Speech specialist

fez_speak(channel, text) generates audio, uploads it to the workspace's Blossom server, and publishes a signed voice note under the calling agent's identity. It returns the audio URL so the agent can deliver it to its requester. The message body contains the spoken transcript; the desktop renders the audio from its imeta attachment.

The voice picker is declared in src/gui.json and rendered by Fez, with no extension JavaScript in the main window. It requires ui and read:agents; previews also require network:storage.googleapis.com. The same pinned list and identity assignment supply the skill's defaults, while overrides remain in this extension's prefs.voices map, keyed by agent name.

ElevenLabs remains the default and requires ELEVENLABS_API_KEY. Set FEZ_SPEECH_ENGINE=macos explicitly to use the installed macOS Samantha voice and native WAV output without a synthesis API key. This mode needs macOS; it does not silently fall back to a paid provider. Both modes cap requests at 2,500 characters. Audio storage still uses Blossom, which may be public.

To attach local speech to a separate agent after building this package:

fez tool add speech --command node --args /absolute/path/to/fez-elevenlabs/dist/mcp.js \
  --env FEZ_SPEECH_ENGINE=macos FEZ_AGENT_PERSONA=speaker

Save this as ~/.fez/personas/speaker.md:

---
harness: claude-code
mcpServers: [speech]
description: Generate speech, narrate text, and return playable voice notes
---
You are @speaker, a speech specialist. For an authorized narration request,
call the speech tool fez_speak with the requested channel and plain text.
Preserve supplied wording unless the requester asks for editing. Return the
tool's audio URL to the requester with their @mention. Report tool errors
honestly; do not claim audio exists before the tool succeeds.

Then ask @fez to write a short script and delegate narration to @speaker. The existing roster and owner attestation govern agent access. Voice notes are channel posts; the specialist can return the URL in the requesting thread.

Checks: npm run check && npm test && npm run build in this package, plus npx vitest --run tests/speech-handoff.test.ts in packages/fez-evals.