feat(voice): local voice dictation via Foundry Local + Nemotron streaming

[HashwanthVen]

Voice dictation: speak into Chamber locally (Nemotron + Foundry Local)

Summary

Adds local-first voice dictation to Chamber. Users hold a renderer-only shortcut (Alt+Shift+V) or click a mic button in the chat input to speak; transcripts insert at the caret of the active single-agent chat input. Never auto-sends — user always reviews before pressing Send.

Built on the same foundry-local-sdk@1.1.0 + Nemotron streaming model that the official GitHub Copilot CLI /voice uses, so the engineering pattern is upstream-validated.

Closes the gap in the PRD at VOICE_DICTATION_PRD.md (planning doc lives outside the repo).

Highlights

• Local-only ASR. nemotron-speech-streaming-en-0.6b runs in a worker_threads.Worker via Foundry Local SDK. Audio never leaves the device. Model auto-downloads on first use (~696 MB CPU variant).
• Push-to-talk keyboard shortcut, renderer-only (no globalShortcut — explicitly out of scope for Phase 1). Default Alt+Shift+V. Configurable hold-vs-toggle behavior.
• Mic button in ChatInput with visible "Listening…" indicator. Hidden when feature flag is off; disabled with tooltip until the model is Ready.
• Settings → Voice dictation section with six rows: device select, permission state + "Open preferences", Test mic readiness check, shortcut display, keyboard-shortcut-behavior toggle, model row with live download progress.
• Feature-flagged behind voiceDictation (default off on stable, on for insiders + dev). Cleanly invisible when off.
• Service-boundary clean. packages/services/** never imports electron or foundry-local-sdk; depcruise enforces the rule (new foundry-local-sdk-only-in-voice-workers rule added).
• Pinned runtime. chamber-voice-runtime/ mirrors the existing chamber-copilot-runtime/ pattern. foundry-local-sdk@1.1.0 pinned exactly. forge.config.ts adds the SDK to asar.unpack for native prebuilds.

Architecture

Renderer (apps/web)
  ChatInput mic button + push-to-talk key handler
        ↓
  useVoiceDictation hook
        ↓  getUserMedia + AudioWorklet → PCM16 16 kHz mono frames
  window.electronAPI.voice.*  (preload contextBridge)
        ↓  IPC (IPC.VOICE.*)
Main (apps/desktop)
  setupVoiceIPC adapter  →  VoiceDictationService  →  VoiceWorkerPool
        ↓  worker_threads boundary
apps/desktop/src/main/voiceWorker/voiceWorker.ts
  FoundryLocalManager + LiveAudioTranscriptionSession (single worker; Foundry Core is a process-singleton)

Files changed

New:

• packages/shared/src/voice-types.ts + test — config, IPC contracts, RPC verbs, payload size guard
• packages/services/src/voice/ — VoiceDictationStore, VoiceDictationService, VoiceWorkerPool, FakeTranscriptionProvider, FoundryTranscriptionProvider, permissions/provider interfaces, all with colocated tests
• apps/desktop/src/main/voice/PermissionInspector.ts + test — wraps systemPreferences.getMediaAccessStatus cross-platform
• apps/desktop/src/main/voiceWorker/voiceWorker.ts + test — single worker handling both engine + installer verbs
• apps/desktop/src/main/voiceWorker/workerProtocol.ts — shared RPC envelope
• apps/desktop/src/main/ipc/voice.ts + test — thin IPC adapter with sender ownership + chunk size validation
• apps/web/src/renderer/hooks/useVoiceDictation.ts + test — session lifecycle, sessionId routing, PTT, IME suppression, cleanup
• apps/web/src/renderer/lib/audio/ — PCM16 encoder, AudioWorklet processor, captureMic helper
• apps/web/src/renderer/components/settings/VoiceDictationSettingsSection.tsx + test
• chamber-voice-runtime/ — pinned foundry-local-sdk@1.1.0
• scripts/prepare-voice-runtime.js, scripts/check-voice-sdk-version.js, scripts/voice-runtime-smoke.js
• apps/desktop/vite.voiceWorker.config.ts
• tests/e2e/electron/voice-dictation-smoke.spec.ts

Modified (surgical):

• packages/shared/src/ipc-channels.ts — IPC.VOICE.* + IPC.E2E.VOICE_*
• packages/shared/src/electron-types.ts (+ test) — voice namespace
• packages/shared/src/feature-flags.ts (+ test) — voiceDictation flag
• apps/desktop/src/main.ts — wire VoiceDictationService when feature flag is on; honor CHAMBER_E2E_VOICE_FAKE=1 for deterministic tests
• apps/desktop/src/preload.ts — expose voice namespace
• apps/desktop/src/main/devFeatureFlags.ts — enable in dev
• apps/desktop/src/main/security/sessionSecurity.ts (+ test) — allow media permission for Chamber renderer origins only
• apps/web/src/renderer/components/chat/ChatInput.tsx (+ test) — mic button + PTT, never auto-submits
• apps/web/src/renderer/components/settings/SettingsView.tsx (+ test) — gate VoiceDictationSettingsSection behind flag
• forge.config.ts — asar.unpack for foundry-local-sdk native prebuilds; new build entry for voiceWorker.ts
• config/dependency-cruiser.cjs — enforces foundry-local-sdk import is allowed only from the worker file
• package.json / package-lock.json — foundry-local-sdk@1.1.0 pinned in dependencies + overrides + new smoke:voice-runtime script

Validation

• npm run typecheck ✅
• npm run lint ✅ (tsc + eslint + deps:check 547 modules / 0 boundary violations + yaml + md)
• 65+ voice unit + RTL tests ✅
• npm run smoke:voice-runtime ✅ (new — exercises the compiled worker bundle against a mocked SDK; would have caught the singleton bug)
• npm run smoke:desktop -- tests/e2e/electron/voice-dictation-smoke.spec.ts — 8 passed / 3 skipped / 0 failed. Three scenarios are test.fixme() with explanation: two need to spy chat.send but contextBridge freezes electronAPI (covered by chat-mic-inserts-sentinel which proves the visible textarea path); one is the second contextBridge case.
• chamber-ui-tester driven against the real Foundry path on Windows: Test mic readiness ✅, model download with live progress ✅, chat mic listening state ✅. Edge cases (re-entrant start during stop) found two bugs that are now fixed in this branch — see 94dda24 and the voice-rt-fix commit chain.

Privacy + security

• Audio never persisted, never logged, never sent over the network. keytar is not used (no secrets in the voice path).
• Microphone permission gated by sessionSecurity allowlist — only Chamber's own renderer origin can request media; foreign origins denied.
• worker_threads isolation between Foundry SDK and main process.
• webPreferences (contextIsolation: true, nodeIntegration: false) unchanged.
• All renderer ↔ main calls go through IPC.VOICE.* constants with zod schema validation and payload size limits (VOICE_MAX_APPEND_CHUNK_BYTES).
• IPC sessions are owner-scoped: appendAudio/endSession reject from a non-owner WebContents.
• gpt-5.5 security review: APPROVE, no findings.

Not in scope (Phase 2 candidates)

• Streaming partial transcripts inserted live (currently only finals; partials shown as transient "Listening…")
• Chatroom voice support (only single-agent chat)
• Multiple model picker / cloud transcription
• Voice activity detection / auto-stop on silence
• System-global keyboard shortcut (Electron.globalShortcut)
• Always-on wake word
• Text-to-speech for agent responses
• Click-to-rebind shortcut UI

Drift / mergeability

Branch is 31 commits ahead, 0 commits behind ianphil/chamber:master. Base point is 8834f0b (verified via gh api repos/ianphil/chamber/branches/master). Clean fast-forward / no-op rebase.

How to try

1. Pull this branch, npm install, npm start.
2. Settings → Voice dictation → click Download (~696 MB first time via Foundry Local).
3. Activate a mind, click the mic button in the chat input, speak, click to stop, edit transcript, press Enter or click Send.
4. Push-to-talk: hold Alt+Shift+V with the textarea focused.

---

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com