earakely-scale's picture
REA computer-use environment, v0.1.0
a033dfe verified
|
Raw History Blame Contribute Delete
3.5 kB
# LibreOffice Calc computer app
Exported by REA's `export-desktop-application` from a capture of the `libreoffice-calc` package
(capture Evidence `ev_8c5f728dff342fed12aa62c5dbeb9fe41179fe34d33284e313ce8a1bde75ce1c`). It was checked against agentenv-framework
0.9.1431.
- `envs/libreoffice-calc/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
`amd64` and `arm64`. It installs `libreoffice-calc` from the distribution archive, launches `sh -c cd "$HOME" 2>/dev/null; exec 'libreoffice' '--calc' "$HOME"/'Documents/q1-sales.ods'` at boot and is ready
when a window of class `libreoffice-calc` appears.
- `libreoffice-common`, which owns the program the desktop entry runs, is installed with it.
- At launch the app opens `~/Documents/q1-sales.ods`, the seed's one document.
- Owned paths: `~/.config/libreoffice`, `~/Documents`.
`data/get` returns their files and a typed JSON view of each; `data/reset` restores the seed.
- Seed: 584 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/Thunar`.
To own them, capture again naming them with `state_paths` on `start_build_capture`, or `--state-path` on `capture-desktop-application`.
- Seeded as captured, though the app keeps them for its own running, so a stale one can change how it starts:
caches `~/.config/libreoffice/4/cache`, `~/.config/libreoffice/4/user/uno_packages/cache`.
- `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/libreoffice-calc/seed/` before publishing the bundle.
## Task `fix-february-sales`
`tasks/fix-february-sales.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/fix-february-sales-check/verify.py`, which reads the app's `data/get` views and needs
no model. `evals/libreoffice-calc.toml` names it, so a run of the bundle runs it.
> February's sales figure in my quarterly sales spreadsheet, Documents/q1-sales.ods, is wrong: it should be 23,100. Fix it and save the spreadsheet.
Checks, each of which must pass:
- `february`: `{"kind":"text_contains","file":"Documents/q1-sales.ods","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:spreadsheet]/children/[name=table:table][@table:name=Q1 Sales]","value":"February23100March"}`
- `sheet`: `{"kind":"text_equals","file":"Documents/q1-sales.ods","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:spreadsheet]/children/[name=table:table][@table:name=Q1 Sales]","value":"MonthSales (USD)January18450February23100March19875Q1 Total61425"}`
REA kept the task because its checks fail on the seed (`february`, `sheet`) 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-february-sales --agent claude-code --model MODEL