Mr Review

MR Review: Code Analysis + Pipeline Diagnostics

Review a GitLab merge request end to end. The output is a structured assessment that covers both what the code does and whether the pipeline is healthy enough to merge.

Input Parsing

Accept any of:

• Full GitLab URL: https://community.opengroup.org/osdu/platform/system/search-service/-/merge_requests/845
• Short reference: search-service MR 845, search !845
• MR number alone: 845, !845 (requires current repo context or prior context)

URL Parsing

From https://community.opengroup.org/osdu/platform/system/search-service/-/merge_requests/845:

• Host: community.opengroup.org
• Project path: osdu/platform/system/search-service
• MR number: 845

CRITICAL: glab defaults to gitlab.com. For OSDU, ALL glab api commands MUST use:

GITLAB_HOST=community.opengroup.org glab api ...

Or use --hostname community.opengroup.org if supported by the subcommand.

Prerequisites

• glab authenticated to the target GitLab instance
• If glab is not authenticated, fall back to web fetch for public repos

Phase 1: Fetch MR Context

Gather everything needed before analysis. Run these in parallel where possible.

1a. MR Metadata

glab api "projects/<url-encoded-path>/merge_requests/<mr-number>" \
  --hostname community.opengroup.org

Extract: title, author, description, source branch, target branch, labels, milestone, state.

1b. Diff

glab api "projects/<url-encoded-path>/merge_requests/<mr-number>/changes?access_raw_diffs=true" \
  --hostname community.opengroup.org

If the diff is too large for a single API call, use the .diff endpoint:

# Via web fetch as fallback
https://<host>/<project-path>/-/merge_requests/<number>.diff

1c. Pipeline Status

glab api "projects/<url-encoded-path>/merge_requests/<mr-number>/pipelines" \
  --hostname community.opengroup.org

Take the most recent pipeline and fetch its jobs:

glab api "projects/<url-encoded-path>/pipelines/<pipeline-id>/jobs?per_page=100" \
  --hostname community.opengroup.org

Child/downstream pipelines (common in OSDU trusted-tests pattern): Look for jobs where the stage is trigger or the job name contains trigger-trusted. These bridge jobs have a downstream_pipeline object in the JSON response containing the child pipeline ID. Fetch the child pipeline's jobs using that ID. The real test results are usually in the child pipeline, not the parent.

Phase 2: Code Analysis

Analyze the diff with these goals:

Change Categorization

Group changes into areas:

Code Review Checklist

Scan for:

• Security: hardcoded secrets, credential handling, injection vectors
• Behavioral changes: null handling differences, exception type changes, new error paths
• API contract: header changes, endpoint modifications, response shape changes
• Cross-provider consistency: if one provider is changed, are others updated where needed
• Test coverage: are new code paths tested, are test assertions meaningful

Summary

Write a concise summary of what the MR does and why. Lead with the problem being solved, then describe the approach. Keep it to 3-5 sentences max.

Phase 3: Pipeline Diagnostics

This is the differentiator. For each failed job in the pipeline:

3a. Get Job Logs

glab api "projects/<url-encoded-path>/jobs/<job-id>/trace" \
  --hostname community.opengroup.org

Read the last 80 lines to find the failure reason. If the tail only shows cascading failures (e.g., "BUILD FAILURE" with no root cause), search upward for the first error. Look for patterns like ERROR, FAILED, Exception, error:, or test assertion failures.

3b. Classify Each Failure

Categorize every failed job:

How to classify:

• MR-caused: The failure message references files or logic changed in the MR, or a metric (like coverage) dropped because of newly added code.
• Pre-existing: Check the target branch's latest pipeline. If the same job fails with the same error there, it is pre-existing.
• Environment: The error is about infrastructure (DNS, cluster connectivity, service timeouts, certificate errors) not about the code being tested.
• Transient: Similar to environment but the error pattern is intermittent (resource contention, network blip). If the same job passed in a previous pipeline run on the same branch, it is likely transient.

For MR-caused failures, explain what specifically in the MR triggered them and what the fix would look like.

3c. Pipeline Summary Table

Pipeline #XXXXX: <status>

| Job                    | Status  | Classification | Details                    |
|------------------------|---------|----------------|----------------------------|
| compile                | passed  |                |                            |
| unit-tests             | passed  |                |                            |
| core_code_coverage     | failed  | MR-caused      | 78% vs 80% threshold       |
| azure-acceptance-test  | failed  | Environment    | 400s on record ingestion   |
| ibm-deploy             | failed  | Environment    | DNS resolution failure     |

3d. Blockers vs Non-Blockers

Clearly separate:

• Blockers: failures the MR author needs to fix before merge
• Non-blockers: failures unrelated to the MR (environment, pre-existing, transient)

Phase 4: Verdict

Provide a clear recommendation:

Include specific next actions for anything that isn't "approve."

Phase 5: Draft Comment

Generate a comment for the MR. Do NOT post it automatically. Show the draft to the user and wait for approval or edits.

Comment Tone

The comment must read like a developer wrote it. Follow these rules strictly:

• Write short, direct sentences. No filler.
• Do not use em dashes as punctuation. Use periods or commas instead.
• Do not open with praise ("Great work!", "Nice MR!"). Just get into it.
• Do not use phrases like "It's worth noting", "I noticed that", "Just a thought".
• No excessive bullet lists when a short paragraph works fine.
• Technical, factual, concise. The kind of comment a senior developer leaves in a real review.
• If there's nothing meaningful to say, don't pad the comment. Shorter is better.

Good example:

The cache fix looks correct. Each provider now resolves cluster settings by partition ID instead of pulling from request scope, which eliminates the poisoning vector.

One pipeline blocker: core_code_coverage fails at 78% (threshold 80%). Your new code is fully covered, the gap is pre-existing untested classes. I've submitted !850 with tests for the three cache delegation wrappers to bring it above the threshold.

Bad example:

Great work on this MR! I noticed that the changes look solid — the fix for the cache poisoning issue is well-implemented. It's worth noting that the pipeline has some failures, but most of them aren't caused by your changes. Here are my detailed findings:

• The code changes are consistent across all providers
• The test coverage is good
• ...

Comment Structure

For simple reviews: a short paragraph or two is fine.

For complex reviews with pipeline issues:

1. Brief assessment of the code (2-3 sentences)
2. Pipeline status with focus on blockers
3. Any specific action items or offers to help

Fallback: No glab Auth

If glab is not authenticated for the target instance:

1. Use web fetch to get the MR page and diff (works for public OSDU repos)
2. Pipeline details may be limited. Note this in the output.
3. Posting comments requires glab auth. Draft the comment but tell the user they'll need to post it manually or authenticate glab first.

Integration

After the review:

• If the user wants to contribute fixes, suggest the contribute skill
• If the user says "allow it" or "sync trusted", hand off to the maintainer skill
• If the user wants to store the review, offer the brain skill for vault storage

Posting the Comment

Only when the user explicitly approves:

glab mr note <mr-number> \
  --repo <project-path> \
  -m "<approved comment text>"