fix(agents): auto-title crash on metadata=None#16
Merged
Conversation
The /auto-title endpoint passes metadata=None because session-title generation isn't part of an agent invocation and must not show up in Langfuse. But LLMClient._build_langfuse_metadata accessed call_meta.analytics_consent unconditionally, so every auto-title call raised AttributeError. The endpoint's outer try/except hid the failure — every new chat session silently fell back to a 60-char slice of the first user message instead of the LLM-generated 3-6 word title, and logs filled with "auto-title LLM call failed: 'NoneType' object has no attribute analytics_consent". Make metadata optional in acompletion's signature and have _build_langfuse_metadata return None when call_meta is None (same behavior as analytics_consent="off"). Adds regression test.
When the frontend aborts a streaming endpoint (tab close, navigation,
network blip), uvicorn cancels the ASGI task with
asyncio.CancelledError. That hits get_db while it's running
session.commit() after the yield. Our `except Exception:` did NOT catch
CancelledError (it's a BaseException, not Exception), so:
1. commit() bubbles the CancelledError out of the try.
2. The async-with __aexit__ kicks in and tries to terminate the
asyncpg connection.
3. Connection terminate ALSO runs in the cancelled scope, raises
CancelledError again — SQLAlchemy logs "Exception terminating
connection ..." at ERROR level.
4. Then the original CancelledError bubbles all the way up, uvicorn
logs "Exception in ASGI application" with a multi-screen traceback.
Net effect: every SSE disconnect produces a wall of red in prod logs
even though the cancellation is correct and resources still get cleaned
up. Restructure get_db to catch BaseException, do a best-effort rollback
(suppressed because the scope may already be torn down), and re-raise.
Move commit() into the else branch so it only runs on normal completion.
Self-hosted Langfuse setups often use LANGFUSE_BASE_URL in their .env (matching the public Langfuse SDK docs in some Docker examples). Our code reads only LANGFUSE_HOST (LiteLLM convention), so a misnamed env var silently disables tracing — `is_langfuse_configured()` returns False, callbacks never register, and operators see no traces with no loud error. Use Pydantic AliasChoices so langfuse_host accepts either env name. LANGFUSE_HOST wins when both are set (canonical first). Document the alias in .env.example and README. Adds two regression tests that instantiate a fresh Settings() with each env-var combination.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
LLMClient._build_langfuse_metadataraisedAttributeErrorwhenever the caller passedmetadata=None. The/agents/sessions/{id}/auto-titleendpoint always does this (title generation isn't part of an agent invocation, must not show in Langfuse), so every fresh chat session silently fell back to a 60-char slice of the first user message instead of the LLM-generated 3-6 word title, withauto-title LLM call failed: 'NoneType' object has no attribute 'analytics_consent'in prod logs.metadata: LLMCallMetadata | Noneinacompletion. Have_build_langfuse_metadatareturnNonewhencall_meta is None(same path asanalytics_consent="off").test_langfuse_metadata_none_returns_noneregression.Notes on the
notificationserrors in the same prod logThe
relation "notifications" does not existerrors stopped on prod between 09:07 and 10:54 UTC today — that's the deploy of #14 picking up thea1f8c9d2b3e4_repair_notifications_tablemigration (CREATE TABLE IF NOT EXISTS). No further action needed unless those errors come back.Test plan
pytest tests/agents/test_llm.py -q— 19 passedauto-title LLM call failed— should be gone