earakely-scale's picture
REA computer-use environment, v0.1.0
fbbbd86 verified
|
Raw History Blame Contribute Delete
3.15 kB
# gLabels Label Designer 3 computer app
Exported by REA's `export-desktop-application` from a capture of the `glabels` package
(capture Evidence `ev_92a1ee88a892a24749a52e7954d21e09e8086fbf72c122cc198ac3d6f65be773`). It was checked against agentenv-framework
0.9.1431.
- `envs/glabels/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
`amd64` and `arm64`. It installs `glabels` from the distribution archive, launches `sh -c cd "$HOME" 2>/dev/null; exec 'glabels-3' "$HOME"/'Documents/address-label.glabels'` at boot and is ready
when a window of class `Glabels-3` 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/address-label.glabels`, the seed's one document.
- Owned paths: `~/Documents`.
`data/get` returns their files and a typed JSON view of each; `data/reset` restores the seed.
- Seed: 1 file(s) made for the task, as saved at checkpoint `seed`, and the application's dconf keys.
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/glabels/seed/Documents/address-label.glabels`.
- `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/glabels/seed/` before publishing the bundle.
dconf keys are compiled into the user's dconf database, with `dconf-cli`, once when the app first starts; `data/reset` does not restore them.
## Task `fix-label-name`
`tasks/fix-label-name.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/fix-label-name-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/glabels.toml` names it, so a run of the bundle runs it.
> The name on my address label in Documents/address-label.glabels is misspelled. It should read Marisol Penhallow. Fix the name in gLabels and save the label.
Checks, each of which must pass:
- `check-1`: `{"kind":"text_equals","file":"Documents/address-label.glabels","pointer":"/root/children/[name=Objects]/children/[name=Object-text]/children/[name=Span]","value":"Marisol Penhallow418 Juniper CourtBellmont, OR 97215"}`
REA kept the task because its checks fail on the seed (`check-1`) 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-label-name --agent claude-code --model MODEL