earakely-scale's picture
REA computer-use environment, v0.1.0
4a88c89 verified
|
Raw History Blame Contribute Delete
3.44 kB

Freeplane computer app

Exported by REA's export-desktop-application from a capture of the freeplane package (capture Evidence ev_ed047dd05a74c681a3fe606f991b8081a4a994d756fda748a31d69b1c7d93176). It was checked against agentenv-framework 0.9.1431.

  • envs/freeplane/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs freeplane from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec '/usr/bin/freeplane' "$HOME"/'Documents/weekend-trip.mm' at boot and is ready when a window of class org-knopflerfish-framework-BundleThread 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: openjdk-17-jre.
  • At launch the app opens ~/Documents/weekend-trip.mm, the seed's one document.
  • Owned paths: ~/.config/freeplane, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
  • Seed: 10 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 ~/.java. 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/freeplane/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/freeplane.toml names it, so a run of the bundle runs it.

In my Weekend Trip mind map (Documents/weekend-trip.mm) in Freeplane, add a child topic called Water filter under Packing, and save the map.

Checks, each of which must pass:

  • water-filter-topic: {"kind":"line_count","file":"Documents/weekend-trip.mm","contains":"<node TEXT=\"Water filter\" ID=\"","equals":1}
  • no-leaf-became-parent: {"kind":"line_count","file":"Documents/weekend-trip.mm","exact":"</node>","equals":4}
  • sleeping-bags-kept: {"kind":"lines_present","file":"Documents/weekend-trip.mm","lines":["<node TEXT=\"Sleeping bags\" ID=\"ID_1227863932\" CREATED=\"1793955122470\" MODIFIED=\"1793955125127\"/>"]}

REA kept the task because its checks fail on the seed (water-filter-topic) 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