Voice mode + web UI for @open-gitagent/gitagent.

This package was split out of @open-gitagent/gitagent in core v2.0.0 so the CLI/SDK tarball stays slim and supply-chain scanners stop blocking it. The voice runtime (OpenAI Realtime + Gemini Live adapters, the file browser, the Composio toolkit bridge, the scheduler UI, the 200 KB single-file web UI) all live here.

# Both packages — voice mode works out of the box
npm install -g @open-gitagent/gitagent @open-gitagent/voice
# Then launch with voice:
gitagent --voice -d ~/assistant

The gitagent CLI ships in @open-gitagent/gitagent. --voice dynamically loads this package; without it installed, the CLI prints an install hint and exits.

• startVoiceServer(opts) — HTTP + WebSocket server (default :3333) with the web UI, file viewer, settings tab, skills/integrations/scheduler tabs.
• OpenAI Realtime adapter — GA endpoint, GA session.update shape, GA event names.
• Gemini Live adapter — Google's Live API.
• Composio bridge — toolkit browser and action execution from the UI.

import { startVoiceServer } from "@open-gitagent/voice";
const stop = await startVoiceServer({
 adapter: "openai-realtime",
 adapterConfig: { apiKey: process.env.OPENAI_API_KEY! },
 agentDir: "/path/to/agent",
});
// ...later
await stop();

startVoiceServer uses @open-gitagent/gitagent (peer dep) for the agent loader, SDK query(), the message-type protocol, and the standalone scheduler. Versions are decoupled — voice declares ^2.0.0 on core and follows core's semver.

The published @open-gitagent/gitagent@1.5.x tarball was being blocked with HTTP 403 by supply-chain scanners. The trigger was the bundled dist/voice/ui.html (~3,860 LOC of inline HTML/JS/CSS) plus the unused baileys dependency. Splitting voice out:

• core tarball: 179.5 kB → 85.7 kB packed (−52%), 310 → 211 deps
• voice tarball: ~150 kB packed (separate install, you opt in)

MIT.