Voice

Browser speech-to-text and read-aloud, added 2026-10-09 (voice-providers.ts + routes in server.ts). The desktop agent never sees audio — transcripts arrive as plain text in the prompt.

Configuration

Per-user config lives on the desktops row (migration 0002): stt_/tts_ provider, account, key, model, tts_voice, and voice_language. Managed via:

GET  /api/voice      # keys are never returned to the browser
PUT  /api/voice
DELETE /api/voice

The Config page's Voice tab (VoiceProvider.tsx) renders the form. Providers can fall back to a configured vault entry: Workers AI reuses the cloudflare slug, xAI → xai, OpenAI → openai.

Provider	STT	TTS
xAI	`POST /v1/tts` for speech; Groq-compatible whisper for STT	✓
OpenAI	`/audio/speech`	✓
Cloudflare Workers AI	`@cf/openai/whisper`	`@cf/myshell-ai/melotts`

Endpoints

POST /api/stt — multipart field audio, hard cap 25 MB, returns {text}.
POST /api/tts — JSON {text}, hard cap 4000 chars, returns audio bytes.

UI

Hold-to-speak mic in the composer — hold, talk, release to send the transcript as the prompt.
Read-aloud toggle — synthesizes assistant replies. speakableText in chat-view.ts strips markdown and plugin cards before synthesis.

Voice is core (not dev-gated). Existing desktops need bun run images && bun run update after pulling the feature.