earakely-scale/KmPlotFromRequestComputerEnv-REA

KmPlot: a computer-use environment built from a one-line request

The real KmPlot 23.08 function plotter running on an Ubuntu 24.04 desktop, as an agentenv computer environment. An agent completes the task only through screenshots, mouse and keyboard, and a verifier grades the saved files with no model.

Claude Code built this environment from a one-line request in about 21 minutes, with only REA's build skill and MCP tools; How it was made has the request and the steps. KmPlot opens a small plot file, Documents/lesson_plots.fkt, at launch: three functions, f(x) = x^2 − 4, g(x) = 2sin(x) and h(x) = 0.5x + 1. The agent has to change h's formula to h(x) = 2x + 1 and save the plot file. It sees only the screen and acts only with the mouse and keyboard.

Watch

• Agent runs (Claude Code with Claude Sonnet 5.5): run 1, run 2, run 3
• Witness, which proves the task can be done through the computer: witness.mp4

Each video is the computer's own screen recording at 10 frames per second.

The task

In KmPlot, change the formula of the function h in Documents/lesson_plots.fkt to h(x) = 2x + 1 and save the plot file.

bundle/artifacts/change-h-formula-check/verify.py reads the app's files through the computer's data/get view and passes only if every check holds:

Check	Kind	File
h-new-formula	line_count	Documents/lesson_plots.fkt
h-old-formula-gone	line_count	Documents/lesson_plots.fkt
g-unchanged	line_count	Documents/lesson_plots.fkt
f-unchanged	line_count	Documents/lesson_plots.fkt
three-functions	line_count	Documents/lesson_plots.fkt

A .fkt file is XML, but it opens with <!DOCTYPE kmpdoc> and no XML declaration, and REA read it as text lines rather than XML. So the checks count lines: exactly one must contain >h(x) = 2x + 1< and none >h(x) = 0.5x + 1<, g's and f's equations must each still appear once, and the file must still hold three <function lines. The f check reads only >f(x) = x^2, because KmPlot stores the minus in x^2 − 4 as U+2212, not a hyphen. REA's draft suggested checks on each function's whole line, whose attributes Qt writes in a different order in every process, and a guard that failed a correct edit in a fresh process for that reason; Claude Code narrowed two suggestions to the equation text, discarded the guard and wrote the rest.

Results

3 of 3 runs passed, about $0.68 per run.

Run	Outcome	Turns	Cost	Video
1	passed	11	$1.02	video
2	passed	14	$0.55	video
3	passed	10	$0.47	video

Run it

You need Docker. agentenv-framework builds the computer: Ubuntu 24.04 with the app installed from Ubuntu's archive (kmplot 23.08.5).

pip install agentenv-framework==0.9.1431
hf download earakely-scale/KmPlotFromRequestComputerEnv-REA --repo-type dataset --local-dir kmplot-request
agent-env run kmplot-request/bundle --task change-h-formula --agent claude-code --model <model>

The agent needs a model provider configured for agentenv. --task smoke only boots the computer, with no agent. bundle/README.md describes the bundle in detail.

How it was made

Claude Code built this environment from a one-line request, given only REA's build skill and REA's MCP tools:

Build me a KmPlot environment where the agent changes the formula of one plotted function and saves the plot file.

It asked only the skill's upfront questions, wrote no app-specific scripts, and finished in about 21 minutes. The steps were REA's usual ones:

1. Capture. REA launched KmPlot on a Linux desktop and recorded its package, windows and opened files while the edit was made. It also recorded the app's state folders before and after.
2. Export. REA turned the capture into an agentenv computer_app. It installs the app, seeds the starting files, starts the app, and waits for its main window. It also serves typed views of the app's files for verifiers.
3. Task. The task's checks were validated against the capture: they fail on the starting state and pass on the captured end state.
4. Witness. In a fresh computer, the edit was made through computer actions alone, and the checks were read at four stages. They held on the first try:

Stage	Checks
seed	fail, as expected: h is still 0.5x + 1; the kept checks pass
after_edit	pass
near_miss	fail, as expected: the formula was typed as h(x) = 2x + 2
after_reset	fail, as expected: data/reset restored the seed

The agent runs came after the build: Claude Code with Claude Sonnet 5.5 (--model anthropic/claude-sonnet-5-5 --no-emulate, on an amd64 host) solved the task 3 times, each run recorded.

What's inside

bundle/      the agentenv bundle: envs/ (app module, seed files), tasks/, artifacts/ (verifier), evals/
media/       witness and agent-run recordings (MP4) with short GIF previews
results.json the runs above, with each check's result

Limits

• Because REA read the file as text, the checks are line counts, not checks on the XML's function elements. They need the equation exactly as the prompt writes it, h(x) = 2x + 1; an equivalent formula written differently, such as 2*x + 1, fails.
• KmPlot stores the minus sign as U+2212 (−), not a hyphen, so a check on f's whole equation would need that character; the f check reads only the part before it.
• The computer installs whatever kmplot Ubuntu 24.04's archive holds at build time (23.08.5 for these runs); it is not pinned.
• Three runs is a small sample: the solve rate shows the task is solvable, not how hard it is.

Licences

The bundle code, the task and the synthetic seed files are Apache-2.0. KmPlot itself is not included: the computer installs it from Ubuntu's archive (kmplot 23.08.5) at build time, under its own licence (GPL-2.0-or-later).