earakely-scale's picture
REA computer-use environment, v0.1.0
35fa020 verified
|
Raw History Blame Contribute Delete
2.17 kB

Mousepad computer app

Exported by REA's export-desktop-application from a capture of the mousepad package (capture Evidence ev_22dd18d1ecd38dbb96e8ab160a67bdc00a95aa24a6af4da49f87db3f31bb3669). It was checked against agentenv-framework 0.9.1431.

  • envs/mousepad/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs mousepad from the distribution archive, launches sh -c exec 'mousepad' "$HOME"/'Documents/todo.txt' at boot and is ready when a window of class Mousepad appears.
  • Owned paths: ~/.config/Mousepad, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
  • Seed: 1 file(s) from the user's home before the capture, with the captured home path respelled in text files.
  • tasks/smoke.json deploys the app.

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

The seed is synthetic test data made for this showcase; it holds no one's real files.

Task add-milk

tasks/add-milk.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/add-milk-check/verify.py, which reads the app's data/get views and needs no model. evals/mousepad.toml names it, so a run of the bundle runs it.

Add buy milk as a new line at the end of the to-do list open in Mousepad, then save it.

Checks, each of which must pass:

  • milk: {"kind":"lines_present","file":"Documents/todo.txt","lines":["buy milk"]}
  • bread: {"kind":"lines_present","file":"Documents/todo.txt","lines":["buy bread"]}
  • one-milk: {"kind":"line_count","file":"Documents/todo.txt","exact":"buy milk","equals":1}

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