Minitap published mobile-use under Apache 2.0 as a research experiment in getting agents to reliably drive phones. In September the team opened Google’s Artemis repository and recognised their own work immediately: “What the fuck? We wrote this.” Nicolas Dehandschoewercker’s essay is a first-person account of what was taken, and an argument that open source depends on people being willing to keep sharing — which depends on how that work is treated once it travels.
What they say they found, with comparisons published at pinned commits:
- Android device-connection code matching theirs nearly line for line
- The
Hopperagent’s instruction file identical word for word — including the name, chosen only because one engineer liked Minecraft - A messaging example with the same Happy New Year task for Alice, Bob and Charlie, the same comments, same cleanup steps
- A shared historical bug: a helper wrote a results file, then failed reading its own output on the next run
Then the part that is harder to explain as coincidence:
- An earlier Artemis
pyproject.tomllisted three Minitap engineers as authors - The version that replaced it dropped all three and substituted another name — the author list was the only change, and GitHub’s activity record shows a force push in August, before Minitap looked
- The README they checked credited
mobile-usenowhere; Minitap published a factual record with the diffs and timeline
The licence sets a floor, and ordinary practice sets a higher one:
- Apache 2.0 section 4 asks redistributors to preserve copyright and attribution notices, identify changes, and retain upstream
NOTICEattribution - A fork links back to the original; an imported copy documents its source; a README says which parts came from elsewhere and what the new team added
- Attribution is not just credit — it is how the next person finds the maintainers, understands a design decision, or reports a bug
A second complaint sits alongside the code: months of unanswered emails asking AndroidWorld’s leaderboard to reflect newer results (91.4% accepted in December, 94.8% and 100% submitted in January with four emails and no reply), and an Artemis comparison chart that omitted mobile-use while listing an unrelated project of the same name. The author is explicit that he has no evidence connecting the leaderboard silence to the removed names — he separates what is documented from what is merely felt.
That restraint is what makes the rest land. The remedy for a small team here is its own unpaid labour: recovering detached revisions, diffing files, publishing a record, opening an issue. Minitap’s ask is narrow — acknowledge the derivation, credit the people, put the attribution right. The case for doing it is not reputational but structural: attribution is the return that makes publishing worth it, and an ecosystem where maintainers have to chase their own names is one that quietly stops sharing.