Updates

README.md

@@ -90,7 +90,8 @@ curl -fsSL https://git.ssh.surf/snxraven/gnome-jarvis/raw/branch/main/packaging/
 Omit `--enable` to install files without starting the service. With
 `--enable`, the installer writes default settings if needed, starts
 `jarvisd`, and enables the GNOME extension. The default wake phrase is
-**hey jarvis**; change it in Settings. Spoken replies are on.
+**hey jarvis** with a built-in CPU detector; change it in Settings. Spoken
+replies are on.
 
 If GNOME says the extension does not exist immediately after installation, log
 out and back in once. GNOME Shell only refreshes its user extension catalogue
@@ -124,9 +125,9 @@ bash packaging/install.sh --enable
 ~~~
 
 The installer writes `~/.config/jarvis/config.json` when it is missing, with
-wake phrase `hey jarvis`, model profile `laptop-16gb`, and TTS enabled. Change
+wake phrase `hey jarvis`, CPU wake detector `jarvis-wake-bridge`, model profile
-those later in Settings. Diagnostics stay available as `npm run gpu-doctor`,
+`laptop-16gb`, and TTS enabled. Change those later in Settings. Diagnostics stay
-`npm run voice-doctor`, and `npm run cu-doctor`.
+available as `npm run gpu-doctor`, `npm run voice-doctor`, and `npm run cu-doctor`.
 
 Inspect the installed daemon with:
 

apps/gnome-extension/jarvis@qvac.local/extension.js

@@ -243,8 +243,9 @@ export default class JarvisExtension extends Extension {
       const voice = status.voice;
       this._eachView((view) => view.setConnectionStatus(voice ? 'local' : 'voice-unavailable'));
       if (voice) {
-        const input = voice.asr && voice.capture;
+        const input = Boolean(voice.capture);
-        this._eachView((view) => view.setVoiceStatus({ tts: voice.tts, input, wake: voice.wake }));
+        const tts = Boolean(voice.tts || voice.speech);
+        this._eachView((view) => view.setVoiceStatus({ tts, input, wake: voice.wake }));
       }
     } catch { this._eachView((view) => view.setConnectionStatus('voice-unavailable')); }
   }

apps/gnome-extension/jarvis@qvac.local/settings-catalog.json

@@ -481,9 +481,9 @@
       "key": "wakeCommand",
       "title": "Wake detector command",
       "group": "Wake and privacy",
-      "default": "",
+      "default": "jarvis-wake-bridge",
       "type": "string",
-      "description": "Advanced: local program that receives microphone audio. Leave empty to use Hold Talk.",
+      "description": "Local program that receives microphone audio. jarvis-wake-bridge is the built-in CPU detector. Leave it unless you have openWakeWord or sherpa.",
       "aliases": [
         "wake_command"
       ]
@@ -524,6 +524,17 @@
       "max": 240,
       "step": 1
     },
+    {
+      "key": "freeVramOnIdle",
+      "title": "Free GPU memory when idle",
+      "group": "Wake and privacy",
+      "default": true,
+      "type": "boolean",
+      "description": "Unload speech and chat models after the idle delay. Hey Jarvis stays ready on the CPU.",
+      "aliases": [
+        "free_vram_on_idle"
+      ]
+    },
     {
       "key": "vadThreshold",
       "title": "Speech detection threshold",
@@ -740,6 +751,32 @@
       "max": 20,
       "step": 1,
       "restart": true
+    },
+    {
+      "key": "fsAccess",
+      "title": "File access",
+      "group": "Agent limits",
+      "default": "workspace",
+      "type": "choice",
+      "description": "Requires restart. The agent still runs as your user account and asks before writing. Shell commands can already reach other paths.",
+      "options": [
+        {
+          "value": "workspace",
+          "label": "Jarvis workspace only"
+        },
+        {
+          "value": "home",
+          "label": "Home directory"
+        },
+        {
+          "value": "filesystem",
+          "label": "Entire filesystem"
+        }
+      ],
+      "aliases": [
+        "fs_access"
+      ],
+      "restart": true
     }
   ]
 }

daemon/harness-bridge.js

@@ -1,24 +1,34 @@
 import { voiceSettings } from './voice-settings.js';
 import { EventEmitter } from 'node:events';
+import os from 'node:os';
 import { acquireQvac, closeQvac, releaseQvac, Agent, QVAC_MASTER } from './qvac-master.js';
 import { createRuntimeTools } from '../skills/runtime-tools.js';
-import { createPhase2Tools } from '../skills/phase2-tools.js';
+import { createPhase2Tools, filesystemRoots } from '../skills/phase2-tools.js';
 import { createQvacTools } from '../skills/qvac-tools.js';
 import { VOICE_SYSTEM_PROMPT, parseHudSidecar } from '../skills/voice-prompt.js';
 import { createComputerObserveTools } from '../skills/computer-observe.js';
 import { createComputerActTools } from '../skills/computer-act.js';
 import { createPhase9GatewayTool } from '../skills/phase9-tools.js';
 
+export function harnessRoots(fsAccess) {
+  if (fsAccess === 'filesystem') return ['/'];
+  if (fsAccess === 'home') return [os.homedir()];
+  return [];
+}
+
 export class HarnessBridge extends EventEmitter {
-  constructor({ cwd = process.cwd(), model = QVAC_MASTER.model, tools = [], computer, observer, actuator, permissionMode = 'ask' } = {}) {
+  constructor({ cwd = process.cwd(), model = QVAC_MASTER.model, tools = [], computer, observer, actuator, permissionMode = 'ask', fsAccess } = {}) {
     super();
     const settings = voiceSettings();
+    const access = fsAccess ?? settings.fsAccess;
+    const roots = harnessRoots(access);
     this.options = {
       cwd,
+      roots,
       model,
       tools: [
         ...createRuntimeTools({ computer }),
-        ...createPhase2Tools({ cwd, computer }),
+        ...createPhase2Tools({ cwd, computer, roots: filesystemRoots(access, cwd) }),
         ...createComputerObserveTools({ computer, observer }),
         ...createComputerActTools({ actuator }),
         ...createQvacTools(),

daemon/index.js

@@ -8,7 +8,7 @@ import { HarnessBridge } from './harness-bridge.js';
 import { ComputerUseSession } from '../computer-use/session.js';
 import { VoiceStateMachine } from './voice-state.js';
 import { QvacScheduler } from './qvac-scheduler.js';
-import { cancelQvac, resumeQvac, suspendQvac, callQvac, cancelQvacRequest, qvacStatus } from './qvac-master.js';
+import { cancelQvac, closeQvac, resumeQvac, suspendQvac, callQvac, cancelQvacRequest, qvacStatus } from './qvac-master.js';
 import { PrivacyLog } from './privacy-log.js';
 import { VoiceLoop } from './voice-loop.js';
 import { QvacVoiceAdapter } from './voice-adapters.js';
@@ -37,7 +37,7 @@ export class JarvisDaemon extends EventEmitter {
     this.perception = new QvacPerception();
     this.observer = new DesktopObserver({ normalizer: new FrameNormalizer({ maxLongEdge: this.settings.screenshotMaxEdge, quality: this.settings.screenshotQuality }), ocr: (image) => this.perception.ocr(image) });
     this.actuator = new ComputerActuator({ session: this.computer, input: this.input, find: ({ ref }) => this.observer.lastTree.filter((node) => node.ref === ref), atspiAction: (target, action) => this.observer.atspi.action(target, action), highlight: async (target, action) => this.emit('ComputerHighlight', JSON.stringify({ rect: target?.rect || null, label: `${action} ${target?.name || ''}` })) , audit: this.audit });
-    this.harness = new HarnessBridge({ cwd: process.cwd(), computer: this.computer, observer: this.observer, actuator: this.actuator });
+    this.harness = new HarnessBridge({ cwd: process.cwd(), computer: this.computer, observer: this.observer, actuator: this.actuator, fsAccess: this.settings.fsAccess });
     this.log = new PrivacyLog();
     this.locked = false;
     this.lastReply = '';
@@ -57,8 +57,32 @@ export class JarvisDaemon extends EventEmitter {
   }
 
   setState(state) { this.state = state; try { this.recovery.save({ state, mode: this.mode }); } catch (error) { this.emit('Error', 'RECOVERY_WRITE', error.message); } this.emit('StateChanged', state); this.log.record('state', { state }).catch(() => {}); }
-  async arm() { if (this.locked) return; await resumeQvac().catch(() => {}); this.voice.wake(); this.setState('LISTENING'); }
+  async arm() {
-  async sleep() { this.cancel(); this.voice.sleep(); this.setState('SLEEPING'); await suspendQvac().catch((error) => this.emit('Error', 'QVAC_SUSPEND', error.message)); }
+    if (this.locked) return;
+    await resumeQvac().catch(() => {});
+    await this.ensureAsr();
+    this.voice.wake();
+    this.setState('LISTENING');
+  }
+  async sleep() {
+    if (this._activeAsk) {
+      await suspendQvac().catch((error) => this.emit('Error', 'QVAC_SUSPEND', error.message));
+      return;
+    }
+    this.cancel();
+    this.voice.sleep();
+    this.setState('SLEEPING');
+    if (this.settings.freeVramOnIdle !== false) await this.parkModels();
+    else await suspendQvac().catch((error) => this.emit('Error', 'QVAC_SUSPEND', error.message));
+  }
+  async parkModels() {
+    try { await this.voiceLoop?.parkModels?.(); }
+    catch (error) { this.emit('Error', 'VOICE_UNLOAD', error.message); }
+    try { await this.harness.resetContext(); }
+    catch (error) { this.emit('Error', 'QVAC_UNLOAD', error.message); }
+    try { await closeQvac(); }
+    catch (error) { this.emit('Error', 'QVAC_CLOSE', error.message); }
+  }
   say(text) {
     const spoken = spokenReply(text) || spokenReply(this.lastReply) || 'There is nothing to repeat.';
     this.lastReply = spoken;
@@ -145,18 +169,21 @@ export class JarvisDaemon extends EventEmitter {
     if (!settings.ttsEnabled) { this.voiceLoop.status.tts = false; return; }
     if (this.voiceLoop.status.tts) return;
     if (!this.voiceLoop.tts) this.voiceLoop.tts = new QvacVoiceAdapter({ role: 'tts', settings });
-    try {
+    const ready = await this.voiceLoop.ensureTts();
-      await this.voiceLoop.tts.start();
+    if (ready) console.log('jarvisd: voice: TTS ready');
-      this.voiceLoop.status.tts = true;
+    else if (this.voiceLoop.status.errors.tts) {
-      delete this.voiceLoop.status.errors.tts;
+      const message = this.voiceLoop.status.errors.tts;
-      console.log('jarvisd: voice: TTS ready');
-    } catch (error) {
-      const message = String(error?.message || error);
-      this.voiceLoop.status.errors.tts = message;
       console.error(`jarvisd: voice: TTS: ${message}`);
       this.emit('Error', 'TTS', message);
     }
   }
+  async ensureAsr() {
+    await this._voiceStarting;
+    if (!this.voiceLoop || !this.settings.microphoneEnabled) return;
+    if (this.voiceLoop.status?.asr) return;
+    if (!this.voiceLoop.asr) this.voiceLoop.asr = new QvacVoiceAdapter({ role: 'asr', settings: this.settings });
+    await this.voiceLoop.ensureAsr?.();
+  }
   cancel() { this._askGeneration += 1; this.voiceLoop?.interrupt?.(); this.harness.cancel(); this.scheduler.cancelQueued((job) => job.lane === 'voice'); this.computerRevoke(); this.voice.cancel(); this.setState('ARMED'); cancelQvac().catch((error) => this.emit('Error', 'QVAC_CANCEL', error.message)); }
   computerGrant(persist = false) { const result = this.computer.grant({ persist }); this.emit('ComputerStep', JSON.stringify({ action: 'grant', ...result })); if (this.settings.computerMode === 'observe') return result; this.input.grant({ persist }).then((backend) => { this.computer.setBackend(backend.backend); this.emit('ComputerStep', JSON.stringify({ action: 'backend', ...backend })); }).catch((error) => this.emit('Error', 'CU_GRANT', error.message)); return result; }
   computerRevoke() { this.input.revoke(); this.computer.revoke(); this.emit('ComputerStep', JSON.stringify({ action: 'revoke' })); }
@@ -182,8 +209,10 @@ export class JarvisDaemon extends EventEmitter {
     try { await loop.start(); this.emit('StateChanged', this.state); } catch (error) {
       this.voiceLoop = null; await loop.stop?.().catch(() => {}); this.emit('Error', 'VOICE_UNAVAILABLE', error.message); throw error;
     }
-    if (loop.status.tts) console.log('jarvisd: voice: TTS ready');
+    if (loop.status.wake) console.log('jarvisd: voice: wake detector ready');
-    else if (loop.status.errors.tts) console.error(`jarvisd: voice: TTS: ${loop.status.errors.tts}`);
+    else if (loop.status.errors.wake) console.error(`jarvisd: voice: wake: ${loop.status.errors.wake}`);
+    if (loop.status.capture) console.log('jarvisd: voice: microphone capture ready');
+    else if (loop.status.errors.capture) console.error(`jarvisd: voice: capture: ${loop.status.errors.capture}`);
   }
   async reloadSettings() {
     if (this._settingsReload) return this._settingsReload;
@@ -206,7 +235,7 @@ export class JarvisDaemon extends EventEmitter {
     Object.assign(this.observer.normalizer, { maxLongEdge: next.screenshotMaxEdge, quality: next.screenshotQuality });
     await this.startVoice();
     this.voice.cancel(); this.setState('ARMED');
-    return JSON.stringify({ applied: true, voice: this.voiceLoop.status, environmentOverrides: ['JARVIS_TTS_MODEL', 'JARVIS_ASR_MODEL', 'JARVIS_WAKE_COMMAND', 'JARVIS_QVAC_MODEL'].filter(key => process.env[key]), restartRequired: ['chatModel', 'modelProfile', 'maxTurns', 'maxShellCalls', 'maxToolRounds'].filter(key => next[key] !== this.startupSettings[key]) });
+    return JSON.stringify({ applied: true, voice: this.voiceLoop.status, environmentOverrides: ['JARVIS_TTS_MODEL', 'JARVIS_ASR_MODEL', 'JARVIS_WAKE_COMMAND', 'JARVIS_QVAC_MODEL'].filter(key => process.env[key]), restartRequired: ['chatModel', 'modelProfile', 'maxTurns', 'maxShellCalls', 'maxToolRounds', 'fsAccess'].filter(key => next[key] !== this.startupSettings[key]) });
   }
   async previewVoice(text) {
     if (this.locked) throw new Error('Unlock the desktop to preview a voice');

daemon/voice-doctor.js

@@ -1,5 +1,6 @@
 import { voiceSettings } from './voice-settings.js';
 import { ttsConfiguration } from './tts-config.js';
+import { isBuiltinWakeCommand } from './wake-engine.js';
 import { access } from 'node:fs/promises';
 
 const commandAvailable = async (command) => {
@@ -14,6 +15,7 @@ const report = {
   pipewireCapture: pipewire,
   wakeEngine: settings.listeningMode !== 'ptt' && Boolean(settings.command),
   wakeCommand: settings.command || null,
+  wakeBuiltin: settings.listeningMode !== 'ptt' && isBuiltinWakeCommand(settings.command),
   sampleRate: 16000,
   channels: 1,
   ttsPlayback: pipewire,

daemon/voice-loop.js

@@ -29,7 +29,7 @@ export class VoiceLoop extends EventEmitter {
     capture.on('close', () => { this.status.capture = false; if (this.running) { this.status.errors.capture = 'Microphone stream closed'; this._scheduleCaptureRetry(); } });
     wake.on('unavailable', () => { this.status.wake = false; });
     wake.on('error', (error) => { this.status.wake = false; this.emit('error', error); });
-    this.status = { asr: false, tts: false, capture: false, wake: false, errors: {} };
+    this.status = { asr: false, tts: false, capture: false, wake: false, microphone: this.asr != null, speech: this.tts != null, errors: {} };
     wake.on('wake', (phrase) => this.wakeHeard(phrase));
     vad.on('level', (rms) => daemon?.emit('ListeningLevel', rms));
     vad.on('utterance', (audio) => {
@@ -41,23 +41,10 @@ export class VoiceLoop extends EventEmitter {
 
   async start() {
     if (this.running) return;
-    for (const [name, adapter] of [['tts', this.tts], ['asr', this.asr]]) {
+    if (!this.tts) this.status.errors.tts = 'Spoken replies are disabled';
-      try {
-        if (adapter) { await adapter.start?.(); this.status[name] = true; }
-        else if (name === 'tts') this.status.errors.tts = 'Spoken replies are disabled';
-      } catch (error) {
-        const message = String(error?.message || error);
-        this.status.errors[name] = message;
-        this.emit('error', new Error(`${name.toUpperCase()}: ${message}`));
-      }
-    }
     this.running = true;
-    if (this.status.asr) {
-      this._armCapture();
     this._armWake();
-    } else if (this.asr) {
+    this._armCapture();
-      this._scheduleAsrRetry();
-    }
   }
   _notifyStatus() { try { this.daemon?.emit?.('StateChanged', this.daemon.state || 'ARMED'); } catch {} }
   _armWake() {
@@ -65,12 +52,12 @@ export class VoiceLoop extends EventEmitter {
     catch (error) { this.status.errors.wake = error.message; this.emit('error', error); }
   }
   _armCapture() {
-    if (!this.running || !this.status.asr || this.status.capture) return;
+    if (!this.running || this.status.capture || this.asr == null) return;
     try { this.capture.start(); this.status.capture = true; delete this.status.errors.capture; this._notifyStatus(); }
     catch (error) { this.status.errors.capture = error.message; this.emit('error', error); this._scheduleCaptureRetry(); }
   }
   _scheduleCaptureRetry() {
-    if (!this.running || this._captureRetry || this.status.capture || !this.status.asr) return;
+    if (!this.running || this._captureRetry || this.status.capture || this.asr == null) return;
     this._captureRetry = setTimeout(() => {
       this._captureRetry = 0;
       this._armCapture();
@@ -82,19 +69,65 @@ export class VoiceLoop extends EventEmitter {
     this._asrRetry = setTimeout(() => {
       this._asrRetry = 0;
       if (!this.running || this.status.asr || !this.asr) return;
-      this.asr.start().then(() => {
+      this.ensureAsr().catch(() => {});
-        if (!this.running || this.status.asr) return;
+    }, this.asrRetryMs);
+    this._asrRetry.unref?.();
+  }
+  async ensureAsr() {
+    if (!this.asr) return false;
+    if (this.status.asr) return true;
+    if (this._asrStarting) return this._asrStarting;
+    this._asrStarting = (async () => {
+      try {
+        await this.asr.start();
+        if (!this.running) return false;
         this.status.asr = true;
         delete this.status.errors.asr;
         this._armCapture();
         this._armWake();
         this._notifyStatus();
-      }).catch((error) => {
+        return true;
-        this.status.errors.asr = String(error?.message || error);
+      } catch (error) {
+        const message = String(error?.message || error);
+        this.status.errors.asr = message;
+        this.emit('error', new Error(`ASR: ${message}`));
         this._scheduleAsrRetry();
-      });
+        return false;
-    }, this.asrRetryMs);
+      } finally {
-    this._asrRetry.unref?.();
+        this._asrStarting = null;
+      }
+    })();
+    return this._asrStarting;
+  }
+  async ensureTts() {
+    if (!this.tts) { this.status.errors.tts = 'Spoken replies are disabled'; return false; }
+    if (this.status.tts) return true;
+    if (this._ttsStarting) return this._ttsStarting;
+    this._ttsStarting = (async () => {
+      try {
+        await this.tts.start();
+        this.status.tts = true;
+        delete this.status.errors.tts;
+        return true;
+      } catch (error) {
+        const message = String(error?.message || error);
+        this.status.errors.tts = message;
+        this.emit('error', new Error(`TTS: ${message}`));
+        return false;
+      } finally {
+        this._ttsStarting = null;
+      }
+    })();
+    return this._ttsStarting;
+  }
+  async parkModels() {
+    if (this._asrRetry) { clearTimeout(this._asrRetry); this._asrRetry = 0; }
+    this._asrStarting = null;
+    this._ttsStarting = null;
+    await this.asr?.stop?.();
+    if (this.tts && this.tts !== this.asr) await this.tts.stop?.();
+    this.status.asr = false;
+    this.status.tts = false;
   }
   async stop() {
     this.running = false;
@@ -102,17 +135,34 @@ export class VoiceLoop extends EventEmitter {
     if (this._asrRetry) { clearTimeout(this._asrRetry); this._asrRetry = 0; }
     this.interrupt(); this.capture.stop(); this.wake.close(); this.vad.reset(); this.playback.stop(); await Promise.allSettled([this._speechQueue, ...this._transcriptions]); await this.asr?.stop?.(); if (this.tts !== this.asr) await this.tts?.stop?.();
   }
-  setPushToTalk(pressed) { this.ptt = Boolean(pressed) && this.status.asr && this.status.capture; if (this.ptt) this.wakeHeard('push-to-talk'); else this.vad.end(); }
+  setPushToTalk(pressed) {
+    if (!pressed) {
+      this._pttHeld = false;
+      this.ptt = false;
+      this.vad.end();
+      return Promise.resolve();
+    }
+    this._pttHeld = true;
+    return this.ensureAsr().then((ok) => {
+      if (!this._pttHeld) return;
+      this.ptt = Boolean(ok && this.status.capture);
+      if (this.ptt) this.wakeHeard('push-to-talk');
+    }).catch((error) => this.emit('error', error));
+  }
   pushAudio(chunk) {
-    if (!this.running || this.daemon?.locked || this.daemon?.state === 'SLEEPING') return;
+    if (!this.running || this.daemon?.locked) return;
-    if (this.isSpeaking || this.now() < this.cooldownUntil) { this.metrics.feedbackDrop(); return; }
+    const sleeping = this.daemon?.state === 'SLEEPING';
+    if (!sleeping && (this.isSpeaking || this.now() < this.cooldownUntil)) { this.metrics.feedbackDrop(); return; }
+    if (!sleeping) {
       try { this.daemon?.emit('ListeningLevel', Math.min(1, Math.max(0, chunk?.length ? 0.01 : 0))); } catch {}
+    }
     if (!this.ptt && this.listeningMode !== 'ptt') this.wake.push(chunk);
+    if (sleeping) return;
     if (this.ptt || (this.listeningMode !== 'ptt' && this.daemon?.state === 'LISTENING')) this.vad.push(chunk);
   }
   wakeHeard(phrase) {
     if (!this.running || this.daemon?.locked) return;
-    this.metrics.wakeAccepted(); this.daemon?.emit('WakeHeard', phrase); this.daemon?.arm?.(); this.emit('wake', phrase);
+    this.metrics.wakeAccepted(); this.daemon?.emit('WakeHeard', phrase); this.ensureAsr().catch((error) => this.emit('error', error)); this.daemon?.arm?.(); this.emit('wake', phrase);
   }
   async transcribe(audio) {
     if (!this.status.asr || !this.asr?.transcribeAudio) return;
@@ -147,6 +197,7 @@ export class VoiceLoop extends EventEmitter {
   }
   async speak(text) {
     const generation = this._generation || 0;
+    if (this.tts && !this.status.tts && typeof this.tts.start === 'function') await this.ensureTts();
     if (!isSpeakable(text) || !this.tts?.speak || !this.status.tts) {
       this._releaseSpeaking();
       return;

daemon/wake-engine.js

@@ -1,12 +1,20 @@
 import { EventEmitter } from 'node:events';
 import { spawn } from 'node:child_process';
+import path from 'node:path';
+
+export const BUILTIN_WAKE_COMMAND = 'jarvis-wake-bridge';
+const FRAME_BYTES = 640;
+const SPEECH_RMS = 0.018;
+const SILENCE_FRAMES = 8;
+const MIN_SPEECH_FRAMES = 16;
+const MAX_SPEECH_FRAMES = 90;
+const WAKE_COOLDOWN_MS = 1500;
 
 /**
  * Wake engines consume PCM frames. A production wake model can be attached
- * through `detect(frame)`, keeping the daemon independent from a Python or
+ * through `detect(frame)` or a stdin bridge. The built-in CPU detector looks
- * native openWakeWord installation. The default detector is intentionally
+ * for a short two-to-four-peak cadence matching “hey jarvis” without loading
- * disabled until a local model command is configured; it never pretends that
+ * Whisper or any GPU model.
- * an energy spike is the wake phrase.
  */
 export class WakeEngine extends EventEmitter {
   constructor({ phrases = ['hey jarvis', 'jarvis', 'okay jarvis'], detect } = {}) {
@@ -31,6 +39,100 @@ export class WakeEngine extends EventEmitter {
   close() { this.pause(); this.removeAllListeners(); }
 }
 
+function frameRms(buf) {
+  const samples = buf.length / 2;
+  if (!samples) return 0;
+  let sum = 0;
+  for (let i = 0; i < buf.length; i += 2) {
+    const sample = buf.readInt16LE(i) / 32768;
+    sum += sample * sample;
+  }
+  return Math.sqrt(sum / samples);
+}
+
+function countPeaks(series) {
+  const smooth = series.map((value, index, items) => {
+    const prev = items[index - 1] ?? value;
+    const next = items[index + 1] ?? value;
+    return (prev + value + next) / 3;
+  });
+  const max = Math.max(0, ...smooth);
+  if (max < SPEECH_RMS) return 0;
+  const floor = Math.max(SPEECH_RMS * 1.2, max * 0.45);
+  let peaks = 0;
+  let raised = false;
+  for (let i = 1; i < smooth.length - 1; i++) {
+    if (smooth[i] >= floor && smooth[i] >= smooth[i - 1] && smooth[i] >= smooth[i + 1]) {
+      if (!raised) { peaks += 1; raised = true; }
+    } else if (smooth[i] < floor * 0.65) {
+      raised = false;
+    }
+  }
+  return peaks;
+}
+
+/** CPU keyword spotter. Keeps GPU ASR unloaded until a real wake. */
+export class KeywordWakeEngine extends WakeEngine {
+  constructor(options = {}) {
+    super(options);
+    this.command = options.command || BUILTIN_WAKE_COMMAND;
+    this.detect = (frame) => this._ingest(frame);
+    this._pending = Buffer.alloc(0);
+    this._rms = [];
+    this._silence = 0;
+    this._lastWake = 0;
+  }
+
+  start() { this.active = true; return this; }
+
+  push(frame) {
+    if (!this.active) return false;
+    return this._ingest(frame);
+  }
+
+  _ingest(frame) {
+    this._pending = Buffer.concat([this._pending, Buffer.from(frame)]);
+    let woke = false;
+    while (this._pending.length >= FRAME_BYTES) {
+      const next = this._pending.subarray(0, FRAME_BYTES);
+      this._pending = this._pending.subarray(FRAME_BYTES);
+      if (this._acceptFrame(next)) woke = true;
+    }
+    return woke;
+  }
+
+  _acceptFrame(frame) {
+    const rms = frameRms(frame);
+    if (rms >= SPEECH_RMS) {
+      this._rms.push(rms);
+      this._silence = 0;
+      if (this._rms.length > MAX_SPEECH_FRAMES + SILENCE_FRAMES) this._rms.shift();
+      return false;
+    }
+    if (!this._rms.length) return false;
+    this._rms.push(rms);
+    this._silence += 1;
+    if (this._silence < SILENCE_FRAMES) return false;
+    const series = this._rms;
+    this._rms = [];
+    this._silence = 0;
+    return this._maybeWake(series);
+  }
+
+  _maybeWake(series) {
+    const voiced = series.filter((value) => value >= SPEECH_RMS).length;
+    if (voiced < MIN_SPEECH_FRAMES || voiced > MAX_SPEECH_FRAMES) return false;
+    const peaks = countPeaks(series);
+    if (peaks < 2 || peaks > 4) return false;
+    const now = Date.now();
+    if (now - this._lastWake < WAKE_COOLDOWN_MS) return false;
+    const phrase = this.phrases[0] || 'hey jarvis';
+    this._lastWake = now;
+    this.emit('wake', phrase);
+    return true;
+  }
+}
+
 /** Adapter for a local openWakeWord/sherpa bridge. The bridge receives raw
  * PCM on stdin and prints one detected phrase per line on stdout. */
 export class ProcessWakeEngine extends WakeEngine {
@@ -58,8 +160,19 @@ export class ProcessWakeEngine extends WakeEngine {
   close() { super.close(); this.process?.kill('SIGTERM'); this.process = null; }
 }
 
+export function isBuiltinWakeCommand(command) {
+  const value = String(command || '').trim();
+  if (!value) return false;
+  const executable = value.split(/\s+/)[0];
+  const base = path.basename(executable);
+  return value === BUILTIN_WAKE_COMMAND || base === BUILTIN_WAKE_COMMAND || base === 'wake-bridge.js';
+}
+
 export function createWakeEngine(options = {}) {
-  return options.command || process.env.JARVIS_WAKE_COMMAND ? new ProcessWakeEngine({ ...options, command: options.command || process.env.JARVIS_WAKE_COMMAND }) : new WakeEngine(options);
+  const command = options.command || process.env.JARVIS_WAKE_COMMAND || '';
+  if (isBuiltinWakeCommand(command)) return new KeywordWakeEngine({ ...options, command: BUILTIN_WAKE_COMMAND });
+  if (command) return new ProcessWakeEngine({ ...options, command });
+  return new WakeEngine(options);
 }
 
 export function normalizeWakePhrases(value) {

docs/ROADMAP.md

@@ -173,9 +173,10 @@ runtime.
 Implementation is complete in `daemon/audio-pipewire.js`, `daemon/wake-engine.js`,
 `daemon/vad.js`, `daemon/voice-adapters.js`, `daemon/voice-loop.js`, and
 `daemon/audio-playback.js`. The SDK speech models are loaded as auxiliary
-models by the same QVAC master and use GPU settings. Live wake-word accuracy
+models by the same QVAC master and use GPU settings. Live wake-word detection
-requires a local detector bridge configured with `JARVIS_WAKE_COMMAND`; the
+uses the built-in CPU keyword spotter by default (`jarvis-wake-bridge`). A
-repository does not silently substitute cloud or CPU inference.
+custom openWakeWord or sherpa-onnx command can replace it; cloud wake is not
+supported.
 
 Exit gate: implementation complete. In a configured GNOME session, “Hey
 Jarvis” starts a local GPU-backed turn, speaks a response, and does not

docs/security-privacy.md

@@ -28,6 +28,13 @@ flowchart TD
 - QVAC binds to localhost; bearer tokens, when used, come from user-owned
   configuration and are not logged.
 
+## Filesystem access
+
+Path tools default to the Jarvis workspace (`~/.local/share/jarvis-qvac` when
+installed). Settings can widen that to your home directory or the entire
+filesystem. This is still your user account: Jarvis does not gain root, and
+writes plus shell commands still go through the confirmation gate.
+
 ## Computer-use controls
 
 Computer use requires a spoken or HUD grant, shows a visible cursor/target, and

docs/settings.md

@@ -112,9 +112,10 @@ and WebP quality tune observation detail and processing cost.
 
 - **Wake phrase** — `wakePhrase`, default `"hey jarvis"`. Must match a phrase supported by your local wake detector.
 - **Wake aliases** — `aliases`, default `["jarvis", "okay jarvis"]`. Additional detector phrases, separated by commas.
-- **Wake detector command** — `wakeCommand`, default `""`. Advanced: local program that receives microphone audio. Leave empty to use Hold Talk.
+- **Wake detector command** — `wakeCommand`, default `"jarvis-wake-bridge"`. Built-in CPU detector for Hey Jarvis. Replace it only if you have openWakeWord or sherpa. Empty disables the detector and uses Hold Talk.
 - **Listening behavior** — `listeningMode`, default `"conversation"`. Hold Talk only disables the wake detector and automatic follow-up listening. Choices: `conversation`, `single`, `ptt`.
 - **Idle sleep delay (minutes)** — `idleMinutes`, default `30`. Sleep after this much listening inactivity. Range: 1–240.
+- **Free GPU memory when idle** — `freeVramOnIdle`, default `true`. Unload speech and chat models after the idle delay. The microphone and Hey Jarvis detector stay ready on the CPU. The first reply after a long idle reloads models.
 
 ### Detection tuning
 
@@ -150,6 +151,7 @@ and WebP quality tune observation detail and processing cost.
 - **Maximum reasoning turns** — `maxTurns`, default `6`. Requires restart. Limits how long the assistant works on one request. Range: 1–30.
 - **Shell commands per request** — `maxShellCalls`, default `1`. Requires restart. Commands still require normal permissions. Range: 1–10.
 - **Tool rounds per request** — `maxToolRounds`, default `4`. Requires restart. Limits repeated tool use. Range: 1–20.
+- **File access** — `fsAccess`, default `"workspace"`. Requires restart. Limits `read_file`, `list_dir`, `grep`, and the `fs_*` tools. The agent still runs as your user account, not root, and asks before writing. Shell commands can already reach other paths. Choices: `workspace`, `home`, `filesystem`.
 
 ## GNOME appearance and boundaries
 

docs/troubleshooting.md

@@ -10,10 +10,11 @@ GPU.
 
 ## Voice doctor finds no wake engine
 
-Install a local openWakeWord or sherpa-onnx bridge and set
+Fresh installs use the built-in CPU detector (`jarvis-wake-bridge`). If
-`JARVIS_WAKE_COMMAND`. The bridge must read raw 16 kHz mono PCM from stdin and
+`listeningMode` is Hold Talk only, or `wakeCommand` is empty, the detector is
-write detected phrases, one per line, to stdout. Push-to-talk remains useful
+disabled. A custom openWakeWord or sherpa-onnx bridge must read raw 16 kHz mono
-for typed or manually triggered testing.
+PCM from stdin and write detected phrases, one per line, to stdout. Push-to-talk
+remains useful for typed or manually triggered testing.
 
 ## Portal consent succeeds but input is unavailable
 

docs/voice-acceptance.md

@@ -1,10 +1,10 @@
 # Phase 4 voice acceptance
 
-Run `npm run voice-doctor` in the target GNOME session. Set
+Run `npm run voice-doctor` in the target GNOME session. Fresh installs use the
-`JARVIS_WAKE_COMMAND` to the local openWakeWord or sherpa-onnx bridge command.
+built-in CPU detector (`jarvis-wake-bridge`). To use openWakeWord or sherpa-onnx
-The command receives raw 16 kHz mono PCM on stdin and emits one detected phrase
+instead, set `wakeCommand` or `JARVIS_WAKE_COMMAND` to a program that receives
-per line on stdout; the daemon passes it through `ProcessWakeEngine`. No cloud
+raw 16 kHz mono PCM on stdin and emits one detected phrase per line on stdout.
-wake service is supported.
+No cloud wake service is supported.
 
 The capture and playback nodes are both named `Jarvis`, so they can be routed
 in Helvum or qpwgraph. The loop gates capture while TTS is active and for 400ms

docs/voice-pipeline.md

@@ -31,9 +31,12 @@ stateDiagram-v2
   SLEEPING --> LISTENING: wake phrase
 ```
 
-The wake stage is intentionally separate from QVAC ASR. Configure a local
+The wake stage is intentionally separate from QVAC ASR. A built-in CPU detector
-openWakeWord or sherpa-onnx bridge through `JARVIS_WAKE_COMMAND`; no cloud wake
+listens for “hey jarvis” (and aliases) without loading Whisper. You can replace
-service is supported. During TTS, capture submission is gated and a 300–500 ms
+it with a local openWakeWord or sherpa-onnx bridge through `wakeCommand` /
+`JARVIS_WAKE_COMMAND`; no cloud wake service is supported. Microphone audio is
+still sent to the wake detector while Jarvis is sleeping so the phrase can
+restore listening. During TTS, capture submission is gated and a 300–500 ms
 cooldown prevents playback from becoming a new utterance. Trash transcripts and
 very short fragments are discarded.
 
@@ -45,16 +48,18 @@ replies. Raw audio is not persisted by default.
 
 ## Release voice readiness
 
-ASR and TTS initialize independently in the same QVAC runtime. Voice models do
+ASR, TTS, and the chat model stay unloaded until wake, Hold Talk, or a typed
-not require the chat model to load first. The runtime status includes separate
+question. Idle sleep unloads them again so GPU memory is freed; only PipeWire
-`asr`, `tts`, `capture`, and `wake` flags plus per-component errors. Without a
+capture and the CPU wake detector stay resident. Runtime status includes
-wake bridge, use Hold Talk (mouse, Space, or Enter) in the GNOME overlay.
+separate `asr`, `tts`, `capture`, and `wake` flags plus per-component errors.
+The tray shows **WAKE ON** when the detector is alive. Hold Talk (mouse, Space,
+or Enter) still works in the GNOME overlay.
 
 The Control Center's wake command and spoken-reply setting are read from
 `~/.config/jarvis/config.json` on daemon startup. Restart `jarvisd.service` after
-changing them. `JARVIS_WAKE_COMMAND` overrides the saved command. A wake command
+changing them. `JARVIS_WAKE_COMMAND` overrides the saved command. A custom wake
-must consume 16 kHz mono signed PCM and emit detected phrases on stdout; simply
+command must consume 16 kHz mono signed PCM and emit detected phrases on stdout.
-setting a phrase does not install a wake detector.
+The default `jarvis-wake-bridge` value selects the built-in CPU detector.
 
 QVAC TTS numeric arrays contain signed 16-bit samples, not bytes. Supertonic
 plays at 44.1 kHz. Captured utterances use the non-streaming transcription API

packaging/install.sh

@@ -200,11 +200,14 @@ else
   cat > "${CONFIG_DIR}/config.json" <<'EOF'
 {
   "wakePhrase": "hey jarvis",
+  "wakeCommand": "jarvis-wake-bridge",
   "modelProfile": "laptop-16gb",
-  "ttsEnabled": true
+  "ttsEnabled": true,
+  "fsAccess": "workspace",
+  "freeVramOnIdle": true
 }
 EOF
-  log "created ${CONFIG_DIR}/config.json with default wake phrase, laptop-16gb profile, and TTS enabled"
+  log "created ${CONFIG_DIR}/config.json with default wake phrase, CPU wake detector, laptop-16gb profile, and TTS enabled"
 fi
 if [[ ! -f "${CONFIG_DIR}/qvac.config.json" ]]; then
   sed "s#__JARVIS_CACHE__#${CACHE_DIR}#g" "${ROOT_DIR}/packaging/qvac.config.template.json" > "${CONFIG_DIR}/qvac.config.json"

packaging/wake-bridge.js

@@ -0,0 +1,8 @@
+#!/usr/bin/env node
+import { KeywordWakeEngine, normalizeWakePhrases } from '../daemon/wake-engine.js';
+
+const phrases = normalizeWakePhrases(process.env.JARVIS_WAKE_PHRASES || process.argv.slice(2).join(',') || 'hey jarvis,jarvis,okay jarvis');
+const engine = new KeywordWakeEngine({ phrases });
+engine.on('wake', (phrase) => process.stdout.write(`${phrase}\n`));
+process.stdin.on('data', (chunk) => engine.push(chunk));
+process.stdin.on('end', () => engine.close());

skills/phase2-tools.js

@@ -6,10 +6,21 @@ import { qvacStatus } from '../daemon/qvac-master.js';
 
 const PERMISSIONS = Object.freeze({ read: 'read', write: 'write', dangerous: 'dangerous', computerUse: 'computer-use' });
 
+export function filesystemRoots(fsAccess, cwd = process.cwd()) {
+  if (fsAccess === 'filesystem') return ['/'];
+  if (fsAccess === 'home') return [os.homedir()];
+  return [path.resolve(cwd)];
+}
+
+function isInside(root, candidate) {
+  const relative = path.relative(root, candidate);
+  return relative !== '..' && !relative.startsWith(`..${path.sep}`) && !path.isAbsolute(relative);
+}
+
 // Resolve symlinks as well as lexical paths before accessing workspace files.
-async function workspacePath(cwd, file, { write = false } = {}) {
+export async function workspacePath(cwd, file, { write = false, roots } = {}) {
-  const root = await realpath(path.resolve(cwd));
+  const allow = (roots?.length ? roots : [cwd]).map((root) => path.resolve(root));
-  const target = path.resolve(cwd, file);
+  const target = path.isAbsolute(file) ? path.resolve(file) : path.resolve(cwd, file);
   let resolved;
   try { resolved = await realpath(target); }
   catch (error) {
@@ -18,8 +29,10 @@ async function workspacePath(cwd, file, { write = false } = {}) {
     if (entry?.isSymbolicLink()) throw new Error('path is outside the Jarvis workspace or is a dangling symlink');
     resolved = path.join(await realpath(path.dirname(target)), path.basename(target));
   }
-  const relative = path.relative(root, resolved);
+  const resolvedRoots = await Promise.all(allow.map(async (root) => {
-  if (relative === '..' || relative.startsWith(`..${path.sep}`) || path.isAbsolute(relative)) throw new Error('path is outside the Jarvis workspace');
+    try { return await realpath(root); } catch { return root; }
+  }));
+  if (!resolvedRoots.some((root) => isInside(root, resolved))) throw new Error('path is outside the Jarvis workspace');
   return resolved;
 }
 
@@ -60,7 +73,9 @@ async function searchFiles(root, query, limit = 50) {
   return hits;
 }
 
-export function createPhase2Tools({ cwd = process.cwd(), computer } = {}) {
+export function createPhase2Tools({ cwd = process.cwd(), computer, roots } = {}) {
+  const allow = roots?.length ? roots : [cwd];
+  const bound = (file, options) => workspacePath(cwd, file, { ...options, roots: allow });
   return [
     {
       name: 'app_list', permission: PERMISSIONS.read,
@@ -69,21 +84,21 @@ export function createPhase2Tools({ cwd = process.cwd(), computer } = {}) {
     },
     {
       name: 'fs_search', permission: PERMISSIONS.read,
-      description: 'Search local file and directory names below the Jarvis workspace.',
+      description: 'Search local file and directory names inside the allowed filesystem roots.',
       parameters: { type: 'object', properties: { query: { type: 'string' }, root: { type: 'string' }, limit: { type: 'number' } }, required: ['query'] },
-      execute: async ({ query, root = cwd, limit = 50 }) => searchFiles(await workspacePath(cwd, root), query, Math.min(100, Number(limit) || 50)),
+      execute: async ({ query, root = cwd, limit = 50 }) => searchFiles(await bound(root), query, Math.min(100, Number(limit) || 50)),
     },
     {
       name: 'fs_read', permission: PERMISSIONS.read,
-      description: 'Read a UTF-8 local text file below the Jarvis workspace, capped at 400 KiB.',
+      description: 'Read a UTF-8 local text file inside the allowed filesystem roots, capped at 400 KiB.',
       parameters: { type: 'object', properties: { file: { type: 'string' } }, required: ['file'] },
-      execute: async ({ file }) => { const target = await workspacePath(cwd, file); return (await readFile(target, 'utf8')).slice(0, 400 * 1024); },
+      execute: async ({ file }) => { const target = await bound(file); return (await readFile(target, 'utf8')).slice(0, 400 * 1024); },
     },
     {
       name: 'fs_write', permission: PERMISSIONS.write,
       description: 'Write a local text file only after an explicit confirmation flag is supplied.',
       parameters: { type: 'object', properties: { file: { type: 'string' }, contents: { type: 'string' }, confirmed: { type: 'boolean' } }, required: ['file', 'contents', 'confirmed'] },
-      execute: async ({ file, contents, confirmed }) => { if (confirmed !== true) return { confirmation_required: true, action: 'write', file }; const target = await workspacePath(cwd, file, { write: true }); await writeFile(target, String(contents), 'utf8'); return { ok: true, file: target, bytes: Buffer.byteLength(String(contents)) }; },
+      execute: async ({ file, contents, confirmed }) => { if (confirmed !== true) return { confirmation_required: true, action: 'write', file }; const target = await bound(file, { write: true }); await writeFile(target, String(contents), 'utf8'); return { ok: true, file: target, bytes: Buffer.byteLength(String(contents)) }; },
     },
     {
       name: 'memory_recall', permission: PERMISSIONS.read,

test/daemon.test.js

@@ -7,7 +7,7 @@ import assert from 'node:assert/strict';
 import { VoiceStateMachine } from '../daemon/voice-state.js';
 import { QvacScheduler } from '../daemon/qvac-scheduler.js';
 import { JarvisDaemon } from '../daemon/index.js';
-import { HarnessBridge } from '../daemon/harness-bridge.js';
+import { HarnessBridge, harnessRoots } from '../daemon/harness-bridge.js';
 import { spokenReply } from '../skills/voice-prompt.js';
 import { voiceSettings } from '../daemon/voice-settings.js';
 
@@ -67,11 +67,13 @@ test('ask extracts harness reply text instead of stringifying the object', async
 });
 
 test('harness bridge caps voice shell chaining', () => {
-  const bridge = new HarnessBridge();
+  const bridge = new HarnessBridge({ fsAccess: 'workspace' });
   assert.equal(bridge.options.origin, 'jarvis-qvac');
   assert.equal(bridge.options.voice, true);
   assert.equal(bridge.options.maxShellCalls, 1);
   assert.equal(bridge.options.maxTurns, 6);
+  assert.deepEqual(bridge.options.roots, []);
+  assert.deepEqual(harnessRoots('filesystem'), ['/']);
   assert.deepEqual(bridge.options.builtinTools, [
     'read_file',
     'list_dir',
@@ -150,6 +152,20 @@ test('QVAC scheduler prioritizes voice and keeps one active job', async () => {
   assert.deepEqual(scheduler.status(), { running: 0, queued: [] });
 });
 
+test('idle sleep unloads GPU models instead of only suspending', async () => {
+  const daemon = new JarvisDaemon();
+  const calls = [];
+  daemon.settings = { ...daemon.settings, freeVramOnIdle: true };
+  daemon.parkModels = async () => { calls.push('park'); };
+  daemon.voiceLoop = { interrupt() {} };
+  daemon.harness = { cancel() {}, resetContext: async () => {}, close: async () => {} };
+  try {
+    await daemon.sleep();
+    assert.equal(daemon.state, 'SLEEPING');
+    assert.deepEqual(calls, ['park']);
+  } finally { await daemon.close(); }
+});
+
 test('shutdown still closes the harness when recovery or voice cleanup fails', async () => {
   const daemon = new JarvisDaemon(); let closed = false;
   daemon.recovery.save = () => { throw new Error('disk full'); };

test/gnome-extension.test.js

@@ -255,6 +255,21 @@ test('session panel stays closed while Jarvis speaks unless expanded', () => {
   assert.match(session.view.transcript.children[0].text, /still talking/);
 });
 
+test('lazy ASR still shows the microphone as available when capture is up', () => {
+  const { ConversationView } = harness();
+  const popup = new ConversationView({ compact: true });
+  popup.setVoiceStatus({ tts: true, input: true, wake: true });
+  assert.match(popup.status.text, /SPEECH ON/);
+  assert.match(popup.status.text, /WAKE ON/);
+  assert.doesNotMatch(popup.status.text, /MIC UNAVAILABLE/);
+  assert.equal(popup.talk.label, 'Hold to talk');
+  popup.setVoiceStatus({ tts: false, input: false, wake: false });
+  assert.match(popup.status.text, /MIC UNAVAILABLE/);
+  assert.equal(popup.talk.label, 'Mic unavailable');
+  assert.match(extensionSource, /Boolean\(voice\.capture\)/);
+  assert.doesNotMatch(extensionSource, /voice\.asr && voice\.capture/);
+});
+
 test('voice status lives in the panel chip, not a floating overlay', () => {
   const { JarvisOsd, SessionPanel, timers, chrome } = harness();
   const osd = new JarvisOsd();

test/install.test.js

@@ -167,8 +167,11 @@ test('install.sh seeds default config when missing and does not require first-ru
 
     const config = JSON.parse(await readFile(path.join(home, '.config/jarvis/config.json'), 'utf8'));
     assert.equal(config.wakePhrase, 'hey jarvis');
+    assert.equal(config.wakeCommand, 'jarvis-wake-bridge');
     assert.equal(config.modelProfile, 'laptop-16gb');
     assert.equal(config.ttsEnabled, true);
+    assert.equal(config.fsAccess, 'workspace');
+    assert.equal(config.freeVramOnIdle, true);
     assert.match(result.stdout, /Jarvis is ready/);
     assert.doesNotMatch(result.stdout, /first-run/);
     const installSource = readFileSync(new URL('../packaging/install.sh', import.meta.url), 'utf8');

test/review-regressions.test.js

@@ -1,19 +1,20 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import { EventEmitter } from 'node:events';
-import { mkdtemp, mkdir, writeFile, readFile, symlink } from 'node:fs/promises';
+import { mkdtemp, mkdir, writeFile, readFile, symlink, rm } from 'node:fs/promises';
-import { tmpdir } from 'node:os';
+import os, { tmpdir } from 'node:os';
 import path from 'node:path';
 import { createRequire } from 'node:module';
 import { ComputerUseSession } from '../computer-use/session.js';
 import { ComputerActuator } from '../computer-use/actuator.js';
 import { PortalInputBackend } from '../computer-use/portal-input.js';
 import { createComputerObserveTools } from '../skills/computer-observe.js';
-import { createPhase2Tools } from '../skills/phase2-tools.js';
+import { createPhase2Tools, filesystemRoots } from '../skills/phase2-tools.js';
 import { VoiceStateMachine } from '../daemon/voice-state.js';
 import { assertLocalEndpoint } from '../daemon/network-policy.js';
 const require = createRequire(import.meta.url);
 const custom = require('../vendor/agent-harness/agent/custom-tools.js');
+const sandbox = require('../vendor/agent-harness/agent/sandbox.js');
 
 test('custom tool permissions survive registration and default to confirmation', () => {
   const id = 'review-permissions';
@@ -29,6 +30,28 @@ test('custom tool permissions survive registration and default to confirmation',
   } finally { custom.clear(id); }
 });
 
+test('filesystem access lets path tools reach home while workspace stays jailed', async () => {
+  const dir = await mkdtemp(path.join(tmpdir(), 'jarvis-fs-access-'));
+  const workspace = path.join(dir, 'workspace'); await mkdir(workspace);
+  const homeProbe = await mkdtemp(path.join(os.homedir(), '.jarvis-fs-test-'));
+  try {
+    await writeFile(path.join(homeProbe, 'note.txt'), 'from-home');
+    const jailed = Object.fromEntries(createPhase2Tools({ cwd: workspace, roots: filesystemRoots('workspace', workspace) }).map(t => [t.name, t]));
+    await assert.rejects(jailed.fs_read.execute({ file: path.join(homeProbe, 'note.txt') }), /outside/);
+    const opened = Object.fromEntries(createPhase2Tools({ cwd: workspace, roots: filesystemRoots('filesystem', workspace) }).map(t => [t.name, t]));
+    assert.equal(await opened.fs_read.execute({ file: path.join(homeProbe, 'note.txt') }), 'from-home');
+    const home = Object.fromEntries(createPhase2Tools({ cwd: workspace, roots: filesystemRoots('home', workspace) }).map(t => [t.name, t]));
+    assert.equal(await home.fs_read.execute({ file: path.join(homeProbe, 'note.txt') }), 'from-home');
+    sandbox.setGrants({ roots: [path.resolve(workspace)] });
+    assert.equal(sandbox.isAllowed('jarvis-qvac', path.join(homeProbe, 'note.txt')), false);
+    sandbox.setGrants({ roots: [path.resolve(workspace), '/'] });
+    assert.equal(sandbox.isAllowed('jarvis-qvac', path.join(homeProbe, 'note.txt')), true);
+  } finally {
+    await rm(homeProbe, { recursive: true, force: true });
+    await rm(dir, { recursive: true, force: true });
+  }
+});
+
 test('workspace tools reject outside roots and symlink escapes and count UTF-8 bytes', async () => {
   const dir = await mkdtemp(path.join(tmpdir(), 'jarvis-path-test-'));
   const root = path.join(dir, 'workspace'); await mkdir(root);

test/settings.test.js

@@ -171,5 +171,13 @@ test('a disabled microphone never starts capture while speech output remains ava
   const capture = new EventEmitter(); capture.start = () => assert.fail('microphone should be disabled');
   const loop = new VoiceLoop({ capture, asr: null, tts: { start: async () => {} } });
   await loop.start();
-  assert.equal(loop.status.capture, false); assert.equal(loop.status.asr, false); assert.equal(loop.status.tts, true);
+  assert.equal(loop.status.capture, false); assert.equal(loop.status.asr, false); assert.equal(loop.status.tts, false);
+  assert.equal(await loop.ensureTts(), true); assert.equal(loop.status.tts, true);
+});
+
+test('catalog defaults enable CPU wake, workspace files, and idle VRAM unload', () => {
+  const settings = voiceSettings({});
+  assert.equal(settings.fsAccess, 'workspace');
+  assert.equal(settings.freeVramOnIdle, true);
+  assert.equal(settings.wakeCommand, 'jarvis-wake-bridge');
 });

test/voice-phase4.test.js

@@ -117,7 +117,7 @@ test('PCM packing uses even s16le sample pairs', () => {
   assert.equal(fromBytes.samples[0], 256);
 });
 
-test('ASR failure preserves speech output and reports unavailable microphone', async () => {
+test('ASR failure keeps the microphone hot and still allows speech output', async () => {
   const capture = new EventEmitter(); let captured = false;
   capture.start = () => { captured = true; }; capture.stop = () => {};
   let spoken = '';
@@ -127,12 +127,14 @@ test('ASR failure preserves speech output and reports unavailable microphone', a
     playback: { play: async () => {}, stop() {} },
   });
   loop.on('error', () => {});
-  await loop.start(); await loop.speak('Speech still works.');
+  await loop.start();
-  assert.equal(spoken, 'Speech still works.'); assert.equal(captured, false);
+  assert.equal(captured, true);
-  assert.equal(loop.status.tts, true); assert.equal(loop.status.asr, false);
+  assert.equal(loop.status.asr, false);
-  spoken = '';
+  assert.equal(loop.status.tts, false);
-  await loop.speak('OK.');
+  await loop.speak('Speech still works.');
-  assert.equal(spoken, 'OK.');
+  assert.equal(spoken, 'Speech still works.');
+  assert.equal(loop.status.tts, true);
+  assert.equal(await loop.ensureAsr(), false);
   assert.match(loop.status.errors.asr, /Whisper/); await loop.stop();
 });
 
@@ -142,8 +144,11 @@ test('TTS failure preserves ASR and push to talk without wake command', async ()
     asr: { start: async () => {} }, tts: { start: async () => { throw new Error('TTS failed'); } },
     playback: { stop() {} },
   });
-  loop.on('error', () => {}); await loop.start(); loop.setPushToTalk(true);
+  loop.on('error', () => {}); await loop.start();
+  assert.equal(loop.status.asr, false);
+  await loop.setPushToTalk(true);
   assert.equal(loop.status.asr, true); assert.equal(loop.ptt, true);
+  assert.equal(await loop.ensureTts(), false);
   assert.equal(loop.status.tts, false); assert.equal(loop.status.wake, false); await loop.stop();
 });
 
@@ -169,7 +174,7 @@ test('voice loop retries microphone capture after PipeWire comes up', async () =
   });
   loop.on('error', () => {});
   await loop.start();
-  assert.equal(loop.status.asr, true);
+  assert.equal(loop.status.asr, false);
   await new Promise((resolve) => setTimeout(resolve, 5));
   assert.equal(loop.status.capture, false);
   await new Promise((resolve) => setTimeout(resolve, 40));
@@ -194,7 +199,10 @@ test('voice loop retries ASR after a first-login GPU miss', async () => {
   loop.on('error', () => {});
   await loop.start();
   assert.equal(loop.status.asr, false);
-  assert.equal(capture.started, undefined);
+  assert.equal(capture.started, true);
+  assert.equal(attempts, 0);
+  assert.equal(await loop.ensureAsr(), false);
+  assert.equal(attempts, 1);
   await new Promise((resolve) => setTimeout(resolve, 50));
   assert.equal(attempts, 2);
   assert.equal(loop.status.asr, true);
@@ -207,3 +215,61 @@ test('voice adapters load canonical ASR and TTS plugins', () => {
   assert.match(source, /whispercpp-transcription/);
   assert.match(source, /tts-ggml/);
 });
+
+function tonePcm(ms, amplitude = 12_000, hz = 220) {
+  const samples = Math.floor(16_000 * ms / 1000);
+  const buf = Buffer.alloc(samples * 2);
+  for (let i = 0; i < samples; i++) buf.writeInt16LE(Math.floor(amplitude * Math.sin(2 * Math.PI * hz * i / 16_000)), i * 2);
+  return buf;
+}
+
+test('built-in CPU wake detector fires on a two-peak hey-jarvis cadence', async () => {
+  const { KeywordWakeEngine, createWakeEngine } = await import('../daemon/wake-engine.js');
+  const engine = createWakeEngine({ command: 'jarvis-wake-bridge', phrases: ['hey jarvis'] });
+  assert.equal(engine instanceof KeywordWakeEngine, true);
+  const heard = [];
+  engine.on('wake', (phrase) => heard.push(phrase));
+  engine.push(Buffer.concat([tonePcm(80, 0), tonePcm(220), tonePcm(90, 0), tonePcm(280), tonePcm(200, 0)]));
+  assert.deepEqual(heard, ['hey jarvis']);
+  engine.push(tonePcm(800));
+  assert.equal(heard.length, 1);
+});
+
+test('sleeping voice loop still feeds the wake detector and skips VAD', () => {
+  const capture = new EventEmitter(); capture.start = () => {}; capture.stop = () => {};
+  let wakeFrames = 0; let vadFrames = 0;
+  const wake = new WakeEngine({ detect: () => null });
+  wake.push = () => { wakeFrames += 1; };
+  const vad = new VadSegmenter();
+  vad.push = () => { vadFrames += 1; };
+  const daemon = new EventEmitter(); daemon.state = 'SLEEPING';
+  const loop = new VoiceLoop({ daemon, capture, wake, vad });
+  loop.running = true;
+  loop.pushAudio(Buffer.alloc(640));
+  assert.equal(wakeFrames, 1);
+  assert.equal(vadFrames, 0);
+});
+
+test('voice start keeps the microphone hot without loading ASR or TTS', async () => {
+  const capture = new EventEmitter(); let captured = false;
+  capture.start = () => { captured = true; }; capture.stop = () => {};
+  let asrStarts = 0; let ttsStarts = 0;
+  const loop = new VoiceLoop({
+    capture,
+    asr: { start: async () => { asrStarts += 1; } },
+    tts: { start: async () => { ttsStarts += 1; } },
+    wake: new WakeEngine({ detect: () => null }),
+  });
+  await loop.start();
+  assert.equal(captured, true);
+  assert.equal(loop.status.capture, true);
+  assert.equal(loop.status.asr, false);
+  assert.equal(loop.status.tts, false);
+  assert.equal(asrStarts, 0);
+  assert.equal(ttsStarts, 0);
+  assert.equal(loop.status.wake, true);
+  await loop.parkModels();
+  assert.equal(loop.status.asr, false);
+  assert.equal(loop.status.tts, false);
+  await loop.stop();
+});