earakely-scale/ParleyFromRequestComputerEnv-REA

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

The real Parley 23.08 vocabulary trainer 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 25 minutes, with only REA's build skill and MCP tools; How it was made has the request and the steps. Parley opens a small vocabulary collection, Documents/spanish.kvtml, at launch: Spanish Basics, with one lesson of six English–Spanish word pairs, dog/perro, cat/gato, house/casa, water/agua, book/libro and apple/manzana. The agent has to add the pair library/biblioteca and save the collection. 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 Parley, add a new word pair to my Spanish Basics vocabulary collection in Documents/spanish.kvtml: the English word "library" with the Spanish "biblioteca". Then save the collection.

bundle/artifacts/add-library-word-pair-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
pair-spanish	text_equals	Documents/spanish.kvtml
pair-english	text_equals	Documents/spanish.kvtml
one-more-entry	member_count	Documents/spanish.kvtml
apple-kept	text_equals	Documents/spanish.kvtml
book-kept	text_equals	Documents/spanish.kvtml

A .kvtml file is XML, with one entry per word pair and a translation per language (id 0 English, id 1 Spanish). Parley numbers entries in save order, so the checks find entries by their translations ([>translation=library], a selector on a child element's text): the library entry's Spanish must read "biblioteca" and the biblioteca entry's English "library", the collection must hold seven entries (member_count), and book and apple must keep libro and manzana. REA's draft keyed the new entry by its id; Claude Code re-keyed it and the two guards by their translations, kept the entry count as REA wrote it, and discarded two suggestions on the lesson's list of entry references.

Results

3 of 3 runs passed, about $0.57 per run.

Run	Outcome	Turns	Cost	Video
1	passed	9	$0.99	video
2	passed	9	$0.50	video
3	passed	9	$0.23	video

Run it

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

pip install agentenv-framework==0.9.1431
hf download earakely-scale/ParleyFromRequestComputerEnv-REA --repo-type dataset --local-dir parley-request
agent-env run parley-request/bundle --task add-library-word-pair --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 Parley environment where the agent adds a new word pair to a Spanish vocabulary collection and saves it.

It asked only the skill's upfront questions, wrote no app-specific scripts, and finished in about 25 minutes. The steps were REA's usual ones:

1. Capture. REA launched Parley 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 pair was added through computer actions alone, and the checks were read at four stages. They held on the first try:

Stage	Checks
seed	fail, as expected: there is no library pair and six entries; the kept checks pass
after_edit	pass
near_miss	fail, as expected: the Spanish was retyped as "bibloteca"
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

• The new pair is found by its words, which must be exact and in the right columns, "library" in English and "biblioteca" in Spanish. Whether it belongs to Lesson 1 is not checked.
• Only book and apple are checked as kept; the other four pairs are not.
• The computer installs whatever parley 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. Parley itself is not included: the computer installs it from Ubuntu's archive (parley 23.08.5) at build time, under its own licence (GPL-2.0-or-later).