earakely-scale's picture
REA computer-use environment, v0.1.0
913eb00 verified
|
Raw History Blame Contribute Delete
4.22 kB

Dia computer app

Exported by REA's export-desktop-application from a capture of the dia package (capture Evidence ev_ef628427ef0fd3bae36814ad04b2e6050cd5bcac7e2a355cdc51aeda7b0db60a). It was checked against agentenv-framework 0.9.1431.

  • envs/dia/ is a computer app on the ubuntu-24.04 base with the reference driver, for amd64 and arm64. It installs dia from the distribution archive, launches sh -c cd "$HOME" 2>/dev/null; exec 'dia' "$HOME"/'Documents/order-flow.dia' at boot and is ready when a window of class Dia appears.
  • At launch the app opens ~/Documents/order-flow.dia, the seed's one document.
  • Owned paths: ~/.cache/dconf, ~/Documents. data/get returns their files and a typed JSON view of each; data/reset restores the seed.
  • Seed: 1 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/dia/seed/ before publishing the bundle.

Task rename-stock-box

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

In my order flowchart, Documents/order-flow.dia, rename the "Check stock" box to "Verify inventory" and save it.

Checks, each of which must pass:

  • renamed: {"kind":"text_equals","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children/[name=dia:object][@id=O1]/children/[name=dia:attribute][@name=text]/children/[name=dia:composite]/children/[name=dia:attribute][@name=string]","value":"#Verify inventory#"}
  • one-box-named: {"kind":"member_count","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children","matching":"[name=dia:object][@type=Flowchart - Box][>dia:attribute=#Verify inventory#]","equals":1}
  • four-boxes: {"kind":"member_count","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children","matching":"[name=dia:object][@type=Flowchart - Box]","equals":4}
  • guard-receive: {"kind":"text_equals","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children/[name=dia:object][@id=O0]/children/[name=dia:attribute][@name=text]/children/[name=dia:composite]/children/[name=dia:attribute][@name=string]","value":"#Receive order#"}
  • guard-pack: {"kind":"text_equals","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children/[name=dia:object][@id=O2]/children/[name=dia:attribute][@name=text]/children/[name=dia:composite]/children/[name=dia:attribute][@name=string]","value":"#Pack items#"}
  • guard-ship: {"kind":"text_equals","file":"Documents/order-flow.dia","pointer":"/root/children/[name=dia:layer][>dia:object=#Receive order#]/children/[name=dia:object][@id=O3]/children/[name=dia:attribute][@name=text]/children/[name=dia:composite]/children/[name=dia:attribute][@name=string]","value":"#Ship order#"}

REA kept the task because its checks fail on the seed (renamed, one-box-named) 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 rename-stock-box --agent claude-code --model MODEL