# NAS deployment — 2026-09-08

The `voice` container on `192.168.86.2` hosts LiveKit, the voice agent,
token service, music API and nginx. The app connects to
`https://192.168.86.2:9178`. Docker uses host networking and an
`unless-stopped` restart policy.

- Image: `hope-voice-api-voice:68f4c46-ios-audio`.
- LLM: **qwen2.5-vl-3b**, through OpenWebUI at `http://192.168.86.2:3005/api`.
- Speech: existing Azure configuration, unchanged.
- Persistent certs, skills and music: `/mnt/user/dev/hope-voice-api` mounts.
- Build source: `/mnt/user/dev/hope-voice-deploy-68f4c46`, a separate clean
  archive plus the iOS audio compatibility patch. The pre-existing dirty
  backend checkout was not overwritten.
- Local compatibility source: sibling `hope-voice-api/web/media_server.py`;
  regression test: `hope-voice-api/tests/test_ios_audio.py`.

The music service retains original WebM files. An explicit `?format=m4a`
request generates and caches AAC audio under `media/files/ios`, serving
byte ranges for iOS playback and seeking. Browser requests remain unchanged.

## Health and rollback

```sh
ssh root@192.168.86.2 'docker inspect --format "{{.State.Status}} {{.State.Health.Status}}" voice'
```

The stopped container `voice-before-ios-audio` retains the baseline deployment.
If rollback is needed, first inspect both containers, then stop the current
`voice`, rename it to a unique recovery name, rename `voice-before-ios-audio`
to `voice`, and start it. This restores the baseline without deleting images,
containers or the music library; it does not include iOS AAC conversion.

`scripts/deploy-backend.sh` records the launch configuration and refuses to
replace an existing `voice` container. Run it on the NAS only after building
the named image and explicitly handling the existing container. It reads
credentials on the NAS and never prints them. No source changes were pushed;
future CI deployments must include the compatibility patch to preserve it.