fix: re-trigger deferred context frame push on UserStoppedSpeakingFrame

[elliottventures]

What

When LLMAssistantAggregator._handle_function_call_result runs while self._user_speaking is True, the outer conditional currently drops the context push silently — unlike the bot-speaking branch, which sets _push_context_on_bot_stopped_speaking = True and re-triggers on BotStoppedSpeakingFrame. This PR mirrors that pattern for the user-side.

Why

A short user trigger utterance whose transcription-driven turn-start races the function-call result hits this window:

1. [Transcription:user] [Next.] arrives
2. _run_function_call fires (driven by the same transcription)
3. _on_user_turn_started sets self._user_speaking = True
4. FunctionCallResultFrame arrives at the aggregator
5. Outer if run_llm and not self._user_speaking: evaluates False → push silently dropped
6. _on_user_turn_stopped fires a few ms later, unsetting _user_speaking — but nothing re-evaluates the dropped push

Result: the tool response never reaches the LLM service. For Gemini Live specifically, this manifests as the model sitting silent waiting for a response to its function call, until the user speaks again and the interruption path kicks in. TTFB on affected turns can reach minutes.

The bot-speaking branch has had this covered via _push_context_on_bot_stopped_speaking + re-trigger in BotStoppedSpeakingFrame for a while. The user-side was missing the equivalent.

The fix

Five small edits to LLMAssistantAggregator, all mirroring the existing bot-side pattern:

1. New flag _push_context_on_user_stopped_speaking in __init__
2. Reset it in reset() and push_context_frame() alongside the bot flag
3. Outer condition at the run_llm check changed from if run_llm and not self._user_speaking: to if run_llm:, with a new elif self._user_speaking: branch inside that sets the flag (before the existing elif self._bot_speaking:)
4. process_frame's UserStoppedSpeakingFrame handler flushes the deferred push if the flag is set and the bot isn't speaking

Total +18 / -1 lines.

Testing

• Existing tests/test_context_aggregators_universal.py — 37/37 pass pre- and post-change.
• Verified against the same production voice agent referenced in fix: send realtime input after tool_result so Gemini 3.x runs inference #4366. Before this fix (subclass monkeypatch locally): every compound user utterance (e.g., "Freezer is on top. Next.") stalled silently. After this fix: clean turn completion across 20+ test turns.

Happy to add a targeted unit test for the race if a maintainer has a preferred shape — it needs a bit of plumbing to simulate the specific UserStartedSpeakingFrame → FunctionCallResultFrame → UserStoppedSpeakingFrame ordering.

Related

Companion to #4366. Both surfaced during the same Gemini Live function-calling diagnosis.

[elliottventures]

@markbackman Wanted to make sure you guys saw this one as well