(expressive mode): pass expressions at segment start#6359
Conversation
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| if self._is_delta_stream: | ||
| remaining = self._stripper.flush() | ||
| tags = self._stripper.tags | ||
| if not self._writer_expression_sent: | ||
| extra_attributes = expression_attribute(self._pending_expressions()) |
There was a problem hiding this comment.
🔍 Late expression tags arriving without subsequent text are silently dropped
When an expression tag arrives after text has been written (_writer_has_text = True) but the stripper buffers it (e.g. partial tag split across chunks), and then flush() is called before more text arrives, the expression is lost. In flush() at livekit-agents/livekit/agents/voice/room_io/_output.py:607, self._writer_expression_sent is True (from the opening header), so _pending_expressions() is never consulted for the closing header. The buffered tag completes during self._stripper.flush() and its expression is added to self._stripper.tags, but it's never attached to any wire segment. This is arguably correct (an expression with no following text is semantically empty), but it means a tag that was split across the last two chunks of a turn is silently discarded rather than riding the closing header.
Was this helpful? React with 👍 or 👎 to provide feedback.
| async def _rotate_writer(self, pending: list[ExpressiveTag]) -> None: | ||
| """Finalize the current wire segment and open a new one led by the pending expression.""" | ||
| assert self._writer is not None | ||
| attributes = {ATTRIBUTE_TRANSCRIPTION_FINAL: "true"} | ||
| if self._track_id: | ||
| attributes[ATTRIBUTE_TRANSCRIPTION_TRACK_ID] = self._track_id | ||
| await self._writer.aclose(attributes=attributes) | ||
|
|
||
| self._current_id = utils.shortuuid("SG_") | ||
| self._writer = await self._create_text_writer( | ||
| extra_attributes=expression_attribute(pending) | ||
| ) | ||
| self._writer_expression_sent = True | ||
| self._writer_has_text = False |
There was a problem hiding this comment.
🔍 If _rotate_writer fails mid-way, the old writer is closed but self._writer is stale
In _rotate_writer at line 484-497, the old writer is closed (await self._writer.aclose(...)) before the new one is created. If _create_text_writer raises (e.g. room disconnected), self._writer still references the old, now-closed writer. The exception propagates to capture_text's try/except (line 564), which logs a warning. On the next capture_text call, self._writer is the closed writer, and self._writer is None is False, so the code may attempt to write to the closed writer. This is an error-recovery edge case that also existed in the old code (writer creation failure left inconsistent state), so it's pre-existing rather than newly introduced.
Was this helpful? React with 👍 or 👎 to provide feedback.
1 participant
No description provided.