earakely-scale's picture
REA computer-use environment, v0.1.0
c2f89ec verified
|
Raw History Blame Contribute Delete
3.38 kB

Xournal++ computer app

Exported by REA's export-desktop-application from a capture of the xournalpp package (capture Evidence ev_ae6b6473d4c3536078503ae95cc6b6ff3e22d29881d18f518b21b338a9e6142b). It was checked against agentenv-framework 0.9.1431.

  • envs/xournalpp/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs xournalpp from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'xournalpp' "$HOME"/'Documents/biology-notes.xopp' at boot and is ready when a window of class Com.github.xournalpp.xournalpp 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/biology-notes.xopp, the seed's one document.
  • Owned paths: ~/.config/xournalpp, ~/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. Rewritten: envs/xournalpp/seed/Documents/biology-notes.xopp.
  • 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/xournalpp/seed/ before publishing the bundle.

Task add-quiz-note

tasks/add-quiz-note.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/add-quiz-note-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/xournalpp.toml names it, so a run of the bundle runs it.

In Xournal++, add a text note that says "Quiz on Friday: bring a calculator" to the Cell Biology notebook page in Documents/biology-notes.xopp, and save it.

Checks, each of which must pass:

  • note-text: {"kind":"text_contains","file":"Documents/biology-notes.xopp","pointer":"/root/children/[name=page]/children/[name=layer]","value":"Quiz on Friday: bring a calculator"}
  • one-new-note: {"kind":"member_count","file":"Documents/biology-notes.xopp","pointer":"/root/children/[name=page]/children/[name=layer]/children","matching":"[name=text]","equals":4}
  • notes-kept: {"kind":"text_contains","file":"Documents/biology-notes.xopp","pointer":"/root/children/[name=page]/children/[name=layer]","value":"Cell Biology - Week 3Mitochondria make ATPRibosomes build proteins"}

REA kept the task because its checks fail on the seed (note-text, one-new-note) 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-quiz-note --agent claude-code --model MODEL