Fixes #681 — #681

Problem

Room._on_room_event does an unchecked self.local_participant.track_publications[sid] lookup on the local_track_unpublished branch (room.py). LocalParticipant.unpublish_track removes the publication from _track_publications when its FFI async response is processed, and that response races the local_track_unpublished room event. When the response is handled first, the SID is already gone by the time the event is dispatched and the handler raises KeyError.

_listen_task catches it and logs error running user callback for local_track_unpublished: ... with a full traceback, so the worker does not crash — but every affected disconnect cleanup produces a spurious ERROR log. In a production agent fleet this fired ~18×/24h on early/CLIENT_INITIATED disconnects (~1 in several hundred). The local_track_unpublished event is also silently dropped when the race fires.

This is the only local_track_* handler that crashes on a missing SID — track_unpublished (remote) and local_track_republished (#655) already tolerate it.

Fix

room.py — handler: pop defensively and only emit when present, mirroring the remote track_unpublished and local_track_republished handlers (with a debug log for the dropped case):

sid = event.local_track_unpublished.publication_sid
lpublication = self.local_participant._track_publications.pop(sid, None)
if lpublication is not None:
    self.emit("local_track_unpublished", lpublication)
else:
    logging.debug("local_track_unpublished for untracked publication sid %s", sid)

This also folds removal into the handler, fixing a latent leak where a server-forced unpublish (no unpublish_track call) previously left the publication in the dict forever.

participant.py — unpublish_track: make its pop tolerant.

publication = self._track_publications.pop(track_sid, None)
if publication is not None:
    publication._track = None

This second change is required, not cosmetic. The KeyError is rare (~1 in several hundred), which means the common ordering is the room event being processed while the SID is still present — i.e. before unpublish_track's pop. Once the handler itself starts popping, in that common path the old unchecked pop(track_sid) in unpublish_track would raise a new KeyError on nearly every unpublish. Making it tolerant prevents simply relocating the crash.

Behavior across both race orderings

• Event first (common): handler pops + emits; unpublish_track gets None, skips — no crash, event still delivered.
• Async response first (the rare race that was crashing): unpublish_track pops + clears _track; handler gets None, skips emit — no crash. The event is dropped, but it was crashing (never emitting) before, so this is strictly better.

Testing

• ast-parsed both files; no new lint/type diagnostics.
• The repo has no unit tests around _on_room_event (existing tests are e2e and require a live server), so no test was added. Happy to add one with a mocked participant if preferred.

🤖 Generated with Claude Code