# LiveKit Voice Assistant

This worker runs the Rocky Mountain Vending site voice assistant with LiveKit Agents and xAI realtime voice.

## Local development

1. Start the Next.js app:
   `pnpm dev`
2. Start the LiveKit agent worker in a second terminal:
   `pnpm voice-agent:dev`
3. Open the site and launch the voice bubble.
4. Run the automated regressions:
   `pnpm voice-assistant:test:local`
5. Run the fake-microphone voice regression:
   `pnpm voice-assistant:test:local:voice`

## Environment

The worker loads environment variables from:
- `.env.local`
- `.env.voice-agent.local` (optional override file)

Required variables:
- `LIVEKIT_URL`
- `LIVEKIT_API_KEY`
- `LIVEKIT_API_SECRET`
- `XAI_API_KEY`
- `XAI_REALTIME_MODEL=grok-4-1-fast-non-reasoning`

Recommended variable:
- `VOICE_ASSISTANT_SITE_URL`

## Production

Run the worker as a separate long-lived Node process:

`pnpm voice-agent:start`

Point `VOICE_ASSISTANT_SITE_URL` at the deployed Next.js app so the worker can reuse `/api/contact` and `/api/request-machine`.

## Verification

Automated checks:
- `pnpm voice-assistant:test:local`
- `pnpm voice-assistant:test:local:voice`
- `pnpm voice-assistant:test:live`

Manual spoken smoke checklist:
1. Open the homepage and launch the assistant bubble.
2. Press `Start assistant` and confirm the panel reaches `Text Ready` or `Listening`.
3. Confirm the text box is enabled before microphone access is granted.
4. Press `Enable microphone` and verify the browser permission prompt appears.
5. After granting permission, say a short question such as "What cities in Utah do you serve?"
6. Confirm you see a user transcript, then an assistant reply transcript.
7. Confirm spoken audio playback is audible after `Enable spoken replies` when the browser requires it.
8. Refresh the page and verify reconnect still works.
9. Open `/sign-in` and confirm the assistant stays hidden there.