File size: 2,016 Bytes
4e8f4e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
# LibreCAD computer app

Exported by REA's `export-desktop-application` from a capture of the `librecad` package
(capture Evidence `ev_6da77e96c478fb46bed2c9160b524d44d89f7de57d2ac758feed45feb5482405`). It was checked against agentenv-framework
0.9.1431.

- `envs/librecad/` is a computer app on the `ubuntu-24.04` base with the `reference` driver, for
  `amd64` and `arm64`. It installs `librecad` from the distribution archive, launches `sh -c exec 'librecad' "$HOME"/'Documents/plate.dxf'` at boot and is ready
  when a window of class `LibreCAD` appears.
- Owned paths: `~/.config/LibreCAD`, `~/.local/share/LibreCAD`, `~/Documents`.
  `data/get` returns their files and a typed JSON view of each; `data/reset` restores the seed.
- Seed: 3 file(s) from the user's home before the capture, with the captured home path
  respelled in text files.
- `tasks/smoke.json` deploys the app.

Run the task with `agent-env run <this folder>`, or only the deploy with `--task smoke`.

The seed is synthetic test data made for this showcase; it holds no one's real files.

## Task `add-holes`

`tasks/add-holes.json` deploys the app, has the agent work through the computer from the prompt below, and grades
the outcome with `artifacts/add-holes-check/verify.py`, which reads the app's `data/get` views and needs
no model. `evals/librecad.toml` names it, so a run of the bundle runs it.

> Add four more mounting holes to the plate drawing in LibreCAD and save it.

Checks, each of which must pass:

- `nine-circles`: `{"kind":"line_count","file":"Documents/plate.dxf","exact":"CIRCLE","equals":9}`

REA kept the task because its checks fail on the seed (`nine-circles`) 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-holes --agent claude-code --model MODEL