feat(voice): chat UX polish — busy-aware silence, stop hint, thinking sounds, barge-in fix

apps/desktop/src/app/chat/composer/hooks/use-composer-voice.ts

@@ -7,8 +7,8 @@ import { triggerHaptic } from '@/lib/haptics'
 import { $voiceConversationStartRequest, takeVoiceConversationStart } from '@/store/composer'
 import { resetBrowseState } from '@/store/composer-input-history'
 import { $gateway } from '@/store/gateway'
-import { notifyError } from '@/store/notifications'
-import { $autoSpeakReplies, setAutoSpeakReplies } from '@/store/voice-prefs'
+import { notify, notifyError } from '@/store/notifications'
+import { $autoSpeakReplies, $voiceStopPhrase, setAutoSpeakReplies } from '@/store/voice-prefs'
 import { resumeWakeAfterVoice } from '@/store/wake-word'
 
 import type { ComposerTarget } from '../focus'
@@ -208,6 +208,26 @@ export function useComposerVoice({
     }
   }, [pauseWakeForVoice, resumeWakeIfPaused, voiceConversationActive])
 
+  // 'Say "stop" to end the voice chat.' notice when the conversation starts.
+  // Phrase comes from voice.stop_phrases (first entry) so a custom phrase
+  // renders correctly; a null phrase (stop_phrases: []) shows no notice.
+  useEffect(() => {
+    if (!voiceConversationActive) {
+      return
+    }
+
+    const phrase = $voiceStopPhrase.get()
+
+    if (phrase) {
+      notify({
+        id: 'voice-stop-hint',
+        kind: 'info',
+        icon: 'mic',
+        message: t.notifications.voice.sayStopToEnd(phrase)
+      })
+    }
+  }, [t, voiceConversationActive])
+
   useEffect(() => resumeWakeIfPaused, [resumeWakeIfPaused])
 
   // Explicit start/end for the on-screen conversation controls (the hotkey uses

apps/desktop/src/app/chat/composer/hooks/use-voice-conversation.ts

@@ -1,6 +1,7 @@
 import { useCallback, useEffect, useRef, useState } from 'react'
 
 import { useI18n } from '@/i18n'
+import { startThinkingSound, stopThinkingSound } from '@/lib/thinking-sound'
 import { monitorSpeechDuringPlayback } from '@/lib/voice-barge-in'
 import {
   markVoicePlaybackInterrupted,
@@ -574,6 +575,22 @@ export function useVoiceConversation({
     return () => window.removeEventListener('keydown', onKeyDown, { capture: true })
   }, [enabled, stopTurn])
 
+  // Ambient "thinking" sound: while the agent works (status 'thinking') no
+  // audio flows, which reads as dead air mid-conversation. Calm bubble blips
+  // fill the gap; they stop the INSTANT speech starts, the mic re-arms, or the
+  // conversation ends. Gated by voice.thinking_sound + the shared sound mute.
+  useEffect(() => {
+    if (enabled && !muted && status === 'thinking') {
+      startThinkingSound()
+
+      return stopThinkingSound
+    }
+
+    stopThinkingSound()
+
+    return undefined
+  }, [enabled, muted, status])
+
   // Drive the loop: when a voice-submitted reply appears, open a live speech
   // session (which feeds itself from then on). Otherwise start listening when
   // idle between turns.

apps/desktop/src/app/session/hooks/use-hermes-config.ts

@@ -14,7 +14,7 @@ import {
   setDefaultReasoningEffort,
   setIntroPersonality
 } from '@/store/session'
-import { applyAutoSpeakFromConfig } from '@/store/voice-prefs'
+import { applyAutoSpeakFromConfig, applyThinkingSoundFromConfig, applyVoiceStopPhraseFromConfig } from '@/store/voice-prefs'
 
 const DEFAULT_VOICE_SECONDS = 120
 const FAST_TIERS = new Set(['fast', 'priority', 'on'])
@@ -105,6 +105,8 @@ export function useHermesConfig({ activeSessionIdRef }: HermesConfigOptions) {
         setVoiceMaxRecordingSeconds(recordingLimit(config.voice?.max_recording_seconds))
         setSttEnabled(config.stt?.enabled !== false)
         applyAutoSpeakFromConfig(config)
+        applyVoiceStopPhraseFromConfig(config)
+        applyThinkingSoundFromConfig(config)
       } catch {
         // Config is nice-to-have; chat still works without it.
       }

apps/desktop/src/i18n/ar.ts

@@ -137,6 +137,7 @@ export const ar = defineLocale({
       noSpeechDetected: 'لم يتم اكتشاف كلام',
       playbackFailed: 'فشل تشغيل الصوت',
       recordingFailed: 'فشل التسجيل',
+      sayStopToEnd: phrase => `قل "${phrase}" لإنهاء المحادثة الصوتية.`,
       transcriptionFailed: 'فشل التفريغ النصي',
       transcriptionUnavailable: 'التفريغ النصي غير متاح.',
       tryRecordingAgain: 'حاول التسجيل مرة أخرى.',

apps/desktop/src/i18n/en.ts

@@ -154,6 +154,7 @@ export const en: Translations = {
       noSpeechDetected: 'No speech detected',
       playbackFailed: 'Voice playback failed',
       recordingFailed: 'Voice recording failed',
+      sayStopToEnd: phrase => `Say "${phrase}" to end the voice chat.`,
       transcriptionFailed: 'Voice transcription failed',
       transcriptionUnavailable: 'Voice transcription is not available yet.',
       tryRecordingAgain: 'Try recording again.',

apps/desktop/src/i18n/ja.ts

@@ -155,6 +155,7 @@ export const ja = defineLocale({
       noSpeechDetected: '音声が検出されませんでした',
       playbackFailed: '音声再生に失敗しました',
       recordingFailed: '音声録音に失敗しました',
+      sayStopToEnd: phrase => `「${phrase}」と言うと音声チャットを終了できます。`,
       transcriptionFailed: '音声文字起こしに失敗しました',
       transcriptionUnavailable: '音声文字起こしはまだ利用できません。',
       tryRecordingAgain: 'もう一度録音してください。',

apps/desktop/src/i18n/types.ts

@@ -195,6 +195,7 @@ export interface Translations {
       noSpeechDetected: string
       playbackFailed: string
       recordingFailed: string
+      sayStopToEnd: (phrase: string) => string
       transcriptionFailed: string
       transcriptionUnavailable: string
       tryRecordingAgain: string

apps/desktop/src/i18n/zh-hant.ts

@@ -150,6 +150,7 @@ export const zhHant = defineLocale({
       noSpeechDetected: '未偵測到語音',
       playbackFailed: '語音播放失敗',
       recordingFailed: '語音錄製失敗',
+      sayStopToEnd: phrase => `說「${phrase}」即可結束語音對話。`,
       transcriptionFailed: '語音轉寫失敗',
       transcriptionUnavailable: '語音轉寫暫不可用。',
       tryRecordingAgain: '請再錄製一次。',

apps/desktop/src/i18n/zh.ts

@@ -150,6 +150,7 @@ export const zh: Translations = {
       noSpeechDetected: '没有检测到语音',
       playbackFailed: '语音播放失败',
       recordingFailed: '语音录制失败',
+      sayStopToEnd: phrase => `说“${phrase}”即可结束语音对话。`,
       transcriptionFailed: '语音转写失败',
       transcriptionUnavailable: '语音转写暂不可用。',
       tryRecordingAgain: '请再录一次。',

apps/desktop/src/lib/thinking-sound.test.ts

@@ -0,0 +1,118 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+vi.mock('@/store/haptics', () => ({ $hapticsMuted: { get: vi.fn(() => false) } }))
+vi.mock('@/hermes', () => ({
+  getHermesConfigRecord: vi.fn(async () => ({})),
+  saveHermesConfig: vi.fn(async () => undefined)
+}))
+
+import { $hapticsMuted } from '@/store/haptics'
+import { $thinkingSoundEnabled } from '@/store/voice-prefs'
+
+import { isThinkingSoundActive, startThinkingSound, stopThinkingSound } from './thinking-sound'
+
+class FakeOscillator {
+  type = 'sine'
+  frequency = { exponentialRampToValueAtTime: vi.fn(), setValueAtTime: vi.fn() }
+  connect = vi.fn()
+  start = vi.fn()
+  stop = vi.fn()
+}
+
+class FakeGain {
+  gain = { exponentialRampToValueAtTime: vi.fn(), setValueAtTime: vi.fn() }
+  connect = vi.fn()
+}
+
+const started: FakeOscillator[] = []
+
+class FakeAudioContext {
+  currentTime = 0
+  destination = {}
+  state = 'running'
+
+  createOscillator() {
+    const osc = new FakeOscillator()
+
+    started.push(osc)
+
+    return osc
+  }
+
+  createGain() {
+    return new FakeGain()
+  }
+
+  resume() {
+    return Promise.resolve()
+  }
+}
+
+describe('thinking-sound', () => {
+  beforeEach(() => {
+    vi.useFakeTimers()
+    started.length = 0
+    $thinkingSoundEnabled.set(true)
+    vi.mocked($hapticsMuted.get).mockReturnValue(false)
+    vi.stubGlobal('AudioContext', FakeAudioContext)
+  })
+
+  afterEach(() => {
+    stopThinkingSound()
+    vi.unstubAllGlobals()
+    vi.useRealTimers()
+  })
+
+  it('plays repeating blips while active and stops instantly', () => {
+    startThinkingSound()
+    expect(isThinkingSoundActive()).toBe(true)
+
+    vi.advanceTimersByTime(5_000)
+    expect(started.length).toBeGreaterThanOrEqual(3)
+
+    const count = started.length
+
+    stopThinkingSound()
+    expect(isThinkingSoundActive()).toBe(false)
+    vi.advanceTimersByTime(5_000)
+    expect(started.length).toBe(count) // nothing after stop
+  })
+
+  it('respects the voice.thinking_sound config gate', () => {
+    $thinkingSoundEnabled.set(false)
+    startThinkingSound()
+    expect(isThinkingSoundActive()).toBe(false)
+    vi.advanceTimersByTime(3_000)
+    expect(started.length).toBe(0)
+  })
+
+  it('stays silent while sounds are muted but keeps the loop alive', () => {
+    vi.mocked($hapticsMuted.get).mockReturnValue(true)
+    startThinkingSound()
+    vi.advanceTimersByTime(3_000)
+    expect(started.length).toBe(0)
+
+    // Unmute mid-loop → blips resume without a restart.
+    vi.mocked($hapticsMuted.get).mockReturnValue(false)
+    vi.advanceTimersByTime(3_000)
+    expect(started.length).toBeGreaterThan(0)
+  })
+
+  it('start is idempotent', () => {
+    startThinkingSound()
+    startThinkingSound()
+    vi.advanceTimersByTime(1_300)
+
+    // One loop: at most ~2 blips in 1.3s (first at 400ms, next ≥800ms later).
+    expect(started.length).toBeLessThanOrEqual(2)
+  })
+
+  it('never throws when WebAudio is unavailable', () => {
+    vi.stubGlobal('AudioContext', undefined)
+    expect(() => {
+      startThinkingSound()
+      vi.advanceTimersByTime(2_000)
+    }).not.toThrow()
+    stopThinkingSound()
+  })
+})

apps/desktop/src/lib/thinking-sound.ts

@@ -0,0 +1,108 @@
+// Ambient "thinking" sound for the desktop voice conversation. While the agent
+// works (status === 'thinking') no audio flows, which reads as "it died" during
+// long thinking/tool stretches. A calm, quiet, repeating pair of soft bubble
+// blips fills the gap — same WebAudio oscillator synthesis approach as
+// wake-sound.ts / completion-sound.ts (no asset to ship), mirroring the
+// backend's numpy-synthesized blips in tools/voice_mode.py so CLI and desktop
+// sound alike.
+//
+// Honours the shared sound-mute toggle ($hapticsMuted) and the
+// voice.thinking_sound config gate ($thinkingSoundEnabled). Stops instantly on
+// stopThinkingSound() — callers fire it the moment TTS starts, the mic re-arms,
+// or the conversation ends.
+
+import { $hapticsMuted } from '@/store/haptics'
+import { $thinkingSoundEnabled } from '@/store/voice-prefs'
+
+let ctx: AudioContext | null = null
+let timer: number | null = null
+let blipIndex = 0
+
+function getCtx(): AudioContext | null {
+  if (typeof window === 'undefined') {
+    return null
+  }
+
+  try {
+    if (!ctx) {
+      const Ctor =
+        window.AudioContext || (window as unknown as { webkitAudioContext?: typeof AudioContext }).webkitAudioContext
+
+      if (!Ctor) {
+        return null
+      }
+
+      ctx = new Ctor()
+    }
+
+    if (ctx.state === 'suspended') {
+      void ctx.resume().catch(() => undefined)
+    }
+
+    return ctx
+  } catch {
+    return null
+  }
+}
+
+// One soft "blub": short sine with a gentle downward pitch glide and a smooth
+// attack into an exponential decay — no clicks, deliberately quiet.
+function blub(ac: AudioContext, freq: number) {
+  const t0 = ac.currentTime + 0.01
+  const dur = 0.16
+  const osc = ac.createOscillator()
+  const env = ac.createGain()
+
+  osc.type = 'sine'
+  osc.frequency.setValueAtTime(freq, t0)
+  osc.frequency.exponentialRampToValueAtTime(freq * 0.72, t0 + dur)
+
+  env.gain.setValueAtTime(0.0001, t0)
+  env.gain.exponentialRampToValueAtTime(0.08, t0 + 0.02)
+  env.gain.exponentialRampToValueAtTime(0.0001, t0 + dur)
+
+  osc.connect(env)
+  env.connect(ac.destination)
+  osc.start(t0)
+  osc.stop(t0 + dur + 0.02)
+}
+
+export function isThinkingSoundActive(): boolean {
+  return timer !== null
+}
+
+/** Start the repeating thinking blips (idempotent). Best-effort, never throws. */
+export function startThinkingSound(): void {
+  if (timer !== null || !$thinkingSoundEnabled.get()) {
+    return
+  }
+
+  const tick = () => {
+    if ($hapticsMuted.get() === false) {
+      const ac = getCtx()
+
+      if (ac) {
+        try {
+          // Alternate two calm pitches (G4 / E4), matching the backend blips.
+          blub(ac, blipIndex % 2 === 0 ? 392 : 329.6)
+        } catch {
+          // Audio backend unavailable — stay silent, keep the loop harmless.
+        }
+      }
+    }
+
+    blipIndex += 1
+    // ~0.8-1.2s spacing with slight randomization so it reads organic.
+    timer = window.setTimeout(tick, 800 + Math.random() * 400)
+  }
+
+  timer = window.setTimeout(tick, 400)
+}
+
+/** Stop the thinking blips instantly (idempotent). */
+export function stopThinkingSound(): void {
+  if (timer !== null) {
+    window.clearTimeout(timer)
+    timer = null
+  }
+}