Agent events arrive on multiple threads concurrently. A stale event
with shorter accumulated text was falsely triggering 'text diverged',
causing the streaming TTS to restart with a new WebSocket — resulting
in multiple simultaneous ElevenLabs connections (2-3 voices) and
eventual system TTS fallback when hasReceivedAudio was false.
Fix: if sentFullText.startsWith(fullText), the event is stale (we
already have this text), not diverged. Accept and ignore it.
- Codex P1: setSpeakerEnabled now syncs talkMode.setPlaybackEnabled
so muting the speaker works when ttsOnAllResponses is active.
- Codex P2: abandonAudioFocus() called in stopSpeaking to prevent
audio focus leak after TTS completes or is interrupted.
- Codex P1: streamAndPlayMp3 was computed but never called after PCM
failure. Now properly invoked as fallback.
- Codex P2: MicCaptureManager.speakAssistantReply now skipped when
TalkModeManager.ttsOnAllResponses is active, preventing both
pipelines from speaking the same assistant reply.
Bug fixes:
- @Synchronized on ElevenLabsStreamingTts.sendText/finish to prevent
sentFullText/sentTextLength races across OkHttp and caller threads
- Pre-set pendingRunId via onRunIdKnown callback before chat.send to
eliminate race where gateway events arrive before runId is stored
- Track drain coroutine as Job; cancel prior on rapid mic toggle to
prevent duplicate TTS and stale transcript sends
- Mic button disabled during 2s drain cooldown (micCooldown StateFlow)
Codex review fixes:
- Gate agent streaming TTS on sessionKey to prevent cross-session
audio leaks (P1)
- Clear ElevenLabs credentials when talk.provider is not elevenlabs;
gate streaming TTS on activeProviderIsElevenLabs (P2)
System TTS fallback fixes:
- Null streamingTts immediately in finishStreamingTts so next response
gets a fresh TTS instance
- Add hasReceivedAudio flag to ElevenLabsStreamingTts to detect when
WebSocket connects but returns no audio (invalid key, network error)
- Fall back to playTtsForText when streaming TTS produced no audio
- Track ttsJob to cleanly cancel prior playTtsForText on new response
- Re-throw CancellationException instead of cascading into fallback
attempts that also get cancelled
ChatController:
- final/aborted/error run events now trigger a history refresh regardless of
whether the runId is in pendingRuns; only delta events require the run to be
tracked (prevents voice-initiated responses from being silently dropped)
MicCaptureManager:
- Don't auto-send on onResults silence detection — accumulate transcript
segments and send when mic is toggled off, giving the recognizer time to
finish processing buffered audio
- Capture any partial live transcript if no final segments arrived (2s drain
window before stop)
- Join multi-segment transcripts with sentence-ending punctuation to avoid
run-on text sent to the gateway
TalkModeManager is instantiated lazily in NodeRuntime and drives ElevenLabs
streaming TTS for all assistant responses when the voice screen is active.
MicCaptureManager continues to own STT and chat.send; TalkModeManager is
TTS-only (ttsOnAllResponses = true, setEnabled never called).
- talkMode.ttsOnAllResponses = true when mic is enabled or voice screen active
- Barge-in: tapping the mic button calls stopTts() before re-enabling mic
- Lifecycle: PostOnboardingTabs LaunchedEffect + VoiceTabScreen onDispose both
call setVoiceScreenActive(false) so TTS stops cleanly on tab switch or
app backgrounding
- applyMainSessionKey wires the session key into TalkModeManager so it
subscribes to the correct chat session for TTS
Streams text to the ElevenLabs WebSocket API and plays audio in real-time
via AudioTrack (PCM 24kHz). Key design points:
- sendText(fullText) takes the full accumulated text and only transmits the
new suffix, detecting divergence for restart
- Chunks are queued if the WebSocket isn't yet connected; flushed in onOpen
- finish() sends EOS to ElevenLabs; deferred if called before onOpen fires
- sendText returns true (not false) when finished=true to avoid treating a
normal end-of-stream as a diverge restart
- finishStreamingTts coroutine uses identity check before nulling streamingTts
to prevent a mid-drain restart from orphaning a live TTS session
- eleven_v3 does NOT support WebSocket streaming; use eleven_flash_v2_5
