Function useVoice

src/hooks/useVoice.ts:199–1144

({
  onTranscript,
  onError,
  enabled,
  focusMode,
}: UseVoiceOptions)

}

export function useVoice({
  onTranscript,
  onError,
  enabled,
  focusMode,
}: UseVoiceOptions): UseVoiceReturn {
  const [state, setState] = useState<VoiceState>('idle')
  const stateRef = useRef<VoiceState>('idle')
  const connectionRef = useRef<VoiceStreamConnection | null>(null)
  const accumulatedRef = useRef('')
  const onTranscriptRef = useRef(onTranscript)
  const onErrorRef = useRef(onError)
  const cleanupTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
  const releaseTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
  // True once we've seen a second keypress (auto-repeat) while recording.
  // The OS key repeat delay (~500ms on macOS) means the first keypress is
  // solo — arming the release timer before auto-repeat starts would cause
  // a false release.
  const seenRepeatRef = useRef(false)
  const repeatFallbackTimerRef = useRef<ReturnType<typeof setTimeout> | null>(
    null,
  )
  // True when the current recording session was started by terminal focus
  // (not by a keypress). Focus-driven sessions end on blur, not key release.
  const focusTriggeredRef = useRef(false)
  // Timer that tears down the session after prolonged silence in focus mode.
  const focusSilenceTimerRef = useRef<ReturnType<typeof setTimeout> | null>(
    null,
  )
  // Set when a focus-mode session is torn down due to silence. Prevents
  // the focus effect from immediately restarting. Cleared on blur so the
  // next focus cycle re-arms recording.
  const silenceTimedOutRef = useRef(false)
  const recordingStartRef = useRef(0)
  // Incremented on each startRecordingSession(). Callbacks capture their
  // generation and bail if a newer session has started — prevents a zombie
  // slow-connecting WS from an abandoned session from overwriting
  // connectionRef mid-way through the next session.
  const sessionGenRef = useRef(0)
  // True if the early-error retry fired during this session.
  // Tracked for the tengu_voice_recording_completed analytics event.
  const retryUsedRef = useRef(false)
  // Full audio captured this session, kept for silent-drop replay. ~1% of
  // sessions get a sticky-broken CE pod that accepts audio but returns zero
  // transcripts (anthropics/anthropic#287008 session-sticky variant); when
  // finalize() resolves via no_data_timeout with hadAudioSignal=true, we
  // replay the buffer on a fresh WS once. Bounded: 32KB/s × ~60s max ≈ 2MB.
  const fullAudioRef = useRef<Buffer[]>([])
  const silentDropRetriedRef = useRef(false)
  // Bumped when the early-error retry is scheduled. Captured per
  // attemptConnect — onError swallows stale-gen events (conn 1's
  // trailing close-error) but surfaces current-gen ones (conn 2's
  // genuine failure). Same shape as sessionGenRef, one level down.
  const attemptGenRef = useRef(0)
  // Running total of chars flushed in focus mode (each final transcript is
  // injected immediately and accumulatedRef reset). Added to transcriptChars
  // in the completed event so focus-mode sessions don't false-positive as
  // silent-drops (transcriptChars=0 despite successful transcription).