bundle/README.md

MyPaint computer app

Exported by REA's export-desktop-application from a capture of the mypaint package (capture Evidence ev_d4e952048ab869003600d34e5b053720e38a2c1a37c3c312735c1efed2dd0bfd). It was checked against agentenv-framework 0.9.1431.

• envs/mypaint/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs mypaint from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'mypaint' "$HOME"/'Documents/sunny-day.ora' at boot and is ready when a window of class MyPaint appears.
• Installed beside the app because the export was asked for them with --extra-package, as a build computer asks for the packages its request named: librsvg2-common.
• At launch the app opens ~/Documents/sunny-day.ora, the seed's one document.
• Owned paths: ~/.local/share/mypaint, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
• Seed: 3 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.
• 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/mypaint/seed/ before publishing the bundle.

Task rename-sketch-layer

tasks/rename-sketch-layer.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/rename-sketch-layer-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/mypaint.toml names it, so a run of the bundle runs it.

In my drawing Documents/sunny-day.ora, rename the layer called Sketch to Crescent Moon and save it.

Checks, each of which must pass:

• renamed: {"kind":"value_equals","file":"Documents/sunny-day.ora","pointer":"/entries/[name=stack.xml]/view/root/children/[name=stack]/children/[name=layer][@name=Crescent Moon]/attributes/name","value":"Crescent Moon"}
• sketch-gone: {"kind":"value_absent","file":"Documents/sunny-day.ora","pointer":"/entries/[name=stack.xml]/view/root/children/[name=stack]/children/[name=layer][@name=Sketch]"}
• ground-kept: {"kind":"value_equals","file":"Documents/sunny-day.ora","pointer":"/entries/[name=stack.xml]/view/root/children/[name=stack]/children/[name=layer][@name=Ground]/attributes/name","value":"Ground"}
• night-sky-kept: {"kind":"value_equals","file":"Documents/sunny-day.ora","pointer":"/entries/[name=stack.xml]/view/root/children/[name=stack]/children/[name=layer][@name=Night Sky]/attributes/name","value":"Night Sky"}
• layer-count: {"kind":"member_count","file":"Documents/sunny-day.ora","pointer":"/entries/[name=stack.xml]/view/root/children/[name=stack]/children","matching":"[name=layer]","equals":4}

REA kept the task because its checks fail on the seed (renamed, sketch-gone) 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 rename-sketch-layer --agent claude-code --model MODEL