earakely-scale's picture
REA computer-use environment, v0.1.0
9a66e8a verified
|
Raw History Blame Contribute Delete
3.6 kB

Pencil2D computer app

Exported by REA's export-desktop-application from a capture of the pencil2d package (capture Evidence ev_a0b3c0a8a390f407a387bf3f067af9329627bb107a646980d40e536016c6791d). It was checked against agentenv-framework 0.9.1431.

  • envs/pencil2d/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs pencil2d from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'pencil2d' "$HOME"/'Documents/bounce.pclx' at boot and is ready when a window of class Pencil2D appears.
  • At launch the app opens ~/Documents/bounce.pclx, the seed's one document.
  • Owned paths: ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
  • Seed: 1 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: these files or folders, which changed during the capture, ~/.config/Pencil, ~/.config/QtProject.conf. To own them, capture again naming them with state_paths on start_build_capture, or --state-path on capture-desktop-application.
  • 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/pencil2d/seed/ before publishing the bundle.

Task rename-ball-layer

tasks/rename-ball-layer.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/rename-ball-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/pencil2d.toml names it, so a run of the bundle runs it.

In Pencil2D, rename the "Bitmap Layer" layer of my bouncing ball animation in Documents/bounce.pclx to Ball and save it.

Checks, each of which must pass:

  • layer-renamed: {"kind":"value_equals","file":"Documents/bounce.pclx","pointer":"/entries/[name=main.xml]/view/root/children/[name=object]/children/[name=layer][@id=3]/attributes/name","value":"Ball"}
  • old-name-gone: {"kind":"value_absent","file":"Documents/bounce.pclx","pointer":"/entries/[name=main.xml]/view/root/children/[name=object]/children/[name=layer][@name=Bitmap Layer]"}
  • ground-kept: {"kind":"value_equals","file":"Documents/bounce.pclx","pointer":"/entries/[name=main.xml]/view/root/children/[name=object]/children/[name=layer][@id=4]/attributes/name","value":"Ground"}
  • bitmap-layer-count: {"kind":"member_count","file":"Documents/bounce.pclx","pointer":"/entries/[name=main.xml]/view/root/children/[name=object]/children","matching":"[name=layer][@type=1]","equals":2}

REA kept the task because its checks fail on the seed (layer-renamed, old-name-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-ball-layer --agent claude-code --model MODEL