mixdog 0.9.0 → 0.9.2

package/src/tui/lib/voice-recorder.mjs

@@ -0,0 +1,453 @@
+/**
+ * src/tui/lib/voice-recorder.mjs — TUI-local voice install/toggle orchestration.
+ *
+ * Step1 (/voice command): owns the `voice.enabled` config flag and drives the
+ * managed whisper.cpp + ffmpeg runtime install (voice-runtime-fetcher.mjs) the
+ * first time voice is turned ON.
+ *
+ * Step2 (Ctrl+Space recorder): startRecording/stopRecording/cancelRecording
+ * drive a small idle -> recording -> transcribing -> idle state machine.
+ * Recording spawns ffmpeg against a DirectShow ("dshow") capture device (the
+ * only capture API implemented — mixdog's managed voice runtime targets
+ * win32; the enumerate/record commands below are DirectShow-specific).
+ * Transcription reuses the SAME whisper-server.mjs singleton manager the
+ * channels pipeline uses (ensureReady + transcribe), so a TUI-triggered
+ * transcription and a Discord-attachment transcription share one warm
+ * whisper-server child instead of racing two.
+ *
+ * Config load is a STATIC import here (approved design decision) — unlike
+ * theme.mjs's dist-aware dynamic import of config.mjs, this file is not on the
+ * TUI's hot boot path and esbuild bundling config.mjs's relative dependency
+ * graph (plugin-paths/atomic-file/user-data-guard + the keychain .cjs via
+ * createRequire) into src/tui/dist/index.mjs is harmless: config.mjs's
+ * createRequire('../../lib/keychain-cjs.cjs') resolves relative to
+ * import.meta.url, and src/tui/dist/ sits at the same depth-from-src as
+ * src/runtime/shared/ (2 levels), so the relative path still lands on
+ * src/lib/keychain-cjs.cjs after bundling.
+ *
+ * The whisper.cpp/ffmpeg runtime installer (voice-runtime-fetcher.mjs) and
+ * the whisper-server manager (whisper-server.mjs) are BOTH loaded lazily via
+ * dynamic import — neither should load into memory unless a user actually
+ * asks for voice.
+ */
+import { spawn, spawnSync } from 'node:child_process';
+import { cpus, tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { unlinkSync } from 'node:fs';
+import { readSection, updateSection } from '../../runtime/shared/config.mjs';
+import { resolvePluginData } from '../../runtime/shared/plugin-paths.mjs';
+
+let _voiceRuntimeFetcherPromise = null;
+function loadVoiceRuntimeFetcher() {
+  if (!_voiceRuntimeFetcherPromise) {
+    _voiceRuntimeFetcherPromise = import('../../runtime/channels/lib/voice-runtime-fetcher.mjs');
+  }
+  return _voiceRuntimeFetcherPromise;
+}
+
+let _whisperServerPromise = null;
+// Set once loadWhisperServer()'s import resolves, so disposeRecorder() (a
+// SYNC function called from a React unmount cleanup) can check "was the
+// whisper-server manager ever engaged this session" without triggering a
+// fresh dynamic import just to shut down something that was never started.
+let _whisperServerModule = null;
+function loadWhisperServer() {
+  if (!_whisperServerPromise) {
+    _whisperServerPromise = import('../../runtime/channels/lib/whisper-server.mjs').then((mod) => {
+      _whisperServerModule = mod;
+      return mod;
+    });
+  }
+  return _whisperServerPromise;
+}
+
+// Reentrancy guard: only one install (ensureWhisperRuntime/ensureWhisperModel/
+// ensureFfmpegRuntime sequence) may run at a time per process.
+let _voiceInstallBusy = false;
+
+/** True while a Step1 install sequence is in flight. */
+export function isVoiceInstallBusy() {
+  return _voiceInstallBusy;
+}
+
+/** Read `voice.enabled` from mixdog-config.json (top-level `voice` section). */
+export function isVoiceEnabled() {
+  try {
+    return readSection('voice')?.enabled === true;
+  } catch {
+    return false;
+  }
+}
+
+// Coalesce ensure*'s onProgress ({ phase, downloaded, total } | { phase:'extra', ... })
+// into a throttled (2s) human-readable pushNotice call so a multi-hundred-MB
+// model download doesn't spam the notice toast on every chunk.
+function makeThrottledProgressNotice(pushNotice, intervalMs = 2000) {
+  let lastEmitAt = 0;
+  const phaseLabel = (phase) => {
+    if (phase === 'model') return 'model';
+    if (phase === 'ffmpeg') return 'ffmpeg';
+    if (phase === 'extra') return 'extra data';
+    return 'whisper runtime';
+  };
+  return (progress = {}) => {
+    if (typeof pushNotice !== 'function') return;
+    const now = Date.now();
+    if (now - lastEmitAt < intervalMs) return;
+    lastEmitAt = now;
+    const total = Number(progress.total) || 0;
+    const downloaded = Number(progress.downloaded) || 0;
+    const label = phaseLabel(progress.phase);
+    const text = total > 0
+      ? `Voice: downloading ${label}… ${Math.min(100, Math.round((downloaded / total) * 100))}%`
+      : `Voice: downloading ${label}…`;
+    pushNotice(text, 'info');
+  };
+}
+
+/**
+ * Ensure every managed voice-runtime component (whisper binary+server,
+ * whisper model, ffmpeg) is installed. Installs ONLY the missing pieces —
+ * an already-installed component is never re-fetched. Returns the resolved
+ * runtime descriptor (resolveVoiceRuntime shape) once `installed` is true;
+ * throws on any ensure* failure (manifest fetch, sha256 mismatch, etc.).
+ */
+export async function ensureVoiceRuntimeReady({ dataDir = resolvePluginData(), pushNotice } = {}) {
+  const fetcher = await loadVoiceRuntimeFetcher();
+  let runtime = fetcher.resolveVoiceRuntime(dataDir);
+  if (runtime.installed) return runtime;
+
+  const onProgress = makeThrottledProgressNotice(pushNotice);
+  if (!runtime.binary || !runtime.serverCmd) {
+    await fetcher.ensureWhisperRuntime(dataDir, onProgress);
+  }
+  if (!runtime.model) {
+    await fetcher.ensureWhisperModel(dataDir, onProgress);
+  }
+  if (!runtime.ffmpeg) {
+    await fetcher.ensureFfmpegRuntime(dataDir, onProgress);
+  }
+
+  runtime = fetcher.resolveVoiceRuntime(dataDir);
+  if (!runtime.installed) {
+    throw new Error('voice runtime install did not complete (still missing a required component)');
+  }
+  return runtime;
+}
+
+/**
+ * /voice command entry point. OFF -> immediately flips `voice.enabled` to
+ * false and notices "Voice OFF" (no install work, no busy gate — a disable
+ * must never be blocked by an in-flight install of a DIFFERENT toggle-on).
+ * ON  -> busy-guarded: checks resolveVoiceRuntime, installs only missing
+ * components, then persists `voice.enabled: true` and notices "Voice ON —
+ * Ctrl+Space to record". Any ensure* failure notices the cause and leaves
+ * `voice.enabled` untouched (still off).
+ *
+ * Returns the NEW enabled state (true/false) on success, or null when the
+ * toggle could not run (install already in flight) or failed.
+ */
+export async function toggleVoice({ pushNotice } = {}) {
+  const dataDir = resolvePluginData();
+  if (isVoiceEnabled()) {
+    // Med-5: updateSection is a synchronous file write (readAllForRmW +
+    // atomic write under a file lock) and can throw — a locked/corrupt
+    // config, a permissions error, etc. An unguarded throw here would
+    // propagate out of the OFF branch uncaught (toggleVoice's try/catch
+    // below only wraps the ON path) and crash the caller. Guard it the
+    // same way the ON path is guarded, and report failure via pushNotice
+    // instead of silently pretending the toggle succeeded.
+    try {
+      updateSection('voice', (current) => ({ ...current, enabled: false }));
+    } catch (err) {
+      pushNotice?.(`Voice OFF failed: ${err?.message || err}`, 'error');
+      return { ok: false, error: err?.message || String(err) };
+    }
+    pushNotice?.('Voice OFF', 'info');
+    return false;
+  }
+  if (_voiceInstallBusy) {
+    pushNotice?.('Voice install is already running', 'warn');
+    return null;
+  }
+  _voiceInstallBusy = true;
+  try {
+    await ensureVoiceRuntimeReady({ dataDir, pushNotice });
+    updateSection('voice', (current) => ({ ...current, enabled: true }));
+    pushNotice?.('Voice ON — Ctrl+Space to record', 'info');
+    return true;
+  } catch (err) {
+    pushNotice?.(`Voice setup failed: ${err?.message || err}`, 'error');
+    return null;
+  } finally {
+    _voiceInstallBusy = false;
+  }
+}
+
+// ── Step2: idle -> recording -> transcribing -> idle ────────────────────────
+
+const RECORDER_STATE = Object.freeze({
+  IDLE: 'idle',
+  STARTING: 'starting',
+  RECORDING: 'recording',
+  TRANSCRIBING: 'transcribing',
+});
+
+// Module-scoped: the TUI has exactly one prompt/recorder, so one singleton
+// state machine (mirrors whisper-server.mjs's own single-manager pattern).
+const rec = {
+  state: RECORDER_STATE.IDLE,
+  child: null,     // ffmpeg ChildProcess while RECORDING
+  wavPath: null,    // tmp wav path for the in-flight/just-finished recording
+};
+
+/** Current recorder state — 'idle' | 'recording' | 'transcribing'. */
+export function getRecorderState() {
+  return rec.state;
+}
+
+function newTmpWavPath() {
+  return join(tmpdir(), `mixdog-rec-${Date.now()}.wav`);
+}
+
+// Resolve the ffmpeg binary: prefer the managed runtime (Step1 install),
+// fall back to a bare 'ffmpeg' on PATH (spawn resolves PATH itself when no
+// path separator is present), so a user with a system ffmpeg install still
+// gets Ctrl+Space without running /voice's installer.
+async function resolveFfmpegCmd(dataDir) {
+  const fetcher = await loadVoiceRuntimeFetcher();
+  const managed = fetcher.resolveManagedFfmpegPath(dataDir);
+  if (managed) return managed;
+  const probe = spawnSync(process.platform === 'win32' ? 'where' : 'which', ['ffmpeg'], { windowsHide: true });
+  if (probe.status === 0) return 'ffmpeg';
+  return null;
+}
+
+// Enumerate DirectShow audio-capture devices via
+// `ffmpeg -hide_banner -list_devices true -f dshow -i dummy`, which ffmpeg
+// (by design) always exits non-zero on while printing the device list to
+// stderr. Returns the first "(audio)" device name, or null when none is
+// found / the probe fails. UTF-8 decoded (device names may be non-ASCII on
+// a localized Windows install).
+function listFirstDshowAudioDevice(ffmpegCmd) {
+  const r = spawnSync(ffmpegCmd, ['-hide_banner', '-list_devices', 'true', '-f', 'dshow', '-i', 'dummy'], {
+    windowsHide: true,
+    encoding: 'utf8',
+  });
+  const text = `${r.stderr || ''}${r.stdout || ''}`;
+  // ffmpeg prints device lines like:  "Mic Name" (audio)
+  const lines = text.split(/\r?\n/);
+  for (const line of lines) {
+    const m = /"([^"]+)"\s*\(audio\)/.exec(line);
+    if (m) return m[1];
+  }
+  return null;
+}
+
+/**
+ * Begin recording. IDLE -> RECORDING. No-op (returns null) unless the
+ * recorder is currently idle. Claims the IDLE -> STARTING transition
+ * SYNCHRONOUSLY (before any `await`) so a second Ctrl+Space fired while
+ * ffmpeg/device resolution is still in flight sees STARTING (not IDLE) and
+ * is a no-op rather than racing a second spawn onto the same rec.* fields.
+ * Resolves ffmpeg (managed -> PATH fallback), enumerates the first dshow
+ * audio-capture device, and spawns ffmpeg writing 16kHz mono PCM to a fresh
+ * tmp wav. `-y` overwrites (tmp path is always fresh so this is defensive,
+ * not load-bearing). Any failure on the STARTING path (no ffmpeg, no
+ * device, spawn throw) reverts to IDLE before returning.
+ *
+ * Returns `{ ok: true }` on a successful spawn, `{ ok: false, reason }` on
+ * any failure (no ffmpeg, no device, spawn error) — the caller (App.jsx)
+ * turns `reason` into a user-facing notice — or `null` when the recorder
+ * was not IDLE (already starting/recording/transcribing): a silent no-op,
+ * matching stopRecording's null-for-wrong-state contract.
+ */
+export async function startRecording({ dataDir = resolvePluginData() } = {}) {
+  if (rec.state !== RECORDER_STATE.IDLE) return null;
+  // Claim the slot BEFORE the first await — see High-1 in the doc comment
+  // above. Everything below this line runs against an already-STARTING
+  // rec.state, so a concurrent startRecording() call bails out via the
+  // guard above instead of racing a second ffmpeg spawn.
+  rec.state = RECORDER_STATE.STARTING;
+  const ffmpegCmd = await resolveFfmpegCmd(dataDir);
+  if (!ffmpegCmd) {
+    rec.state = RECORDER_STATE.IDLE;
+    return { ok: false, reason: 'ffmpeg not found — run /voice to install it, or install ffmpeg on PATH' };
+  }
+  const device = listFirstDshowAudioDevice(ffmpegCmd);
+  if (!device) {
+    rec.state = RECORDER_STATE.IDLE;
+    return { ok: false, reason: 'no microphone found (DirectShow audio device enumeration returned none)' };
+  }
+  const wavPath = newTmpWavPath();
+  let child;
+  try {
+    child = spawn(ffmpegCmd, [
+      '-f', 'dshow',
+      '-i', `audio=${device}`,
+      '-ar', '16000',
+      '-ac', '1',
+      '-y', wavPath,
+    ], { stdio: ['pipe', 'ignore', 'pipe'], windowsHide: true });
+  } catch (err) {
+    rec.state = RECORDER_STATE.IDLE;
+    return { ok: false, reason: `failed to start ffmpeg: ${err?.message || err}` };
+  }
+  // EPIPE guard (Med-3): a write to ffmpeg's stdin AFTER it has already
+  // closed the pipe (e.g. process died right as stopRecording() writes 'q')
+  // emits an 'error' on the stream — Node throws uncaught if nothing is
+  // listening. Attached once, here, so every future stdin write/close on
+  // this child is covered (stopRecording's write included).
+  child.stdin?.on('error', () => {});
+  // A spawn error surfaces async (e.g. EACCES on the resolved path even
+  // though the sync probe above found something on PATH) — fall back to
+  // IDLE so a stuck RECORDING state never blocks the next Ctrl+Space.
+  child.once('error', () => {
+    if (rec.child === child) {
+      rec.state = RECORDER_STATE.IDLE;
+      rec.child = null;
+      rec.wavPath = null;
+    }
+  });
+  rec.state = RECORDER_STATE.RECORDING;
+  rec.child = child;
+  rec.wavPath = wavPath;
+  return { ok: true };
+}
+
+// Graceful-stop budget before escalating to a hard kill. ffmpeg needs a
+// moment to flush the wav trailer after receiving 'q' on stdin.
+const RECORDING_STOP_GRACE_MS = 3000;
+
+function waitForExit(child, timeoutMs) {
+  return new Promise((resolve) => {
+    let done = false;
+    const finish = (exited) => {
+      if (done) return;
+      done = true;
+      resolve(exited);
+    };
+    child.once('exit', () => finish(true));
+    setTimeout(() => finish(false), timeoutMs).unref?.();
+  });
+}
+
+/**
+ * Stop the in-flight recording. RECORDING -> TRANSCRIBING -> (transcribe) ->
+ * IDLE. Sends 'q' on ffmpeg's stdin (its documented graceful-quit key —
+ * flushes the wav trailer); if ffmpeg has not exited within
+ * RECORDING_STOP_GRACE_MS it is force-killed (the partial wav is still
+ * attempted). No-op (returns null) unless currently RECORDING.
+ *
+ * On success runs ensureReady + transcribe against the SAME whisper-server
+ * singleton the channels pipeline uses, using `voice.language` from config
+ * (undefined -> whisper auto-detect) and `voice.transcription.threadCount`
+ * (falls back to the same quartered-cpu-count default index.mjs uses).
+ * Returns `{ ok: true, text }` on success or `{ ok: false, reason }` on any
+ * failure; state returns to IDLE either way.
+ *
+ * Med-1: rec.child is INTENTIONALLY kept set (not nulled) for the entire
+ * grace-wait window below — only cleared once the child has actually
+ * exited (or been killed). This lets disposeRecorder(), called mid-grace
+ * (e.g. the user quits the TUI right after Ctrl+Space-stop), still see and
+ * kill the SAME child instead of finding rec.child already null and
+ * orphaning a live ffmpeg process.
+ */
+export async function stopRecording({ dataDir = resolvePluginData() } = {}) {
+  if (rec.state !== RECORDER_STATE.RECORDING || !rec.child) return null;
+  const child = rec.child;
+  const wavPath = rec.wavPath;
+  rec.state = RECORDER_STATE.TRANSCRIBING;
+  try {
+    try { child.stdin?.write('q'); } catch { /* stdin may already be closed */ }
+    const exited = await waitForExit(child, RECORDING_STOP_GRACE_MS);
+    if (!exited) {
+      try { child.kill('SIGKILL'); } catch { /* best-effort */ }
+      await waitForExit(child, RECORDING_STOP_GRACE_MS);
+    }
+    // Only now is the child provably gone (or we gave up trying) — safe to
+    // drop the reference so a concurrent disposeRecorder() no longer needs
+    // (and no longer has) anything to kill.
+    if (rec.child === child) rec.child = null;
+    return await transcribeWav(wavPath, { dataDir });
+  } finally {
+    if (rec.child === child) rec.child = null;
+    rec.state = RECORDER_STATE.IDLE;
+    rec.wavPath = null;
+  }
+}
+
+/**
+ * Cancel the in-flight recording. RECORDING -> IDLE (no transcription).
+ * Force-kills ffmpeg immediately and deletes the tmp wav. No-op (returns
+ * false) unless currently RECORDING.
+ */
+export function cancelRecording() {
+  if (rec.state !== RECORDER_STATE.RECORDING || !rec.child) return false;
+  const child = rec.child;
+  const wavPath = rec.wavPath;
+  try { child.kill('SIGKILL'); } catch { /* best-effort */ }
+  if (wavPath) {
+    try { unlinkSync(wavPath); } catch { /* tmp file may not exist yet */ }
+  }
+  rec.state = RECORDER_STATE.IDLE;
+  rec.child = null;
+  rec.wavPath = null;
+  return true;
+}
+
+/**
+ * Best-effort teardown for TUI unmount/exit: kills any live ffmpeg child and
+ * deletes its tmp wav, regardless of current state (RECORDING or
+ * TRANSCRIBING — the latter can still hold a live child mid-stop-grace).
+ * Never throws.
+ */
+export function disposeRecorder() {
+  const child = rec.child;
+  const wavPath = rec.wavPath;
+  if (child) {
+    try { child.kill('SIGKILL'); } catch { /* best-effort */ }
+  }
+  if (wavPath) {
+    try { unlinkSync(wavPath); } catch { /* best-effort */ }
+  }
+  rec.state = RECORDER_STATE.IDLE;
+  rec.child = null;
+  rec.wavPath = null;
+  // High-2: shut down the whisper-server child too, not just ffmpeg — a
+  // TUI-triggered transcription can leave the managed whisper-server.exe
+  // process running past TUI exit (it's a long-lived singleton by design,
+  // kept warm across transcriptions), which otherwise orphans it. Only
+  // meaningful if loadWhisperServer() actually resolved at least once this
+  // session (_whisperServerModule set) — nothing to stop otherwise, and we
+  // must not trigger a fresh dynamic import from a sync unmount callback.
+  if (_whisperServerModule) {
+    try { void _whisperServerModule.stopVoiceWhisperServer?.(); } catch { /* best-effort */ }
+  }
+}
+
+async function transcribeWav(wavPath, { dataDir }) {
+  // Med-2: the runtime.installed check is now INSIDE the try so an early
+  // return here still hits the finally below and unlinks wavPath — the
+  // original early-return (before try) skipped cleanup and leaked the tmp
+  // wav on every "voice runtime not installed" path.
+  try {
+    const fetcher = await loadVoiceRuntimeFetcher();
+    const runtime = fetcher.resolveVoiceRuntime(dataDir);
+    if (!runtime?.installed) {
+      return { ok: false, reason: 'voice runtime not installed — run /voice to install it' };
+    }
+    const server = await loadWhisperServer();
+    const cfg = readSection('voice');
+    const cpuCount = (() => { try { return cpus().length; } catch { return 2; } })();
+    const threadCount = cfg?.transcription?.threadCount ?? Math.max(1, Math.ceil(cpuCount / 4));
+    const language = cfg?.language || undefined;
+    await server.ensureReady({ serverCmd: runtime.serverCmd, modelPath: runtime.modelPath, threadCount, host: '127.0.0.1' });
+    const text = await server.transcribe(wavPath, { language });
+    return { ok: true, text: text.trim() };
+  } catch (err) {
+    return { ok: false, reason: err?.message || String(err) };
+  } finally {
+    try { unlinkSync(wavPath); } catch { /* best-effort cleanup */ }
+  }
+}