Computer-use agent inference harness. Screenshot → Qwen3-VL → parsed tool call → dispatched as OS input. macOS only (for now).

The artifact is an .app bundle (work.app) that wraps a single CLI binary. Modules inside (model, input, screen, qwen3vl, log, error) are internal — they back main but aren't an external library surface.

./scripts/bundle.sh
open "x-apple.systempreferences:com.apple.preference.security?Privacy_Accessibility"
# drag target/bundle/work.app into the list, toggle ON

CFBundleIdentifier = org.pdoom.work is stable across rebuilds, so the Accessibility grant persists.

MODEL_ENDPOINT=http://localhost:8000 \
MODEL_NAME="Qwen/Qwen3-VL-30B-A3B-Instruct" \
./target/bundle/work.app/Contents/MacOS/work "Click the bluetooth icon."

Env:

• MODEL_ENDPOINT required; no trailing slash
• MODEL_NAME required
• MODEL_API_KEY optional; sent as bearer if set
• WORK_SESSIONS optional; where to write session dirs (default: ./sessions)

Session events.jsonl + frames/ land under WORK_SESSIONS/<timestamp>/.

cargo test                          # offline unit tests
cargo test -- --include-ignored     # adds live-endpoint tests (MODEL_ENDPOINT + MODEL_NAME required)

Edit src/main.rs (or any module), rerun ./scripts/bundle.sh, rerun the binary. No re-granting.