bundle/README.md

LibreOffice Draw computer app

Exported by REA's export-desktop-application from a capture of the libreoffice-draw package (capture Evidence ev_b4e9242c03370c32eebc70b7b7b598c14903544a9fea8e31415b6f59dd7b5532). It was checked against agentenv-framework 0.9.1431.

• envs/libreoffice-draw/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs libreoffice-draw from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'libreoffice' '--draw' "$HOME"/'Documents/order-flow.odg' at boot and is ready when a window of class libreoffice-draw appears.
• Installed beside the app because the export was asked for them with --extra-package, as a build computer asks for the packages its request named: fonts-liberation.
• libreoffice-common, which owns the program the desktop entry runs, is installed with it.
• At launch the app opens ~/Documents/order-flow.odg, the seed's one document.
• Owned paths: ~/.config/libreoffice, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
• Seed: 625 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.
• LibreOffice's list of the documents it had open, which would start it into document recovery instead of its window, is left out of ~/.config/libreoffice/4/user/registrymodifications.xcu.
• Seeded as captured, though the app keeps them for its own running, so a stale one can change how it starts: caches ~/.config/libreoffice/4/cache, ~/.config/libreoffice/4/user/uno_packages/cache.
• Left out of the seed: lock files the app holds while it runs, since a copied one can make it find its profile or document in use and keep it from starting: ~/.config/libreoffice/4/.lock.
• 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/libreoffice-draw/seed/ before publishing the bundle.

Task relabel-pick-box

tasks/relabel-pick-box.json deploys the app, has the agent work through the computer from the prompt below, and grades the outcome with artifacts/relabel-pick-box-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/libreoffice-draw.toml names it, so a run of the bundle runs it.

In my order fulfillment diagram, Documents/order-flow.odg, relabel the "Pick Items" box to "Pick and Scan" and save it.

Checks, each of which must pass:

• relabeled: {"kind":"text_equals","file":"Documents/order-flow.odg","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:drawing]/children/[name=draw:page]/children/[name=draw:custom-shape][@draw:name=Step 2]","value":"Pick and Scan"}
• box-above-kept: {"kind":"text_equals","file":"Documents/order-flow.odg","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:drawing]/children/[name=draw:page]/children/[name=draw:custom-shape][@draw:name=Step 1]","value":"Order Received"}
• box-below-kept: {"kind":"text_equals","file":"Documents/order-flow.odg","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:drawing]/children/[name=draw:page]/children/[name=draw:custom-shape][@draw:name=Step 3]","value":"Pack Items"}
• four-boxes: {"kind":"member_count","file":"Documents/order-flow.odg","pointer":"/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:drawing]/children/[name=draw:page]/children","matching":"[name=draw:custom-shape]","equals":4}

REA kept the task because its checks fail on the seed (relabeled) 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 relabel-pick-box --agent claude-code --model MODEL