fix(chat,curator): unstick chat from silent generation crashes; curator only sees new messages

Two related reliability fixes.

1. routes/chat.py — guard run_generation against uncaught exceptions.

run_generation is launched with asyncio.create_task(); any exception
raised inside the coroutine is silently swallowed by the event loop,
the buffer stays in GenerationState.RUNNING forever, and every
subsequent POST /api/chat/conversations/<id>/messages returns 409
'Generation already in progress' — locking the user out of the chat
with no log trail.

Observed in dev 2026-05-22: assistant message 768 created at 20:36:59
with status=generating, stayed in that state for an hour+, and four
follow-up message attempts returned 409 instantly. The generation
task hung before any internal log line could fire, so the only
diagnostic was the 409 responses themselves.

Wrap run_generation in _run_generation_guarded() that catches
exceptions, logs with full traceback, transitions the buffer to
ERRORED, emits a final 'done' SSE event so any active stream
client closes cleanly, and marks the assistant message status=error
in the DB. After this, a stuck conversation recovers on its own
the next time the user sends a message — no manual DB poke needed.

2. services/curator_scheduler.py — pass last_curator_run_at as 'since'
to the curator so each sweep only sees messages added after the
previous successful pass.

Previously the scheduler called run_curator_for_conversation(conv_id)
with no 'since' argument, so the curator defaulted to its 24h
lookback window. Within an active journal session that meant every
15-min sweep re-extracted beats from messages already captured
on prior sweeps — producing duplicate moments.

_candidate_conversations() now returns (conv_id, last_curator_run_at)
tuples; _sweep() threads the timestamp through. First-run case
(last_curator_run_at IS NULL) falls back to the curator's default
24h window, which is what we want — process recent backlog on
first contact, then only deltas after.

Manual trigger path (POST /api/journal/curator/run/<conv_id>) is
intentionally NOT changed; it still passes since=None so the
24h re-sweep behaviour is preserved for ad-hoc 'reprocess today'
clicks from the UI.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/fabledassistant/routes/chat.py

@@ -178,19 +178,58 @@ async def send_message_route(conv_id: int):
 
     model = await get_setting(uid, "default_model", Config.OLLAMA_MODEL) or Config.OLLAMA_MODEL
 
-    # Launch background generation task (context building happens inside the task)
-    asyncio.create_task(run_generation(
-        buf, history, model,
-        uid, conv_id, conv.title, content,
-        context_note_id=context_note_id,
-        include_note_ids=include_note_ids,
-        excluded_note_ids=excluded_note_ids,
-        think=think,
-        rag_project_id=effective_rag_project_id,
-        workspace_project_id=workspace_project_id,
-        user_timezone=user_timezone,
-        voice_mode=(conv.conversation_type == "voice"),
-    ))
+    # Launch background generation task (context building happens inside the task).
+    #
+    # Wrap in a top-level guard: `run_generation` is fire-and-forget via
+    # asyncio.create_task. Without a wrapper, an unhandled exception inside
+    # the coroutine is swallowed by the event loop, the buffer stays in
+    # GenerationState.RUNNING forever, and every subsequent POST to this
+    # conversation returns 409 — locking the user out of the chat surface
+    # with no log trail. Observed in production 2026-05-22 where the
+    # generation hung before any internal log line could fire.
+    async def _run_generation_guarded():
+        try:
+            await run_generation(
+                buf, history, model,
+                uid, conv_id, conv.title, content,
+                context_note_id=context_note_id,
+                include_note_ids=include_note_ids,
+                excluded_note_ids=excluded_note_ids,
+                think=think,
+                rag_project_id=effective_rag_project_id,
+                workspace_project_id=workspace_project_id,
+                user_timezone=user_timezone,
+                voice_mode=(conv.conversation_type == "voice"),
+            )
+        except Exception:
+            logger.exception(
+                "run_generation crashed for conv %d msg %d (model=%s); "
+                "transitioning buffer to FAILED so the route can be used again",
+                conv_id, assistant_msg.id, model,
+            )
+            try:
+                buf.state = GenerationState.ERRORED
+                buf.append_event(
+                    "done",
+                    {"done": True, "error": "Generation crashed; see server logs"},
+                )
+            except Exception:
+                logger.exception("Failed to mark buffer ERRORED for conv %d", conv_id)
+            try:
+                from fabledassistant.services.generation_task import _update_message
+                await _update_message(
+                    assistant_msg.id,
+                    "",
+                    "error",
+                    tool_calls=None,
+                )
+            except Exception:
+                logger.exception(
+                    "Failed to mark assistant message %d as error after crash",
+                    assistant_msg.id,
+                )
+
+    asyncio.create_task(_run_generation_guarded())
 
     return jsonify({
         "assistant_message_id": assistant_msg.id,

src/fabledassistant/services/curator_scheduler.py

@@ -56,9 +56,15 @@ _loop: asyncio.AbstractEventLoop | None = None
 _running_lock = asyncio.Lock()  # prevents overlapping sweeps if one runs long
 
 
-async def _candidate_conversations() -> list[int]:
-    """Return conversation IDs that have user messages newer than the
-    last curator run. Limited to journal type, capped at the sweep limit.
+async def _candidate_conversations() -> list[tuple[int, datetime.datetime | None]]:
+    """Return (conv_id, last_curator_run_at) for journal conversations that
+    have user messages newer than the last curator run. Capped at the
+    sweep limit.
+
+    Returning the timestamp alongside the id lets `_sweep` pass it to the
+    curator as the `since` cutoff, so each sweep only sees messages added
+    since the previous successful pass — no re-extracting already-captured
+    beats.
     """
     async with async_session() as session:
         # Per-conversation: newest USER message timestamp, vs last_curator_run_at.
@@ -72,7 +78,7 @@ async def _candidate_conversations() -> list[int]:
             .subquery()
         )
         stmt = (
-            select(Conversation.id)
+            select(Conversation.id, Conversation.last_curator_run_at)
             .join(latest_user_msg, latest_user_msg.c.conv_id == Conversation.id)
             .where(Conversation.conversation_type == "journal")
             .where(
@@ -83,7 +89,7 @@ async def _candidate_conversations() -> list[int]:
             .limit(_MAX_CONVERSATIONS_PER_SWEEP)
         )
         rows = await session.execute(stmt)
-        return [row[0] for row in rows.all()]
+        return [(row[0], row[1]) for row in rows.all()]
 
 
 async def _stamp_last_run(conv_id: int, summary: str | None = None) -> None:
@@ -129,9 +135,16 @@ async def _sweep() -> None:
             "Curator sweep: %d journal conversation(s) to process",
             len(candidates),
         )
-        for conv_id in candidates:
+        for conv_id, last_run_at in candidates:
             try:
-                result = await run_curator_for_conversation(conv_id)
+                # Pass last_run_at as the `since` cutoff so the curator only
+                # examines messages added since the previous successful pass.
+                # Without this, every sweep re-loads the curator's default
+                # 24h window — and the LLM re-extracts beats from messages
+                # that were already captured, producing duplicate moments.
+                result = await run_curator_for_conversation(
+                    conv_id, since=last_run_at,
+                )
                 if result.error:
                     logger.warning(
                         "Curator run errored for conv %d (will retry next sweep): %s",