bundle/README.md

Parley computer app

Exported by REA's export-desktop-application from a capture of the parley package (capture Evidence ev_2a09f0a0b4ca6e7585a0fa4ff5e1478608f94e9f58943da83c08c5a96c163d0e). It was checked against agentenv-framework 0.9.1431.

• envs/parley/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs parley from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'parley' "$HOME"/'Documents/spanish.kvtml' at boot and is ready when a window of class parley appears.
• At launch the app opens ~/Documents/spanish.kvtml, the seed's one document.
• Owned paths: ~/.config/parleyrc, ~/.local/share/stalefiles/parley, ~/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: these files or folders, which changed during the capture, ~/.config/QtProject.conf, ~/.config/emaildefaults. 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/parley/seed/ before publishing the bundle.

Task add-library-word-pair

tasks/add-library-word-pair.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/add-library-word-pair-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/parley.toml names it, so a run of the bundle runs it.

In Parley, add a new word pair to my Spanish Basics vocabulary collection in Documents/spanish.kvtml: the English word "library" with the Spanish "biblioteca". Then save the collection.

Checks, each of which must pass:

• pair-spanish: {"kind":"text_equals","file":"Documents/spanish.kvtml","pointer":"/root/children/[name=entries]/children/[name=entry][>translation=library]/children/[name=translation][@id=1]","value":"biblioteca","normalize_whitespace":true}
• pair-english: {"kind":"text_equals","file":"Documents/spanish.kvtml","pointer":"/root/children/[name=entries]/children/[name=entry][>translation=biblioteca]/children/[name=translation][@id=0]","value":"library","normalize_whitespace":true}
• one-more-entry: {"kind":"member_count","file":"Documents/spanish.kvtml","pointer":"/root/children/[name=entries]/children","matching":"[name=entry]","equals":7}
• apple-kept: {"kind":"text_equals","file":"Documents/spanish.kvtml","pointer":"/root/children/[name=entries]/children/[name=entry][>translation=apple]/children/[name=translation][@id=1]","value":"manzana"}
• book-kept: {"kind":"text_equals","file":"Documents/spanish.kvtml","pointer":"/root/children/[name=entries]/children/[name=entry][>translation=book]/children/[name=translation][@id=1]","value":"libro"}

REA kept the task because its checks fail on the seed (pair-spanish, pair-english, one-more-entry) 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 add-library-word-pair --agent claude-code --model MODEL