6 Commits

Author SHA1 Message Date
Boody d88c8cbacf Merge pull request #5313 from RaresKeY/fix/chat-web-search-explicit-deny
fix(chat): require explicit web search enable
2026-07-09 01:56:32 +03:00
Wes Huber a35384e68f fix(copilot): guard request_flags against a non-dict last message (#5274)
request_flags derives (agent, vision) and does last.get("role") after only
a truthy check. A client can send a bare-string message element
("messages": ["hi"]), and the vision loop right below already guards each
element with isinstance — so the .get() on a non-dict last element is an
oversight that raises AttributeError on every Copilot-proxied request with
such a body.

Use isinstance(last, dict) to match the loop's own guard.

Fixes #5273

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-08 23:57:23 +02:00
Miraç Duran 3064819e3d fix(chat): give extensionless image/audio uploads a valid MIME subtype (#5205)
build_user_content derived the data-URL subtype from the file extension
only (image_format = ext[1:]). An extensionless upload (e.g. a pasted
screenshot) has ext == "", producing "data:image/;base64,..." with an
empty subtype (invalid per RFC 2046) that vision/audio endpoints reject,
silently dropping the attachment. Fall back to the resolved MIME subtype
when the extension is missing; present extensions are unchanged.
2026-07-08 21:04:15 +02:00
RaresKeY 54f1d015b5 fix(chat): require explicit web search enable 2026-07-08 17:44:28 +00:00
Steve Holloway 2d8177035b fix(chat): restore missing _explicit_web_intent definition (#5290)
chat_stream() references `_explicit_web_intent` in three places
(disabled-tools gating, global-disabled web allowance, and the
per-turn tool filter) but the assignment was dropped during a
branch merge. Every chat request raised

    NameError: name '_explicit_web_intent' is not defined

at routes/chat_routes.py, surfacing to the client as a bare
"Internal Server Error" before any LLM call was made — chat was
fully broken on dev and main.

Restore the original definition, computed from the already-derived
tool intent, immediately before its first use:

    _explicit_web_intent = bool(_tool_intent and _tool_intent.category == "web")

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-08 19:14:43 +02:00
PewDiePie c67deaa60a Merge pull request #5283 from pewdiepie-archdaemon/sync-main-into-dev-20260707
chore: sync tested main into dev
2026-07-07 11:32:53 +09:00
9 changed files with 362 additions and 72 deletions
+24 -33
View File
@@ -42,7 +42,12 @@ from routes.chat_helpers import (
_enforce_chat_privileges, _enforce_chat_privileges,
) )
from src.action_intents import ToolIntent, 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 from src.tool_policy import (
WEB_TOOL_NAMES,
build_effective_tool_policy,
is_web_search_explicitly_denied,
web_search_enabled_for_turn,
)
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -583,10 +588,7 @@ def setup_chat_routes(
# below). Skill extraction should only learn from real agent sessions, # below). Skill extraction should only learn from real agent sessions,
# not chats we quietly promoted for a notes/calendar intent. # not chats we quietly promoted for a notes/calendar intent.
user_requested_agent = (chat_mode == "agent") user_requested_agent = (chat_mode == "agent")
_search_enabled = ( _search_enabled = web_search_enabled_for_turn(allow_web_search, use_web)
str(allow_web_search).lower() == "true"
or str(use_web).lower() == "true"
)
# Intent auto-escalation: if the user is clearly asking the assistant # Intent auto-escalation: if the user is clearly asking the assistant
# to create a todo, reminder, or calendar event, promote chat → agent # to create a todo, reminder, or calendar event, promote chat → agent
# for this turn so the LLM has access to manage_notes / manage_calendar. # for this turn so the LLM has access to manage_notes / manage_calendar.
@@ -870,23 +872,20 @@ def setup_chat_routes(
# Build disabled-tools set from frontend toggles + user privileges # Build disabled-tools set from frontend toggles + user privileges
disabled_tools = set() disabled_tools = set()
# Only disable bash/web_search when the caller *explicitly* set them # Only disable bash when the caller *explicitly* set it to a falsy
# to a falsy value. When unset (None), defer to per-user privilege # value. When unset (None), defer to per-user privilege checks below.
# checks below — this lets admins with can_use_bash=True use bash # Web search is per-turn opt-in: either the chat pre-search setting
# by default without having to send allow_bash in every request. # (`use_web=true`) or agent web toggle (`allow_web_search=true`) must
# explicitly enable it.
if allow_bash is not None and str(allow_bash).lower() != "true": if allow_bash is not None and str(allow_bash).lower() != "true":
disabled_tools.add("bash") disabled_tools.add("bash")
if ( _explicit_web_intent = bool(_tool_intent and _tool_intent.category == "web")
allow_web_search is not None if is_web_search_explicitly_denied(allow_web_search) or not _search_enabled:
and str(allow_web_search).lower() != "true" disabled_tools.update(WEB_TOOL_NAMES)
):
disabled_tools.add("web_search")
disabled_tools.add("web_fetch")
if _explicit_web_intent: if _explicit_web_intent:
# A direct lookup/search request should not drift into personal # A direct lookup/search request should not drift into personal
# tools or shell fallbacks. We still keep web_search/web_fetch # tools or shell fallbacks. It can only use web_search/web_fetch
# available even when the frontend toggle is stale/falsy because # when the request's explicit web setting enabled them.
# the user's words are the stronger signal.
disabled_tools.update({ disabled_tools.update({
"bash", "python", "bash", "python",
"search_chats", "manage_skills", "manage_memory", "search_chats", "manage_skills", "manage_memory",
@@ -896,11 +895,12 @@ def setup_chat_routes(
"manage_notes", "manage_calendar", "manage_tasks", "manage_notes", "manage_calendar", "manage_tasks",
"api_call", "builtin_browser", "api_call", "builtin_browser",
}) })
disabled_tools.discard("web_search") if _search_enabled:
disabled_tools.discard("web_fetch") disabled_tools.difference_update(WEB_TOOL_NAMES)
else:
disabled_tools.update(WEB_TOOL_NAMES)
elif _search_enabled: elif _search_enabled:
disabled_tools.discard("web_search") disabled_tools.difference_update(WEB_TOOL_NAMES)
disabled_tools.discard("web_fetch")
# Nobody/incognito mode: deny tools that would expose the user's # Nobody/incognito mode: deny tools that would expose the user's
# persistent memory, past chats, or other identity-linked data. # persistent memory, past chats, or other identity-linked data.
@@ -951,13 +951,6 @@ def setup_chat_routes(
from src.settings import get_setting from src.settings import get_setting
_global_disabled = get_setting("disabled_tools", []) _global_disabled = get_setting("disabled_tools", [])
if _global_disabled and isinstance(_global_disabled, list): if _global_disabled and isinstance(_global_disabled, list):
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:
disabled_tools.update(_global_disabled) disabled_tools.update(_global_disabled)
# Light auto-escalation: the user is in chat mode and just expressed a # Light auto-escalation: the user is in chat mode and just expressed a
@@ -1410,10 +1403,8 @@ def setup_chat_routes(
_max_rounds = max(1, min(_max_rounds, 200)) _max_rounds = max(1, min(_max_rounds, 200))
_forced_tools = None _forced_tools = None
if _explicit_web_intent: if _search_enabled:
_forced_tools = {"web_search", "web_fetch"} _forced_tools = set(WEB_TOOL_NAMES)
elif _search_enabled:
_forced_tools = {"web_search", "web_fetch"}
async for chunk in stream_agent_loop( async for chunk in stream_agent_loop(
sess.endpoint_url, sess.endpoint_url,
+10 -11
View File
@@ -24,7 +24,7 @@ from src.model_context import estimate_tokens
from src.settings import get_setting from src.settings import get_setting
from src.prompt_security import untrusted_context_message from src.prompt_security import untrusted_context_message
from src.tool_security import blocked_tools_for_owner, plan_mode_disabled_tools from src.tool_security import blocked_tools_for_owner, plan_mode_disabled_tools
from src.tool_policy import GUIDE_ONLY_DIRECTIVE, ToolPolicy from src.tool_policy import GUIDE_ONLY_DIRECTIVE, WEB_TOOL_NAMES, ToolPolicy
from src.tool_utils import _truncate, get_mcp_manager from src.tool_utils import _truncate, get_mcp_manager
from src.agent_tools import ( from src.agent_tools import (
parse_tool_blocks, parse_tool_blocks,
@@ -321,7 +321,7 @@ _DOMAIN_RULES = {
} }
_DOMAIN_TOOL_MAP = { _DOMAIN_TOOL_MAP = {
"web": {"web_search", "web_fetch"}, "web": set(WEB_TOOL_NAMES),
"documents": {"create_document", "edit_document", "update_document", "suggest_document", "manage_documents"}, "documents": {"create_document", "edit_document", "update_document", "suggest_document", "manage_documents"},
"email": {"list_email_accounts", "list_emails", "read_email", "send_email", "reply_to_email", "bulk_email", "archive_email", "delete_email", "mark_email_read", "resolve_contact", "manage_contact"}, "email": {"list_email_accounts", "list_emails", "read_email", "send_email", "reply_to_email", "bulk_email", "archive_email", "delete_email", "mark_email_read", "resolve_contact", "manage_contact"},
"cookbook": {"download_model", "serve_model", "serve_preset", "list_serve_presets", "list_served_models", "stop_served_model", "tail_serve_output", "list_downloads", "cancel_download", "search_hf_models", "list_cached_models", "list_cookbook_servers", "adopt_served_model"}, "cookbook": {"download_model", "serve_model", "serve_preset", "list_serve_presets", "list_served_models", "stop_served_model", "tail_serve_output", "list_downloads", "cancel_download", "search_hf_models", "list_cached_models", "list_cookbook_servers", "adopt_served_model"},
@@ -2847,13 +2847,12 @@ async def stream_agent_loop(
if "email" in (_intent.get("domains") or set()): if "email" in (_intent.get("domains") or set()):
_relevant_tools.add("ui_control") _relevant_tools.add("ui_control")
if "web" in (_intent.get("domains") or set()): if "web" in (_intent.get("domains") or set()):
_relevant_tools.update({"web_search", "web_fetch"}) _relevant_tools.update(WEB_TOOL_NAMES)
_removed_web_blocks = sorted({"web_search", "web_fetch"} & disabled_tools) _blocked_web_tools = sorted(WEB_TOOL_NAMES & disabled_tools)
if _removed_web_blocks: if _blocked_web_tools:
disabled_tools.difference_update({"web_search", "web_fetch"})
logger.info( logger.info(
"[agent-intent] web turn forced search tools enabled; removed disabled=%s", "[agent-intent] web domain selected but search tools remain disabled=%s",
_removed_web_blocks, _blocked_web_tools,
) )
if "ui" in (_intent.get("domains") or set()): if "ui" in (_intent.get("domains") or set()):
_relevant_tools.add("ui_control") _relevant_tools.add("ui_control")
@@ -2887,9 +2886,9 @@ async def stream_agent_loop(
_relevant_tools = set(ALWAYS_AVAILABLE) _relevant_tools = set(ALWAYS_AVAILABLE)
_relevant_tools.update({"read_file", "grep", "ls", "manage_documents"}) _relevant_tools.update({"read_file", "grep", "ls", "manage_documents"})
# Per-request forced tools are stronger than retrieval. Search toggles and # Per-request forced tools are stronger than retrieval. Explicit search
# explicit lookup turns must make web tools visible even when tool RAG # settings make web tools visible even when tool RAG misses them;
# misses them; route-level disabled_tools decides what else is allowed. # route-level disabled_tools decides what remains allowed.
if not guide_only and forced_tools: if not guide_only and forced_tools:
forced_set = {t for t in forced_tools if t not in disabled_tools} forced_set = {t for t in forced_tools if t not in disabled_tools}
if _relevant_tools is None: if _relevant_tools is None:
+4 -1
View File
@@ -230,7 +230,10 @@ def request_flags(messages) -> tuple:
""" """
msgs = messages or [] msgs = messages or []
last = msgs[-1] if msgs else None last = msgs[-1] if msgs else None
agent = bool(last) and last.get("role") != "user" # A message element can be a non-dict (clients send `"messages": ["hi"]`);
# the vision loop below already guards each element with isinstance, so do
# the same here rather than call .get() on a bare string.
agent = isinstance(last, dict) and last.get("role") != "user"
vision = False vision = False
for m in msgs: for m in msgs:
content = m.get("content") if isinstance(m, dict) else None content = m.get("content") if isinstance(m, dict) else None
+5 -2
View File
@@ -440,7 +440,10 @@ def build_user_content(
try: try:
with open(path, "rb") as image_file: with open(path, "rb") as image_file:
encoded_string = base64.b64encode(image_file.read()).decode("utf-8") encoded_string = base64.b64encode(image_file.read()).decode("utf-8")
image_format = ext[1:] # Extensionless uploads (e.g. a pasted screenshot) have no ext,
# so fall back to the resolved MIME subtype rather than emitting
# an invalid "data:image/;base64," with an empty subtype.
image_format = ext[1:] or (mime.split("/", 1)[1] if mime.startswith("image/") else "png")
content.append({ content.append({
"type": "image_url", "type": "image_url",
"image_url": {"url": f"data:image/{image_format};base64,{encoded_string}"}, "image_url": {"url": f"data:image/{image_format};base64,{encoded_string}"},
@@ -456,7 +459,7 @@ def build_user_content(
try: try:
with open(path, "rb") as audio_file: with open(path, "rb") as audio_file:
encoded_string = base64.b64encode(audio_file.read()).decode("utf-8") encoded_string = base64.b64encode(audio_file.read()).decode("utf-8")
audio_format = ext[1:] audio_format = ext[1:] or (mime.split("/", 1)[1] if mime.startswith("audio/") else "mpeg")
content.append({ content.append({
"type": "audio", "type": "audio",
"audio": {"url": f"data:audio/{audio_format};base64,{encoded_string}"}, "audio": {"url": f"data:audio/{audio_format};base64,{encoded_string}"},
+33
View File
@@ -16,6 +16,39 @@ GUIDE_ONLY_DIRECTIVE = (
"output they will produce locally." "output they will produce locally."
) )
WEB_TOOL_NAMES = frozenset({"web_search", "web_fetch"})
def tool_toggle_enabled(value: object) -> bool:
"""Return true only for explicit true-like tool toggle values."""
return str(value).lower() == "true"
def tool_toggle_explicitly_denied(value: object) -> bool:
"""Return true when a caller explicitly supplied a non-true toggle value."""
return value is not None and not tool_toggle_enabled(value)
def is_web_search_explicitly_denied(allow_web_search: object) -> bool:
"""Whether the web-search agent toggle was explicitly set to false."""
return tool_toggle_explicitly_denied(allow_web_search)
def web_search_enabled_for_turn(allow_web_search: object, use_web: object = None) -> bool:
"""Return true only when this request explicitly enables web search.
Agent mode sends ``allow_web_search``; chat-mode pre-search sends
``use_web``. If both are present, an explicit ``allow_web_search=false``
wins so a stale or conflicting intent path cannot re-enable web tools.
"""
if is_web_search_explicitly_denied(allow_web_search):
return False
return tool_toggle_enabled(allow_web_search) or tool_toggle_enabled(use_web)
_COMMON_TOOL_NAMES = { _COMMON_TOOL_NAMES = {
"api_call", "api_call",
+83 -23
View File
@@ -1,12 +1,11 @@
"""Issue #3229 — allow_bash / allow_web_search must work for JSON API callers """Issue #3229 and explicit web-toggle regressions.
and admin users must get bash enabled by default.
Bug: allow_bash and allow_web_search were only read from form_data, so JSON Bug: allow_bash and allow_web_search were only read from form_data, so JSON
API callers (Content-Type: application/json) always had bash disabled. API callers (Content-Type: application/json) always had bash disabled.
Fix: (1) Read from JSON body as fallback. Fix: (1) Read from JSON body as fallback.
(2) Only add bash/web_search to disabled_tools when explicitly set to a (2) Keep bash on the privilege fallback when unset.
falsy value; when unset (None), defer to per-user privilege checks. (3) Require an explicit per-turn web setting before exposing web tools.
""" """
import ast import ast
@@ -15,6 +14,11 @@ from pathlib import Path
import pytest import pytest
from src.action_intents import classify_tool_intent from src.action_intents import classify_tool_intent
from src.tool_policy import (
WEB_TOOL_NAMES,
is_web_search_explicitly_denied,
web_search_enabled_for_turn,
)
_CHAT_ROUTES = Path(__file__).resolve().parent.parent / "routes" / "chat_routes.py" _CHAT_ROUTES = Path(__file__).resolve().parent.parent / "routes" / "chat_routes.py"
@@ -76,8 +80,7 @@ def test_allow_web_search_reads_from_body_as_fallback():
def test_disabled_tools_respects_missing_vs_explicit_toggles(): def test_disabled_tools_respects_missing_vs_explicit_toggles():
"""When allow_bash is not set (None), bash must NOT be unconditionally """Bash still defers to privileges, but web is an explicit per-turn opt-in.
added to disabled_tools. The per-user privilege check handles it.
""" """
source = _CHAT_ROUTES.read_text(encoding="utf-8") source = _CHAT_ROUTES.read_text(encoding="utf-8")
@@ -88,11 +91,14 @@ def test_disabled_tools_respects_missing_vs_explicit_toggles():
assert "allow_bash is not None" in source, ( assert "allow_bash is not None" in source, (
"disabled_tools check must guard against allow_bash being None" "disabled_tools check must guard against allow_bash being None"
) )
assert "allow_web_search is not None" in source, ( assert "web_search_enabled_for_turn(allow_web_search, use_web)" in source, (
"disabled_tools check must guard against allow_web_search being None" "web tools must be gated through the explicit per-turn web setting"
) )
assert "and not _explicit_web_intent" not in source, ( assert "disabled_tools.update(WEB_TOOL_NAMES)" in source, (
"explicit allow_web_search=false must not be overridden by prompt web intent" "disabled_tools must add web_search/web_fetch when web is not explicitly enabled"
)
assert "_forced_tools = set(WEB_TOOL_NAMES)" in source, (
"web tools should only be forced visible from the explicit web setting"
) )
@@ -102,9 +108,11 @@ def test_disabled_tools_respects_missing_vs_explicit_toggles():
def _build_disabled_tools( def _build_disabled_tools(
allow_bash=None, allow_bash=None,
allow_web_search=None, allow_web_search=None,
use_web=None,
can_use_bash=True, can_use_bash=True,
can_use_browser=True, can_use_browser=True,
explicit_web_intent=False, explicit_web_intent=False,
global_disabled=None,
): ):
"""Replicate the disabled-tools logic from chat_stream for unit testing. """Replicate the disabled-tools logic from chat_stream for unit testing.
@@ -112,21 +120,36 @@ def _build_disabled_tools(
""" """
disabled_tools = set() disabled_tools = set()
# Issue #3229 fix: only disable when explicitly set to a falsy value. # Issue #3229 fix: only disable bash when explicitly set to a falsy value.
if allow_bash is not None and str(allow_bash).lower() != "true": if allow_bash is not None and str(allow_bash).lower() != "true":
disabled_tools.add("bash") disabled_tools.add("bash")
if ( search_enabled = web_search_enabled_for_turn(allow_web_search, use_web)
allow_web_search is not None if is_web_search_explicitly_denied(allow_web_search) or not search_enabled:
and str(allow_web_search).lower() != "true" disabled_tools.update(WEB_TOOL_NAMES)
): if explicit_web_intent:
disabled_tools.add("web_search") disabled_tools.update({
disabled_tools.add("web_fetch") "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",
})
if search_enabled:
disabled_tools.difference_update(WEB_TOOL_NAMES)
else:
disabled_tools.update(WEB_TOOL_NAMES)
elif search_enabled:
disabled_tools.difference_update(WEB_TOOL_NAMES)
# Enforce per-user privileges # Enforce per-user privileges
if not can_use_bash: if not can_use_bash:
disabled_tools.update({"bash", "python", "read_file", "write_file"}) disabled_tools.update({"bash", "python", "read_file", "write_file"})
if not can_use_browser: if not can_use_browser:
disabled_tools.add("builtin_browser") disabled_tools.add("builtin_browser")
if global_disabled and isinstance(global_disabled, list):
disabled_tools.update(global_disabled)
return disabled_tools return disabled_tools
@@ -157,6 +180,20 @@ def test_json_body_allow_web_search_false_disables_web():
assert "web_fetch" in disabled assert "web_fetch" in disabled
def test_chat_mode_use_web_true_enables_web():
"""Chat pre-search sends use_web=true as the explicit web setting."""
disabled = _build_disabled_tools(use_web="true")
assert "web_search" not in disabled
assert "web_fetch" not in disabled
def test_allow_web_search_false_wins_over_use_web_true():
"""The agent web toggle hard-denies web even if another path says use_web=true."""
disabled = _build_disabled_tools(allow_web_search="false", use_web="true")
assert "web_search" in disabled
assert "web_fetch" in disabled
@pytest.mark.parametrize( @pytest.mark.parametrize(
"message", "message",
[ [
@@ -180,6 +217,21 @@ def test_explicit_false_disables_web_despite_prompt_web_intent(message):
assert "web_fetch" in disabled assert "web_fetch" in disabled
def test_prompt_web_intent_does_not_enable_web_without_setting():
"""Prompt-derived web intent alone must not expose web tools."""
intent = classify_tool_intent("look up the latest docs")
assert intent is not None
assert intent.category == "web"
disabled = _build_disabled_tools(
allow_web_search=None,
use_web=None,
explicit_web_intent=True,
)
assert "web_search" in disabled
assert "web_fetch" in disabled
def test_admin_user_gets_bash_enabled_by_default(): def test_admin_user_gets_bash_enabled_by_default():
"""When allow_bash is not set and user has can_use_bash privilege, """When allow_bash is not set and user has can_use_bash privilege,
bash must NOT be disabled. bash must NOT be disabled.
@@ -188,13 +240,11 @@ def test_admin_user_gets_bash_enabled_by_default():
assert "bash" not in disabled assert "bash" not in disabled
def test_admin_user_gets_web_search_enabled_by_default(): def test_web_search_disabled_by_default_without_explicit_turn_setting():
"""When allow_web_search is not set and user has normal privileges, """Missing web settings must not expose web tools by default."""
web_search must NOT be disabled.
"""
disabled = _build_disabled_tools(allow_web_search=None) disabled = _build_disabled_tools(allow_web_search=None)
assert "web_search" not in disabled assert "web_search" in disabled
assert "web_fetch" not in disabled assert "web_fetch" in disabled
def test_non_privileged_user_without_explicit_flag_still_disabled(): def test_non_privileged_user_without_explicit_flag_still_disabled():
@@ -213,6 +263,16 @@ def test_non_privileged_user_explicit_true_overridden_by_privilege():
assert "bash" in disabled assert "bash" in disabled
def test_global_disabled_web_wins_over_explicit_web_enable():
"""Admin-level disabled tools are still a hard deny."""
disabled = _build_disabled_tools(
allow_web_search="true",
global_disabled=["web_search", "web_fetch"],
)
assert "web_search" in disabled
assert "web_fetch" in disabled
def test_form_data_none_body_true_works(): def test_form_data_none_body_true_works():
"""Simulates: form_data has no allow_bash, body has allow_bash=true. """Simulates: form_data has no allow_bash, body has allow_bash=true.
After the fallback (`form_data.get(...) or body.get(...)`), allow_bash After the fallback (`form_data.get(...) or body.get(...)`), allow_bash
+12
View File
@@ -89,6 +89,18 @@ def test_request_flags_vision():
assert vision is True assert vision is True
def test_request_flags_non_dict_last_message_does_not_crash():
# A client can send a bare-string (non-dict) last element; before the
# isinstance guard this raised AttributeError on last.get("role").
assert copilot.request_flags(["hi"]) == (False, False)
assert copilot.request_flags([{"role": "user"}, "trailing"]) == (False, False)
def test_request_flags_empty_and_none():
assert copilot.request_flags([]) == (False, False)
assert copilot.request_flags(None) == (False, False)
def test_apply_request_headers_mutates(): def test_apply_request_headers_mutates():
h = {"X-GitHub-Api-Version": "v"} h = {"X-GitHub-Api-Version": "v"}
copilot.apply_request_headers(h, [{"role": "tool", "content": "x"}]) copilot.apply_request_headers(h, [{"role": "tool", "content": "x"}])
@@ -0,0 +1,74 @@
"""Regression: extensionless image/audio uploads must get a valid MIME subtype.
The data-URL subtype was derived only from the stored file's extension
(`image_format = ext[1:]`). A pasted screenshot or any file whose stored id
carries no extension yields `ext == ""`, so the emitted URL was
`data:image/;base64,...` — an empty MIME subtype (invalid per RFC 2046) that
vision/audio endpoints reject, silently dropping the attachment. When the
extension is missing, fall back to the resolved MIME subtype. Extensions that
are present are unchanged.
"""
class _Handler:
def __init__(self, uploads, image=False, audio=False):
self.uploads = uploads
self._image = image
self._audio = audio
def resolve_upload(self, fid, owner=None):
return self.uploads.get(fid)
def _inside_upload_dir(self, path):
return True
def is_image_file(self, name, mime):
return self._image and (mime or "").startswith("image/")
def is_audio_file(self, name, mime):
return self._audio and (mime or "").startswith("audio/")
def is_document_file(self, name, mime):
return False
def _blocks(content, block_type):
return [b for b in content if isinstance(b, dict) and b.get("type") == block_type]
def test_extensionless_image_uses_mime_subtype(tmp_path):
import src.document_processor as dp
p = tmp_path / ("a" * 32) # bare id, no extension
p.write_bytes(b"\x89PNG\r\n\x1a\nfake")
uploads = {"img": {"path": str(p), "name": "screenshot", "mime": "image/png"}}
content = dp.build_user_content("look", ["img"], str(tmp_path), _Handler(uploads, image=True), owner="t")
imgs = _blocks(content, "image_url")
assert imgs, content
assert imgs[0]["image_url"]["url"].startswith("data:image/png;base64,")
def test_extensionless_audio_uses_mime_subtype(tmp_path):
import src.document_processor as dp
p = tmp_path / ("b" * 32)
p.write_bytes(b"fakeaudio")
uploads = {"aud": {"path": str(p), "name": "recording", "mime": "audio/mpeg"}}
content = dp.build_user_content("listen", ["aud"], str(tmp_path), _Handler(uploads, audio=True), owner="t")
auds = _blocks(content, "audio")
assert auds, content
assert auds[0]["audio"]["url"].startswith("data:audio/mpeg;base64,")
def test_extension_present_is_unchanged(tmp_path):
import src.document_processor as dp
p = tmp_path / "pic.png"
p.write_bytes(b"\x89PNG\r\n\x1a\n")
uploads = {"img": {"path": str(p), "name": "pic.png", "mime": "image/png"}}
content = dp.build_user_content("look", ["img"], str(tmp_path), _Handler(uploads, image=True), owner="t")
imgs = _blocks(content, "image_url")
assert imgs[0]["image_url"]["url"].startswith("data:image/png;base64,")
+116 -1
View File
@@ -6,7 +6,12 @@ from types import SimpleNamespace
import src.agent_loop as al import src.agent_loop as al
from src.agent_tools import ToolBlock from src.agent_tools import ToolBlock
from src.tool_execution import execute_tool_block from src.tool_execution import execute_tool_block
from src.tool_policy import build_effective_tool_policy, detect_guide_only_turn from src.tool_policy import (
WEB_TOOL_NAMES,
build_effective_tool_policy,
detect_guide_only_turn,
web_search_enabled_for_turn,
)
def _collect(gen): def _collect(gen):
@@ -76,6 +81,116 @@ def test_normal_policy_preserves_existing_disabled_tools():
assert not policy.blocks("bash") assert not policy.blocks("bash")
def test_web_search_enabled_for_turn_requires_explicit_enable():
assert web_search_enabled_for_turn(None, None) is False
assert web_search_enabled_for_turn("true", None) is True
assert web_search_enabled_for_turn(None, "true") is True
assert web_search_enabled_for_turn(True, None) is True
assert web_search_enabled_for_turn("false", "true") is False
assert web_search_enabled_for_turn(False, "true") is False
def _schema_names(tools):
return {
tool.get("function", {}).get("name") or tool.get("name")
for tool in (tools or [])
}
def test_agent_loop_web_intent_preserves_disabled_web_tools(monkeypatch):
_patch_loop_basics(monkeypatch)
sent_tools = []
async def _fake_stream(_candidates, messages, **kwargs):
sent_tools.append(kwargs.get("tools"))
yield _delta_chunk("ok")
yield "data: [DONE]\n\n"
monkeypatch.setattr(al, "stream_llm_with_fallback", _fake_stream, raising=False)
_collect(
al.stream_agent_loop(
"https://api.openai.com/v1",
"gpt-test",
[{"role": "user", "content": "please look up the latest CVEs"}],
max_rounds=1,
relevant_tools=set(),
disabled_tools=set(WEB_TOOL_NAMES),
)
)
assert sent_tools
assert WEB_TOOL_NAMES.isdisjoint(_schema_names(sent_tools[0]))
def test_agent_loop_forced_web_tools_filtered_by_disabled_tools(monkeypatch):
_patch_loop_basics(monkeypatch)
sent_tools = []
async def _fake_stream(_candidates, messages, **kwargs):
sent_tools.append(kwargs.get("tools"))
yield _delta_chunk("ok")
yield "data: [DONE]\n\n"
monkeypatch.setattr(al, "stream_llm_with_fallback", _fake_stream, raising=False)
_collect(
al.stream_agent_loop(
"https://api.openai.com/v1",
"gpt-test",
[{"role": "user", "content": "latest Kubernetes release"}],
max_rounds=1,
relevant_tools=set(),
forced_tools=set(WEB_TOOL_NAMES),
disabled_tools=set(WEB_TOOL_NAMES),
)
)
assert sent_tools
assert WEB_TOOL_NAMES.isdisjoint(_schema_names(sent_tools[0]))
def test_agent_loop_policy_blocks_disabled_web_tool_call_before_execution(monkeypatch):
_patch_loop_basics(monkeypatch)
called = False
async def _fake_exec(*args, **kwargs):
nonlocal called
called = True
return ("web_search", {"output": "ran", "exit_code": 0})
async def _fake_stream(_candidates, messages, **kwargs):
yield _delta_chunk('```web_search\n{"query":"current CVEs"}\n```')
yield "data: [DONE]\n\n"
monkeypatch.setattr(al, "execute_tool_block", _fake_exec, raising=False)
monkeypatch.setattr(al, "stream_llm_with_fallback", _fake_stream, raising=False)
policy = build_effective_tool_policy(
disabled_tools=WEB_TOOL_NAMES,
last_user_message="please look up the latest CVEs",
)
chunks = _collect(
al.stream_agent_loop(
"http://local.test/v1",
"local-model",
[{"role": "user", "content": "please look up the latest CVEs"}],
max_rounds=1,
relevant_tools={"web_search"},
disabled_tools=set(policy.all_disabled_names()),
tool_policy=policy,
)
)
events = _events(chunks)
blocked = [event for event in events if event.get("type") == "tool_output"]
assert called is False
assert not any(event.get("type") == "tool_start" for event in events)
assert blocked
assert blocked[0]["tool"] == "web_search"
assert blocked[0]["exit_code"] == 1
def test_executor_policy_backstop_blocks_tools(): def test_executor_policy_backstop_blocks_tools():
policy = build_effective_tool_policy(last_user_message="Do not use tools.") policy = build_effective_tool_policy(last_user_message="Do not use tools.")
desc, result = asyncio.run( desc, result = asyncio.run(