earakely-scale/JabRefFromRequestComputerEnv-REA

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

The real JabRef 3.8 reference manager 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 33 minutes, with only REA's build skill and MCP tools; How it was made has the request and the steps. JabRef opens a small bibliography, Documents/references.bib, at launch: four synthetic references, two articles, a book and a workshop paper. The Okafor and Lindqvist article, Seasonal Sediment Transport in Glacial Rivers, is dated 2017. The agent has to correct its year to 2019 and save. 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 JabRef, correct the year of the Okafor and Lindqvist article "Seasonal Sediment Transport in Glacial Rivers" in Documents/references.bib to 2019 and save it.

bundle/artifacts/correct-sediment-year-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
check-1	lines_present	Documents/references.bib
check-2	lines_absent	Documents/references.bib
guard-1	lines_present	Documents/references.bib

A .bib file is BibTeX text, and REA reads it as lines. The checks need the line year = {2019},, no line year = {2017},, and the article's journal and volume lines kept. All three are REA's draft, kept exactly as written.

Results

3 of 3 runs passed, about $0.73 per run.

Run	Outcome	Turns	Cost	Video
1	passed	11	$0.78	video
2	passed	13	$0.82	video
3	passed	14	$0.60	video

Run it

You need Docker. agentenv-framework builds the computer: Ubuntu 24.04 with the app installed from Ubuntu's archive (jabref 3.8.2).

pip install agentenv-framework==0.9.1431
hf download earakely-scale/JabRefFromRequestComputerEnv-REA --repo-type dataset --local-dir jabref-request
agent-env run jabref-request/bundle --task correct-sediment-year --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 JabRef environment where the agent corrects a year in a small bibliography and saves it.

It asked only the skill's upfront questions, wrote no app-specific scripts, and finished in about 33 minutes.

About 9 of those minutes were spent waiting on agentenv's lock. The steps were REA's usual ones:

1. Capture. REA launched JabRef 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: the article's year is 2017
after_edit	pass
near_miss	fail, as expected: the year was typed as 2016
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 checks read the file as lines and can't name a BibTeX entry. They are tied to the Okafor article only because 2017 appears nowhere else in the seed and 2019 nowhere at all, and the lines' spacing follows JabRef's alignment of the entry's fields, set by its longest field name.
• The seed also carries JabRef's backups of the bibliography, references.bib.bak and references.bib.sav, beside the document; the checks read only references.bib.
• The computer installs whatever jabref Ubuntu 24.04's archive holds at build time (3.8.2 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. JabRef itself is not included: the computer installs it from Ubuntu's archive (jabref 3.8.2) at build time, under its own licence (Expat, the MIT licence).