Code review
Repository guidelines
Teach Codebot your team’s conventions with files in the repository it reviews.
Codebot reads guideline files from each repository and follows them in its reviews, the same way a coding agent reads them when it writes code.
Supported files
| File | Use it for |
|---|---|
AGENTS.md | Conventions for any agent working in the repository: architecture, naming, patterns to follow or avoid. |
CLAUDE.md | The same, if your team already keeps one for Claude Code. |
.codebot/review.md | Rules only for reviews: what to flag, what to skip, how strict to be. |
All three are optional and live at the root of the repository. Codebot reads the ones it finds.
Writing review rules
Write .codebot/review.md the way you’d brief a new reviewer on the team. Be specific about what matters and why.
# Review rules
- Every new API route needs an authorization check. Missing checks are high severity.
- Database migrations must be backwards compatible: no dropped or renamed
columns in the same release as the code that stops using them.
- Don't comment on import order or formatting; CI runs the formatter.
- Tests under `e2e/` are slow by design. Don't suggest speeding them up.For instructions that apply to every repository, use the Instructions field in review settings.
Which version Codebot reads
Codebot reads guideline files from the pull request’s base commit, never from the pull request itself. A pull request can’t rewrite the rules it’s reviewed by; changes to guideline files are reviewed like any other change, and apply to later pull requests once merged.
Limits
- Each file is read up to 20,000 characters, and all of them together up to 40,000.
- Guidelines shape what the reviewer looks for. They can’t change how findings are reported, the severity scale, or the review outcome you set.
- Per-repository settings (skip rules, reviewer model, merge gate) aren’t supported yet; they apply to all repositories.