# JARVIS-QVAC

Local-first Ubuntu GNOME voice assistant. The cognitive core is the
already-extracted `/home/raven/dev/agent-harness`, copied into
`vendor/agent-harness`; Jarvis does not copy or depend on `/home/raven/dev/dlinux`.

## Bootstrap

```bash
npm install
npm test
npm run cu-doctor
npm run gpu-doctor
npx qvac doctor
npm run qvac:serve:diagnostic # optional; never run beside jarvisd
```

The daemon bridge loads the harness in-process with `Agent.engine.load()` and
creates sessions with `Agent.create()`. The optional QVAC HTTP sibling is
configured for `127.0.0.1:11434`. The GNOME extension is an intentionally thin
ESM panel shell; inference, audio, portals, and computer-use actuation belong
in user services.

GPU inference is a hard requirement. `gpu-doctor` must report a QVAC-visible
GPU and `jarvisd` refuses to continue if the loaded model reports `device: cpu`.
This prevents a misleading CPU mode and competing QVAC workers.

The repository contains the daemon, single GPU-only QVAC master, copied agent
harness, voice loop, GNOME ARC extension, GTK4/libadwaita control center,
computer-use observation and safety boundaries, and the full QVAC capability
skill registry. See [docs/ROADMAP.md](docs/ROADMAP.md),
[docs/agent-harness-map.md](docs/agent-harness-map.md), and
[docs/cu-acceptance.md](docs/cu-acceptance.md).

Runtime inference refuses to start until the QVAC SDK reports a usable GPU.
The RemoteDesktop/libei input bridge also remains unavailable until the local
GNOME portal session and a real EIS injector are present; Jarvis never claims
that an action succeeded when the backend is missing.