- Use the [PR template](.github/PULL_REQUEST_TEMPLATE.md). Explain the problem, the resulting behavior, and the checks actually run. Keep routine descriptions short; include more detail when needed to assess correctness or risk.
- Every PR must state whether AI was used. This includes help with investigation, implementation, tests, documentation, translation, review, or the PR description itself. If no AI was used, write `None` in the AI assistance section.
- When AI was used, name each actual model and the tool used to access it. A tool or provider name alone (such as Codex, ChatGPT, or Copilot) is not a model identity. Use the identity supplied by the runtime or contributor; if the exact model or variant is unavailable, say `Unknown (via <tool>)` or identify the known model family and state that the exact variant is unknown. Do not guess.
- Describe each model's role concretely: for example, investigating the failure, implementing a named fix, writing regression tests, reviewing the diff, or drafting the description. Distinguish AI assistance from work checked by the contributor.
- Explain how the AI-assisted output was verified. Report actual commands, manual checks, outcomes, and any checks not run; an AI review alone is not evidence that the change works. Refer to the Validation section rather than repeating test logs.
- An agent drafting or updating a PR must fill in this disclosure for its own contribution and other known AI contributions. Keep it accurate as the scope changes. Do not invent another model's identity, a human review, or a passing test. Do not include private prompts or chat transcripts.
- The person submitting the PR remains responsible for understanding and reviewing the full diff, checking the result, and responding to review. Maintainers may request missing disclosure or verification before proceeding with review.
AGENTS.md ↗