This isn't a statement for or against AI agents - people are using them here either way. It is only there to guide them better, so the project rules (no GNU code, tests required, short PR descriptions, read the docs) are actually followed. See https://github.com/uutils/coreutils/blob/main/CONTRIBUTING.md#ai-policy
1.7 KiB
AGENTS.md
Writing this file is not a statement for or against using AI coding agents. People are using them on this repository either way, so this file exists to make sure that when they do, the rules of the project are actually followed.
If you are driving an agent here, you are responsible for its output. Read the diff before you send it.
Answers to the reviewers should be done by a human, not a agent.
1. Never read or copy GNU code
GNU coreutils is GPLv3; this project is MIT. Any code derived from it - even a few lines, a helper structure, a test fixture or a comment - cannot be accepted.
2. A PR needs tests
New behavior or a bug fix comes with a test, in tests/by-util/test_<util>.rs
or as a unit test next to the code. If a GNU test used to fail and now passes,
add a Rust test so it cannot silently regress.
No tests, no merge.
3. Keep the PR description short
Describe the problem being solved and what changed. That is all.
- No generated walls of text, no bullet-point summaries of every hunk, no emoji-headed sections.
- Title:
<util>: <what changed>, e.g.ls: fix version sort order. - Write issue reports, PR descriptions and replies to reviewers in your own words. The point of review is to check that a human understands the change.
4. Read the docs already in this repo
Before changing anything, look at the Markdown files here - they are the actual rules, this file is only a pointer:
CONTRIBUTING.md- licensing, commit hygiene, what a PR must containDEVELOPMENT.md- build, test and environment setupREADME.md,SECURITY.md,CODE_OF_CONDUCT.mddocs/src/*.md- platforms, performance, l10n, multicall, test coverage- the per-utility
README.md/locales/*.ftlfiles when touching a utility