bundle/README.md

GnuCash computer app

Exported by REA's export-desktop-application from a capture of the gnucash package (capture Evidence ev_5c8a4727e3df209dad648c39ea8d5ddbea0a6ec79cdc9614725db38d676fef15). It was checked against agentenv-framework 0.9.1431.

• envs/gnucash/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs gnucash from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'gnucash' "$HOME"/'Documents/household.gnucash' at boot and is ready when a window of class Gnucash appears.
• At launch the app opens ~/Documents/household.gnucash, the seed's one document.
• Owned paths: ~/.local/share/gnucash, ~/Documents, ~/translog.20261106084748.log. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
• Seed: 5 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.
• 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/gnucash/seed/ before publishing the bundle.

Task record-grocery-purchase

tasks/record-grocery-purchase.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/record-grocery-purchase-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/gnucash.toml names it, so a run of the bundle runs it.

In my household book (Documents/household.gnucash), record a grocery purchase: on November 3, 2026 I paid $63.48 from the Checking Account at Greenleaf Market. Enter it with the description Greenleaf Market and file it under Expenses:Groceries, then save the book.

Checks, each of which must pass:

• greenleaf-recorded-once: {"kind":"member_count","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children","matching":"[name=gnc:transaction][>trn:description=Greenleaf Market]","equals":1}
• groceries-debited: {"kind":"text_equals","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children/[name=gnc:transaction][>trn:description=Greenleaf Market]/children/[name=trn:splits]/children/[name=trn:split][>split:account=98f0ee4c216a43e290133637df088843]/children/[name=split:value]","value":"6348/100"}
• checking-credited: {"kind":"text_equals","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children/[name=gnc:transaction][>trn:description=Greenleaf Market]/children/[name=trn:splits]/children/[name=trn:split][>split:account=63fd8736eb444c57803945f330e68a28]/children/[name=split:value]","value":"-6348/100"}
• posted-nov-3: {"kind":"text_contains","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children/[name=gnc:transaction][>trn:description=Greenleaf Market]/children/[name=trn:date-posted]","value":"2026-11-03"}
• one-transaction-added: {"kind":"member_count","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children","matching":"[name=gnc:transaction]","equals":5}
• corner-grocer-kept: {"kind":"text_equals","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children/[name=gnc:transaction][>trn:description=Corner Grocer]/children/[name=trn:splits]/children/[name=trn:split][>split:account=98f0ee4c216a43e290133637df088843]/children/[name=split:value]","value":"5275/100"}
• guard-1: {"kind":"text_equals","file":"Documents/household.gnucash","pointer":"/root/children/[name=gnc:book]/children/[name=gnc:count-data][@cd:type=account]","value":"64"}

REA kept the task because its checks fail on the seed (greenleaf-recorded-once, groceries-debited, checking-credited, posted-nov-3, one-transaction-added) 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 record-grocery-purchase --agent claude-code --model MODEL