": 0}
+            ok = await asyncio.to_thread(cc._create_contact, name, email, address, phones)
+            detail = email or ", ".join(phones) or address
+            return {"output": f"{'Added' if ok else 'Failed to add'} {name} ({detail}).", "exit_code": 0 if ok else 1}
 
         if action in ("update", "edit"):
             uid = (args.get("uid") or "").strip()
@@ -129,11 +141,12 @@ async def do_manage_contact(content: str, owner: Optional[str] = None) -> Dict:
                 emails = [args["email"]]
             emails = [e.strip() for e in (emails or []) if e and e.strip()]
             phones = [p.strip() for p in (args.get("phones") or []) if p and p.strip()]
-            if not name and not emails:
-                return {"error": "Provide a name or emails to update", "exit_code": 1}
+            address = (args.get("address") or "").strip()
+            if not name and not emails and not phones and not address:
+                return {"error": "Provide a name, emails, phones, or address to update", "exit_code": 1}
             if not name and emails:
                 name = emails[0].split("@")[0]
-            ok = await asyncio.to_thread(cc._update_contact, uid, name, emails, phones)
+            ok = await asyncio.to_thread(cc._update_contact, uid, name, emails, phones, address)
             return {"output": "Contact updated." if ok else "Update failed.", "exit_code": 0 if ok else 1}
 
         if action == "delete":

src/tools/cookbook.py

@@ -315,6 +315,7 @@ async def _cookbook_register_task(
 _MODEL_PROCESS_PATTERNS = [
     ("vLLM",            ["vllm.entrypoints", "vllm serve", "/vllm/", "vllm-openai"]),
     ("SGLang",          ["sglang.launch_server", "sglang/launch_server"]),
+    ("MLX",             ["mlx_lm.server", "mlx-lm"]),
     ("llama.cpp",       ["llama-server", "llama_cpp_server", "llamacppserver"]),
     ("Ollama",          ["ollama serve", "ollama runner", "/ollama "]),
     ("ComfyUI",         ["comfyui/main.py", "/ComfyUI/main.py", "ComfyUI"]),
@@ -590,7 +591,7 @@ async def do_serve_model(content: str, owner: Optional[str] = None) -> Dict:
         hint = ""
         if isinstance(err_msg, str) and "cmd" in err_msg.lower():
             hint = (" — the cmd must START with an allowlisted binary "
-                    "(vllm, python3, llama-server, ollama, sglang, lmdeploy, node, npx). "
+                    "(vllm, python3, llama-server, ollama, sglang, mlx_lm, lmdeploy, node, npx). "
                     "Do NOT prefix with `cd …`, `source …`, or chain with `&&`. "
                     "env_prefix (e.g. `source ~/qwen35-env/bin/activate`) is added "
                     "automatically from the host's saved venv settings.")
@@ -635,7 +636,7 @@ async def do_list_served_models(content: str, owner: Optional[str] = None) -> Di
 
     if not merged:
         return {
-            "output": "No model servers currently running (cookbook task tracker empty; /proc scan found no vLLM / sglang / llama.cpp / Ollama / ComfyUI / A1111 / Fooocus / InvokeAI / TGI / Aphrodite / Triton / Diffusers processes).",
+            "output": "No model servers currently running (cookbook task tracker empty; /proc scan found no vLLM / sglang / MLX / llama.cpp / Ollama / ComfyUI / A1111 / Fooocus / InvokeAI / TGI / Aphrodite / Triton / Diffusers processes).",
             "exit_code": 0,
         }
 

src/tools/notes.py

@@ -27,7 +27,8 @@ async def do_manage_notes(content: str, owner: Optional[str] = None) -> Dict:
 
     # Action aliases — match what models actually emit. `create` is the most
     # common alternative to `add`. Hyphenated forms also accepted.
-    action = (args.get("action") or "").replace("-", "_").strip().lower()
+    raw_action = (args.get("action") or "").replace("-", "_").strip().lower()
+    action = raw_action
     _NOTE_ACTION_ALIASES = {
         "create": "add",
         "new": "add",
@@ -60,37 +61,68 @@ async def do_manage_notes(content: str, owner: Optional[str] = None) -> Dict:
             q = q.filter(Note.owner == owner)
         return q.first()
 
+    def _format_note_list(notes) -> str:
+        lines = []
+        for n in notes:
+            pin = " [PINNED]" if n.pinned else ""
+            typ = " [checklist]" if n.note_type == "checklist" else ""
+            lbl = f" #{n.label}" if n.label else ""
+            title = n.title or "(untitled)"
+            lines.append(f"- [{n.id[:8]}] **{title}**{pin}{typ}{lbl}")
+            if n.note_type == "checklist" and n.items:
+                try:
+                    items = json.loads(n.items)
+                    for i, item in enumerate(items):
+                        mark = "x" if item.get("done") else " "
+                        lines.append(f"  [{mark}] {i}: {item.get('text', '')}")
+                except (json.JSONDecodeError, TypeError):
+                    pass
+            elif n.content:
+                snippet = n.content[:80].replace("\n", " ")
+                lines.append(f"  {snippet}")
+        return "\n".join(lines)
+
     try:
-        if action == "list":
+        if action in ("list", "search", "find"):
             q = db.query(Note)
             if owner is not None:
                 q = q.filter(Note.owner == owner)
-            if args.get("label"):
-                q = q.filter(Note.label == args["label"])
+            label_filter = str(args.get("label") or "").strip()
+            if label_filter and label_filter.lower() != "default":
+                q = q.filter(Note.label == label_filter)
             show_archived = args.get("archived", False)
             q = q.filter(Note.archived == show_archived)
             notes = q.order_by(Note.pinned.desc(), Note.updated_at.desc()).all()
+            if action in ("search", "find"):
+                query = str(
+                    args.get("query")
+                    or args.get("text")
+                    or args.get("title")
+                    or args.get("content")
+                    or ""
+                ).strip().lower()
+                if query:
+                    filtered = []
+                    for n in notes:
+                        haystack = " ".join(
+                            str(part or "")
+                            for part in (n.title, n.content, n.label, n.items)
+                        ).lower()
+                        if query in haystack:
+                            filtered.append(n)
+                    notes = filtered
             if not notes:
                 return {"response": "No notes found.", "exit_code": 0}
-            lines = []
-            for n in notes:
-                pin = " [PINNED]" if n.pinned else ""
-                typ = " [checklist]" if n.note_type == "checklist" else ""
-                lbl = f" #{n.label}" if n.label else ""
-                title = n.title or "(untitled)"
-                lines.append(f"- [{n.id[:8]}] **{title}**{pin}{typ}{lbl}")
-                if n.note_type == "checklist" and n.items:
-                    try:
-                        items = json.loads(n.items)
-                        for i, item in enumerate(items):
-                            mark = "x" if item.get("done") else " "
-                            lines.append(f"  [{mark}] {i}: {item.get('text', '')}")
-                    except (json.JSONDecodeError, TypeError):
-                        pass
-                elif n.content:
-                    snippet = n.content[:80].replace("\n", " ")
-                    lines.append(f"  {snippet}")
-            return {"results": "\n".join(lines)}
+            return {"results": _format_note_list(notes), "exit_code": 0}
+
+        elif action == "view":
+            note_id = args.get("id", "")
+            note = _note_by_prefix(note_id)
+            if not note:
+                return {"error": f"Note '{note_id}' not found", "exit_code": 1}
+            if not _note_visible_to_owner(note, owner):
+                return {"error": "Note not found", "exit_code": 1}
+            return {"results": _format_note_list([note]), "exit_code": 0}
 
         elif action == "add":
             # Accept the various field names models emit: `text` is the most
@@ -120,6 +152,25 @@ async def do_manage_notes(content: str, owner: Optional[str] = None) -> Dict:
             # `new Date()` resolves the right absolute moment regardless of
             # where the user is.
             due_raw = args.get("due_date")
+            if not due_raw:
+                combined_text = " ".join(
+                    str(v or "")
+                    for v in (title, content_raw, text_raw)
+                ).strip()
+                lower_combined = combined_text.lower()
+                looks_like_reminder = (
+                    raw_action in {"remind", "reminder"}
+                    or re.search(r"\bremind(?:er)?\b", lower_combined)
+                )
+                if looks_like_reminder:
+                    temporal = re.search(
+                        r"\b(?:today|tonight|tomorrow|tmrw|yesterday)\b(?:\s+(?:at\s+)?\d{1,2}(?::\d{2})?\s*(?:am|pm)?)?"
+                        r"|\b\d{1,2}(?::\d{2})?\s*(?:am|pm)?\s+(?:today|tonight|tomorrow|tmrw|yesterday)\b"
+                        r"|\bin\s+\d+\s*(?:hour|hr|minute|min|day)s?\b",
+                        lower_combined,
+                    )
+                    if temporal:
+                        due_raw = temporal.group(0)
             due_iso = None
             if due_raw:
                 try:
@@ -170,7 +221,7 @@ async def do_manage_notes(content: str, owner: Optional[str] = None) -> Dict:
             # link with no target, leaving the user with a click that
             # did nothing and uncertainty about whether the note was made.
             return {
-                "response": f"Note created: \"{title or '(untitled)'}\" (id: {note.id[:8]})",
+                "response": f"{'Reminder' if due_iso else 'Note'} created: \"{title or '(untitled)'}\" (id: {note.id[:8]})",
                 "note_id": note.id,
                 "note_title": title or "",
                 "open_url": f"/#open=notes&note={note.id}",
@@ -246,7 +297,7 @@ async def do_manage_notes(content: str, owner: Optional[str] = None) -> Dict:
             return {"response": f"Item '{items[index].get('text', '')}' marked {mark}", "exit_code": 0}
 
         else:
-            return {"error": f"Unknown action: {action}. Use list/add/update/delete/toggle_item", "exit_code": 1}
+            return {"error": f"Unknown action: {action}. Use list/search/view/add/update/delete/toggle_item", "exit_code": 1}
     except Exception as e:
         logger.error(f"manage_notes error: {e}")
         return {"error": str(e), "exit_code": 1}