earakely-scale's picture
REA computer-use environment, v0.1.0
de1f371 verified
|
Raw History Blame Contribute Delete
3.7 kB
# Tiled computer app
Exported by REA's `export-desktop-application` from a capture of the `tiled` package
(capture Evidence `ev_c75137ec6f5efa14916e2f57f2d7c57e0919fd40276fcff15145cc1ea9a4d517`). It was checked against agentenv-framework
0.9.1431.
- `envs/tiled/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
`amd64` and `arm64`. It installs `tiled` from the distribution archive, launches `sh -c cd "$HOME" 2>/dev/null; exec 'tiled' "$HOME"/'Documents/harbor_town.tmx'` at boot and is ready
when a window of class `tiled` appears.
- At launch the app opens `~/Documents/harbor_town.tmx`, the seed's one document.
- Owned paths: `~/.local/share/tiled`, `~/Documents`.
`data/get` returns their files and a typed JSON view of each; `data/reset` restores the seed.
- Seed: 2 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/mapeditor.org`.
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/tiled/seed/` before publishing the bundle.
## Task `rename-old-well`
`tasks/rename-old-well.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/rename-old-well-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/tiled.toml` names it, so a run of the bundle runs it.
> In my Tiled map Documents/harbor_town.tmx, rename the object called Old Well to Wishing Well and save the map.
Checks, each of which must pass:
- `renamed`: `{"kind":"value_equals","file":"Documents/harbor_town.tmx","pointer":"/root/children/[name=objectgroup][@id=2]/children/[name=object][@id=1]/attributes/name","value":"Wishing Well"}`
- `old-name-gone`: `{"kind":"value_absent","file":"Documents/harbor_town.tmx","pointer":"/root/children/[name=objectgroup][@id=2]/children/[name=object][@name=Old Well]"}`
- `door-kept`: `{"kind":"value_equals","file":"Documents/harbor_town.tmx","pointer":"/root/children/[name=objectgroup][@id=2]/children/[name=object][@id=2]/attributes/name","value":"Blacksmith Door"}`
- `gate-kept`: `{"kind":"value_equals","file":"Documents/harbor_town.tmx","pointer":"/root/children/[name=objectgroup][@id=2]/children/[name=object][@id=3]/attributes/name","value":"Harbor Gate"}`
- `four-objects`: `{"kind":"member_count","file":"Documents/harbor_town.tmx","pointer":"/root/children/[name=objectgroup][@id=2]/children","matching":"[name=object]","equals":4}`
REA kept the task because its checks fail on the seed (`renamed`, `old-name-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 rename-old-well --agent claude-code --model MODEL