earakely-scale's picture
REA computer-use environment, v0.1.0
c8604cb verified
|
Raw History Blame Contribute Delete
3.47 kB
# DB Browser for SQLite computer app
Exported by REA's `export-desktop-application` from a capture of the `sqlitebrowser` package
(capture Evidence `ev_c49ceae8acb63a545d68e4b66aedf2c0709a58a9574e43423cca0c8735b7d288`). It was checked against agentenv-framework
0.9.1431.
- `envs/sqlitebrowser/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
`amd64` and `arm64`. It installs `sqlitebrowser` from the distribution archive, launches `sh -c cd "$HOME" 2>/dev/null; exec 'sqlitebrowser' "$HOME"/'Documents/contacts.db'` at boot and is ready
when a window of class `DB Browser for SQLite` appears.
- At launch the app opens `~/Documents/contacts.db`, the seed's one document.
- Owned paths: `~/.config/sqlitebrowser`, `~/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`.
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/sqlitebrowser/seed/` before publishing the bundle.
## Task `fix-customer-email`
`tasks/fix-customer-email.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/fix-customer-email-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/sqlitebrowser.toml` names it, so a run of the bundle runs it.
> In DB Browser for SQLite, the customer contacts database Documents/contacts.db has the wrong email address for Priya Raman. Correct her email address to priya.raman@example.com and save the change.
Checks, each of which must pass:
- `check-1`: `{"kind":"row_present","file":"Documents/contacts.db","table":"customers","where":{"id":3,"email":"priya.raman@example.com"}}`
- `rest-of-row`: `{"kind":"row_present","file":"Documents/contacts.db","table":"customers","where":{"id":3,"name":"Priya Raman","phone":"555-0103","city":"Denver"}}`
- `other-customer`: `{"kind":"row_present","file":"Documents/contacts.db","table":"customers","where":{"id":5,"name":"Grace Whitfield","email":"grace.whitfield@example.com"}}`
- `no-rows-added`: `{"kind":"row_count","file":"Documents/contacts.db","table":"customers","equals":6}`
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-customer-email --agent claude-code --model MODEL