bundle/README.md

ghostwriter computer app

Exported by REA's export-desktop-application from a capture of the ghostwriter package (capture Evidence ev_64d3e46e61948715f1e70b202e79d75e2ee7000678bd0d53216ba596ead216c7). It was checked against agentenv-framework 0.9.1431.

• envs/ghostwriter/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs ghostwriter from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'ghostwriter' "$HOME"/'Documents/camping-trip.md' at boot and is ready when a window of class ghostwriter appears.
• At launch the app opens ~/Documents/camping-trip.md, the seed's one document.
• Owned paths: ~/.local/share/ghostwriter, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
• Seed: 18 file(s) made for the task, as saved at checkpoint seed. Text files and dconf keys spell the capture host's name as localhost, user@host as user@localhost and the home as /home/user, which the computer respells to its own at boot.
• Neither owned nor seeded, so the app starts without what it saved there: files changed during the capture in ~/.config/QtProject.conf. To own them, capture again naming them with state_paths on start_build_capture, or --state-path on capture-desktop-application.
• Seeded as captured, though the app keeps them for its own running, so a stale one can change how it starts: WebKit or Chromium storage ~/.local/share/ghostwriter/QtWebEngine/Default/Local Storage, ~/.local/share/ghostwriter/QtWebEngine/Default/Network Persistent State; caches ~/.local/share/ghostwriter/QtWebEngine/Default/GPUCache.
• Left out of the seed: lock files the app holds while it runs, since a copied one can make it find its profile or document in use and keep it from starting: ~/.local/share/ghostwriter/QtWebEngine/Default/Local Storage/leveldb/LOCK, ~/.local/share/ghostwriter/QtWebEngine/Default/Platform Notifications/LOCK.
• tasks/smoke.json deploys the app.

Run the task with agent-env run <this folder>, or only the deploy with --task smoke.

The seed was made for this task, not taken from a user's own data; read envs/ghostwriter/seed/ before publishing the bundle.

Task add-water-filter

tasks/add-water-filter.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/add-water-filter-check/verify.py, which reads the app's data/get views and needs no model. It grades once the app's files have held for 3 s, so a save the app makes on its own within that time of the agent stopping counts; each result's settled is false when they were still changing 15 s after the first read. evals/ghostwriter.toml names it, so a run of the bundle runs it.

In ghostwriter, add Water filter to the Gear checklist in Documents/camping-trip.md and save it.

Checks, each of which must pass:

• water-filter: {"kind":"lines_present","file":"Documents/camping-trip.md","lines":["- [ ] Water filter"]}
• added-once: {"kind":"line_count","file":"Documents/camping-trip.md","contains":"Water filter","equals":1}
• kept-items-near: {"kind":"lines_present","file":"Documents/camping-trip.md","lines":["- [ ] Camp stove","- [ ] Headlamps"]}
• kept-items-top: {"kind":"lines_present","file":"Documents/camping-trip.md","lines":["- [x] Tent","- [ ] Sleeping bags"]}

REA kept the task because its checks fail on the seed (water-filter, added-once) and all pass on the captured end state.

To measure a solve rate, run the task k times (3 by default) with the agent and model to measure, and count the runs whose summary line says passed; a run that says failed did not finish and is run again rather than counted:

agent-env run <this folder> --task add-water-filter --agent claude-code --model MODEL