Function finishRecording

src/hooks/useVoice.ts:322–522

()

  }, [setVoiceState])

  function finishRecording(): void {
    logForDebugging(
      '[voice] finishRecording: stopping recording, transitioning to processing',
    )
    // Session ending — stale any in-flight attempt so its late onError
    // (conn 2 responding after user released key) doesn't double-fire on
    // top of the "check network" message below.
    attemptGenRef.current++
    // Capture focusTriggered BEFORE clearing it — needed as an event dimension
    // so BigQuery can filter out passive focus-mode auto-recordings (user focused
    // terminal without speaking → ambient noise sets hadAudioSignal=true → false
    // silent-drop signature). focusFlushedCharsRef fixes transcriptChars accuracy
    // for sessions WITH speech; focusTriggered enables filtering sessions WITHOUT.
    const focusTriggered = focusTriggeredRef.current
    focusTriggeredRef.current = false
    updateState('processing')
    voiceModule?.stopRecording()
    // Capture duration BEFORE the finalize round-trip so that the WebSocket
    // wait time is not included (otherwise a quick tap looks like > 2s).
    // All ref-backed values are captured here, BEFORE the async boundary —
    // a keypress during the finalize wait can start a new session and reset
    // these refs (e.g. focusFlushedCharsRef = 0 in startRecordingSession),
    // reproducing the silent-drop false-positive this ref exists to prevent.
    const recordingDurationMs = Date.now() - recordingStartRef.current
    const hadAudioSignal = hasAudioSignalRef.current
    const retried = retryUsedRef.current
    const focusFlushedChars = focusFlushedCharsRef.current
    // wsConnected distinguishes "backend received audio but dropped it" (the
    // bug backend PR #287008 fixes) from "WS handshake never completed" —
    // in the latter case audio is still in audioBuffer, never reached the
    // server, but hasAudioSignalRef is already true from ambient noise.
    const wsConnected = everConnectedRef.current
    // Capture generation BEFORE the .then() — if a new session starts during
    // the finalize wait, sessionGenRef has already advanced by the time the
    // continuation runs, so capturing inside the .then() would yield the new
    // session's gen and every staleness check would be a no-op.
    const myGen = sessionGenRef.current
    const isStale = () => sessionGenRef.current !== myGen
    logForDebugging('[voice] Recording stopped')

    // Send finalize and wait for the WebSocket to close before reading the
    // accumulated transcript.  The close handler promotes any unreported
    // interim text to final, so we must wait for it to fire.
    const finalizePromise: Promise<FinalizeSource | undefined> =
      connectionRef.current
        ? connectionRef.current.finalize()
        : Promise.resolve(undefined)

    void finalizePromise
      .then(async finalizeSource => {
        if (isStale()) return
        // Silent-drop replay: when the server accepted audio (wsConnected),
        // the mic captured real signal (hadAudioSignal), but finalize timed
        // out with zero transcript — the ~1% session-sticky CE-pod bug.
        // Replay the buffered audio on a fresh connection once. A 250ms
        // backoff clears the same-pod rapid-reconnect race (same gap as the
        // early-error retry path below).
        if (