earakely-scale's picture
REA computer-use environment, v0.1.0
db620a0 verified
|
Raw History Blame Contribute Delete
3.31 kB
# LyX computer app
Exported by REA's `export-desktop-application` from a capture of the `lyx` package
(capture Evidence `ev_a1721a9bb396409192540109192c45c9a2186d0ed4a2c1cb329636edf97249ce`). It was checked against agentenv-framework
0.9.1431.
- `envs/lyx/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
`amd64` and `arm64`. It installs `lyx` from the distribution archive, launches `sh -c cd "$HOME" 2>/dev/null; exec 'lyx' "$HOME"/'Documents/garden-report.lyx'` at boot and is ready
when a window of class `lyx` appears.
- At launch the app opens `~/Documents/garden-report.lyx`, the seed's one document.
- Owned paths: `~/.config/LyX`, `~/.lyx`, `~/Documents`.
`data/get` returns their files and a typed JSON view of each; `data/reset` restores the seed.
- Seed: 16 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
`~/.config/QtProject.conf`, `~/.config/enchant`.
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/lyx/seed/` before publishing the bundle.
## Task `fix-section-heading`
`tasks/fix-section-heading.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/fix-section-heading-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/lyx.toml` names it, so a run of the bundle runs it.
> In LyX, the heading of section 2 in Documents/garden-report.lyx says "Soil Preperation". Correct the misspelled word so the heading reads "Soil Preparation", and save the document.
Checks, each of which must pass:
- `heading-fixed`: `{"kind":"lines_present","file":"Documents/garden-report.lyx","lines":["Soil Preparation"]}`
- `typo-gone`: `{"kind":"lines_absent","file":"Documents/garden-report.lyx","lines":["Soil Preperation"]}`
- `three-sections`: `{"kind":"line_count","file":"Documents/garden-report.lyx","exact":"\\begin_layout Section","equals":3}`
- `other-headings-kept`: `{"kind":"lines_present","file":"Documents/garden-report.lyx","lines":["Introduction","Harvest Results"]}`
REA kept the task because its checks fail on the seed (`heading-fixed`, `typo-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 fix-section-heading --agent claude-code --model MODEL