Review
Intelligent code review agent
Overview
Review is a precise, vigilant, detail-obsessed code review agent. It reviews GitLab merge requests thoroughly across six scoring categories, posts inline comments with severity levels, manages discussion threads, and checks CI/CD pipeline results. Review has a real opinion — it does not just agree for the sake of agreeing.
Every review produces a weighted score that determines whether the MR is approved, needs changes, or is blocked. Review resolves its own outdated findings, replies to other reviewers, and never duplicates existing feedback.
How to trigger
Assign Review to a GitLab merge request. Review reads the full MR diff, all existing discussions, and the CI pipeline status before forming any findings.
Scoring rubric
Review scores every review across six categories. Each category has a weight that contributes to the overall score.
| Category | Weight | What it evaluates |
|---|---|---|
| Correctness | 25% | Code logic, error handling, edge cases |
| Product Behavior | 20% | Matches requirements, UX impact |
| Security | 15% | Auth checks, injection prevention, data exposure |
| Code Quality | 15% | Style, naming, patterns, performance |
| Completeness | 15% | Nothing missing, docs updated |
| Tests | 10% | Coverage, quality of tests |
Score thresholds
| Score | Status | Meaning |
|---|---|---|
| ≥ 90 | Approved | Code is high quality and ready to merge (no single category below 40) |
| 75 – 89 | Review Complete | Good quality with minor findings that may not block merge |
| 50 – 74 | Changes Requested | Issues found that should be addressed before merge |
| < 50 | Blocked | Significant issues or any single category below 40 |
What it reads
- MR details (title, description, author, branches, state)
- Full diff of all changed files
- All existing discussion threads and comments (from all reviewers)
- CI/CD pipeline status and job logs for failed jobs
- Source files from the repository when the diff alone lacks sufficient context
What it produces
- Overall weighted score across all six categories
- Individual category scores with a markdown table
- Inline comments on specific lines, each prefixed with a severity level:
- Must-fix — Issues that must be resolved before merge
- Should-fix — Issues that should be addressed but may not block
- Risk assessment with confidence level (high, medium, or low)
- Brief review summary covering what was found, what is good, and what needs work
Discussion management
Review is a collaborative reviewer that respects existing conversations. Before posting any new comments, it builds a coverage map of every existing discussion — who said what, on which file and line, and whether it is resolved or open.
- Own previous comments: Resolves findings that the author has fixed. Replies with follow-up if partially addressed. Never re-posts findings that are still open.
- Other reviewers' comments: Replies with agreement or disagreement. Adds context when relevant. Never resolves another reviewer's thread.
- Deduplication: If an existing discussion already covers the same issue, Review replies to that thread instead of creating a new one. No duplicate noise.
Pipeline checking
After reviewing the code, Review checks the CI/CD pipeline. If jobs have failed, it reads the error logs and determines whether the failure is caused by the MR changes or is a pre-existing/flaky issue. Pipeline failures caused by the MR lower the score, especially for correctness and completeness. Review can retry transient failures (flaky tests, network timeouts) but never retries failures caused by actual code errors.
Status values
approvedScore is 90 or above with no single category below 40.
review_completeReview finished with a score of 75-89.
changes_requestedIssues found. Score is 50-74.
blockedSignificant issues. Score is below 50 or any single category is below 40.
errorSomething went wrong during the review.
Advisory, not authoritative
Review's score is advisory. Human reviewers always make the final call on whether to merge. The score provides a consistent baseline and ensures nothing is missed, but it does not replace human judgment.
Limitations
- GitLab only.
- Reviews code changes only. Review does not inspect deployment configurations, infrastructure, or runtime behavior.
- Only reviews new or changed code in the MR diff. Pre-existing code patterns that were not modified are not flagged.