Review a pull request
SkillDev toolsLets your agent review pull requests in the Marin repo, a review Claude skill that posts findings as GitHub comments.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Review a pull request skill
About this capability
Review an explicitly identified Marin pull request against the repository's correctness and maintainability expectations.
What this skill tells your AI
The instructions your AI receives, as published by marin-community/marin in .agents/skills/review-pr/SKILL.md and read by ahel’s review.
Review the requested pull request at the exact head named by the caller. The
review is read-only except for comments posted when --comment is requested.
Review standard
A reviewable Marin pull request:
- Implements the behavior claimed by its title and description without an introduced correctness, security, state-transition, or error-handling defect.
- Follows the root and path-scoped
AGENTS.mdorCLAUDE.mdinstructions. If tests change, it also follows rootTESTING.mdand any scoped testing guide. - Keeps public interfaces, invariants, dependency direction, configuration, and types coherent across every changed call site. Marin does not preserve backward compatibility unless the request explicitly requires it.
- Uses tests for behavior rather than implementation details, incidental prose, private state, tautologies, or mocks below an I/O boundary.
- Keeps abstraction proportional to demonstrated reuse. Flag a maintainability problem only when the change creates a concrete structural obstacle, hidden assumption, or distant coupling that materially constrains nearby work.
- Keeps documentation synchronized with behavior. The pull-request title and
body must follow
.agents/skills/writing-style/pull-requests.md; the body is the squash-merge commit message, not a template, file inventory, or test log.
Report only actionable defects introduced by the pull request. Validate every
finding against the diff, relevant surrounding code, and applicable
instructions. Do not report style preferences, speculative failures, harmless
duplication in experiments/grug, or pre-existing problems. If the evidence is
uncertain, omit the finding.
Workflow
-
Inspect the pull request metadata, issue comments, inline comments, and current head. Stop without commenting when the pull request is closed, a draft, an automated dependency update, or too trivial to benefit from code review. Still review agent-generated pull requests.
-
Check both issue and inline comments for
<!-- marin-correctness-review -->. Stop when a prior automated correctness review is present unless a maintainer explicitly requested another review. -
Confirm that the checked-out commit and the pull request's current head both match the requested head. If either differs, report the stale review and post nothing.
-
Read the root instructions and every instruction file that scopes a changed path. Read the root and scoped testing guides when tests change. Inspect the complete diff and enough surrounding code to validate behavior and call-site consistency.
-
Review the code against the standard above. For each finding, record its changed file and line, category (
bug,instruction-following, ormaintainability), concrete impact, and the evidence that makes it certain. Separately record title or description problems. -
Emit one best-effort stats event before returning, including clean and non-commenting reviews. Never retry or surface a telemetry failure. Run from the repository root:
cat <<'EOF' | uv run infra/codehealth/log_stats.py { "tool": "review-pr", "invocation": { "trigger": "local", "agent_cli": "codex", "pr_number": <PR>, "agent_exit_code": 0, "timed_out": false }, "findings": [ ["<file>", <line>, "<category>", 1.0, "<first 200 chars of issue description>"] ] } EOFUse an empty
findingsarray when the code review is clean. -
Print the findings or
No issues found.and list pull-request-description problems separately. Without--comment, stop here. -
With
--comment, post one top-level comment for all title or description problems. Begin with🤖, name each concrete problem and fix, and end with<!-- marin-correctness-review -->. -
If the code review is clean and step 8 did not post a description comment, post exactly this top-level comment. Otherwise stop after the description comment.
🤖 Code review No issues found. <!-- marin-correctness-review --> -
Otherwise, post one inline comment per unique code finding. Begin each body with
🤖, explain the defect and its impact, and end with<!-- marin-correctness-review -->. Include a committable suggestion only when that suggestion completely fixes a small, self-contained issue. Use a full-SHA GitHub link with surrounding context when citing repository instructions or related code.
Use gh for GitHub reads and top-level comments. Post inline comments with the
GitHub inline-comment tool and confirmed: true.
Signals
- GitHub stars
- 4k
- Forks
- 303
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
review-pr-marin-community- Source
- github.com/marin-community/marin