bundle/README.md

JabRef computer app

Exported by REA's export-desktop-application from a capture of the jabref package (capture Evidence ev_d57ff8949f0cf1fc8f3d36dba0ab6d7f124300c71c2ed47c91e43b63edb87bdc). It was checked against agentenv-framework 0.9.1431.

• envs/jabref/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs jabref from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'jabref' "$HOME"/'Documents/references.bib' at boot and is ready when a window of class net-sf-jabref-JabRefMain appears.
• At launch the app opens ~/Documents/references.bib, 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: 3 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: these files or folders, which changed during the capture, ~/.java. 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/jabref/seed/ before publishing the bundle.

Task correct-sediment-year

tasks/correct-sediment-year.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/correct-sediment-year-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/jabref.toml names it, so a run of the bundle runs it.

In JabRef, correct the year of the Okafor and Lindqvist article "Seasonal Sediment Transport in Glacial Rivers" in Documents/references.bib to 2019 and save it.

Checks, each of which must pass:

• check-1: {"kind":"lines_present","file":"Documents/references.bib","lines":[" year = {2019},"]}
• check-2: {"kind":"lines_absent","file":"Documents/references.bib","lines":[" year = {2017},"]}
• guard-1: {"kind":"lines_present","file":"Documents/references.bib","lines":[" journal = {Journal of Fluvial Studies},"," volume = {12},"]}

REA kept the task because its checks fail on the seed (check-1, check-2) 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 correct-sediment-year --agent claude-code --model MODEL