fix(stabilization): harden attachment lifecycle and agent guard signals (#5420)

* fix: harden stabilization attachment and agent guards

* fix(uploads): preserve durable references during cleanup

* fix(uploads): close cleanup and compaction races

src/agent_loop.py

@@ -3828,9 +3828,8 @@ async def stream_agent_loop(
                 and _intent_match is not None
                 and len(_intent_text) < 400
                 and "```" not in _intent_text
-                and _intent_nudge_count < _MAX_INTENT_NUDGES
             )
-            if _looks_like_promise:
+            if _looks_like_promise and _intent_nudge_count < _MAX_INTENT_NUDGES:
                 _intent_nudge_count += 1
                 _matched_phrase = _intent_match.group(0).strip()
                 logger.info(f"[agent] intent-without-action nudge #{_intent_nudge_count} on round {round_num}: {_matched_phrase!r}")
@@ -3859,6 +3858,31 @@ async def stream_agent_loop(
                 # Visible signal in the stream so the user knows we caught it.
                 yield f'data: {json.dumps({"type": "agent_step", "round": round_num + 1})}\n\n'
                 continue
+            if _looks_like_promise:
+                _matched_phrase = _intent_match.group(0).strip()
+                _guard_message = (
+                    "The agent stopped because it repeatedly announced a tool "
+                    "action without making the tool call."
+                )
+                logger.warning(
+                    "[agent] intent-without-action guard exhausted on round %d after %d nudges: %r",
+                    round_num,
+                    _intent_nudge_count,
+                    _matched_phrase,
+                )
+                yield (
+                    "data: "
+                    + json.dumps({
+                        "type": "intent_nudge_exhausted",
+                        "reason": "intent_without_action_nudge_cap",
+                        "message": _guard_message,
+                        "round": round_num,
+                        "nudges": _intent_nudge_count,
+                        "matched": _matched_phrase,
+                    })
+                    + "\n\n"
+                )
+                break
             break  # no tools — done
 
         # ── Loop-breaker (Terminus-style stall detector) ──────────────
@@ -3895,6 +3919,21 @@ async def stream_agent_loop(
             reason = (f"calling {_runaway} with identical arguments over and over" if _runaway
                       else "repeating the same tool calls without new progress")
             logger.warning(f"[agent] loop-breaker tripped on round {round_num} ({reason}); sig={_sig[:80]!r}")
+            yield (
+                "data: "
+                    + json.dumps({
+                    "type": "loop_breaker_triggered",
+                    "reason": "loop_breaker_stall",
+                    "message": (
+                        "The loop-breaker detected repeated tool calls without "
+                        "new progress, so the agent is being forced to stop "
+                        "using tools and give its best final answer."
+                    ),
+                    "round": round_num,
+                    "detail": reason,
+                })
+                + "\n\n"
+            )
             # The model has been executing tools, so its results are already
             # in context. Force ONE tool-free round to converge: write the
             # answer from what it has, or state plainly what's blocking it.