Oss Post Pr

Post-PR Review

Your PR got reviewed. Now what? This skill helps you understand what reviewers are actually asking for, why they're asking for it, and how to address their feedback - without writing the code for you.

Purpose

Review comments from maintainers are often terse. "Can you use X pattern instead?" doesn't explain WHY X is preferred, or WHERE to find examples of it. This skill bridges that gap: it researches the reviewer's concern, finds examples in the codebase, and explains the reasoning - so you can address feedback intelligently instead of blindly copy-pasting.

Prerequisites

• A submitted PR with review comments (from oss-submit-pr)
• The repo cloned locally with your working branch checked out

Process

1. Fetch all review feedback

bashCopy code
# Get PR reviews
gh pr view {pr-number} -R {owner}/{repo} --json reviews,comments,reviewDecision,statusCheckRollup

# Get review comments (inline code comments)
gh api repos/{owner}/{repo}/pulls/{pr-number}/comments --jq '.[] | {path: .path, line: .line, body: .body, user: .user.login, state: .state}'

# Get PR conversation comments
gh api repos/{owner}/{repo}/issues/{pr-number}/comments --jq '.[] | {body: .body, user: .user.login, createdAt: .created_at}'

2. Categorize feedback

Sort each piece of feedback into categories:

Present the categorized list:

Copy code
## PR Review Summary: #{pr-number}

**Overall decision**: {approved / changes requested / commented}

### Blocking
1. `src/foo.ts:42` - {reviewer}: "{comment}" → **Must fix**
2. ...

### Suggestions
1. `src/bar.ts:15` - {reviewer}: "{comment}" → **Evaluate**
2. ...

### Questions (you need to answer these)
1. {reviewer}: "{question}" → **Explain your reasoning**

3. Research each blocking comment

For each blocking review comment, investigate what the reviewer actually wants:

bashCopy code
# Find examples of the pattern they're suggesting
grep -rn "pattern_reviewer_mentioned" src/ --include="*.ts" --include="*.py" --include="*.go" --include="*.rs"

# Check if this is a documented convention
grep -r "pattern_name" CONTRIBUTING.md CLAUDE.md .eslintrc* .prettierrc* pyproject.toml

# Look at how similar code handles this elsewhere
# {specific search based on the comment}

Present findings for each blocking comment:

Copy code
### Comment: "{reviewer's comment}" at {file}:{line}

**What they want**: {plain language explanation}
**Why**: {the reasoning - found from codebase patterns, docs, or common practice}
**Examples in codebase**: 
- `src/similar.ts:30` - does it this way
- `src/other.ts:55` - another example
**What you need to change**: {describe the change needed, don't write the code}

4. Thinking gate - user explains each comment

Before the user starts fixing anything:

"For each blocking comment, tell me:

1. What is the reviewer asking for? (Look at the 'What they want' section I researched above)
2. Why do they want it that way? (Check the codebase examples I found - what pattern are they pointing to?)
3. Do you agree, or do you want to discuss it with them?"

This catches misunderstandings BEFORE the user writes code that still doesn't address the comment. Common failure: user makes a surface-level change that doesn't actually address the deeper concern.

If the user misunderstands a comment:

• Point to the specific example in the codebase: "Look at src/similar.ts:30 - see how they handle the same case? The reviewer wants you to follow that pattern because {reason}."
• Don't write the fix - show the pattern and let them figure it out.

5. Handle questions from reviewers

For review comments that are questions ("Why did you do X?"), the user MUST answer these themselves. But help them formulate a clear response:

"The reviewer asked why you chose X. Think back to when we investigated the issue - what was your reasoning? Write a clear response explaining your decision."

If the user struggles to explain their own code, that's a signal. Point them back to the relevant investigation from oss-contribute and ask what they remember.

6. User addresses feedback

The user fixes each blocking comment and responds to questions. During this phase:

What the LLM DOES:

• Answer questions about the codebase patterns the reviewer referenced
• Point to more examples if the user needs them
• Review the user's changes when asked (flag issues, don't fix)
• Help the user write response comments (review, don't write)

What the LLM DOES NOT DO:

• Write the code changes
• Write review response comments for the user
• Dismiss suggestions without the user evaluating them

Review response writing rules (enforce when reviewing user's draft responses):

• Answer the question directly. "I chose X because Y" - done
• No apologizing: "Sorry for the confusion" is noise. Just explain or fix
• No filler: "Great catch!" / "Thanks for pointing that out!" - one line max, then substance
• No AI jargon: "comprehensive", "robust", "leverages" - cut all of it
• If the response is longer than the reviewer's comment, it's probably too long
• Match the reviewer's tone: if they wrote two words, you don't need two paragraphs

7. Track progress

Keep a running status:

Copy code
## Review Status

| # | Comment | Category | Status |
|---|---------|----------|--------|
| 1 | "Use X pattern" | Blocking | ✅ Addressed |
| 2 | "Why not Y?" | Question | ✅ Responded |
| 3 | "Consider Z" | Suggestion | ⏳ User evaluating |
| 4 | "Separate PR" | Scope | ❌ Needs discussion |

8. Re-submit

Once all blocking comments are addressed:

bashCopy code
# Push changes
git push origin {branch-name}

# Respond to review thread (user writes, LLM reviews)
gh pr comment {pr-number} -R {owner}/{repo} --body "{user's response}"

# Request re-review if needed
gh pr edit {pr-number} -R {owner}/{repo} --add-reviewer {reviewer}

9. If significant rework is needed

If the reviewer requests a fundamentally different approach:

1. Don't panic - this is normal and educational
2. Research the suggested approach (same as step 3 but deeper)
3. Present the alternative approach with its trade-offs
4. Let the user decide whether to rework or discuss further with the reviewer
5. If reworking, hand off to → oss-contribute for a fresh investigation of the new approach

Related Skills

• Previous step: ← oss-submit-pr - the PR submission
• If significant rework: → oss-contribute - re-investigate with the new approach
• If minor fixes: → oss-submit-pr - re-verify and push updates
• If CI fails: → oss-debug-ci - diagnose and fix CI pipeline failures
• After merge: → oss-second-contribution - plan your next contribution and build trust
• Loop: This skill may be invoked multiple times per PR as new reviews come in

Common Rationalizations

Red Flags

• User is editing code without being able to say what the reviewer asked for
• A review comment contains a term the user cannot define. stop and use oss-learn-stack
• More than three review rounds on a small diff. the disagreement is about the approach, not the details, and it needs saying out loud
• User pushes back on every comment. it reads as defensive regardless of who is right
• Reviewer has gone quiet after the user argued. the tone landed badly, not the argument

Verification Checklist

• All feedback fetched, including inline comments and review-level bodies (step 1)
• Each comment categorized blocking, suggestion, or question (step 2)
• User explained in their own words what each blocking comment wants (step 4 gate)
• Reviewer questions answered directly rather than deflected into code changes (step 5)
• User wrote the changes. the LLM researched and reviewed (step 6)
• Every thread has a response, including ones declined with a stated reason
• CI green after re-submission (step 8)

Anti-patterns

• DO NOT dismiss review comments - every comment from a maintainer has a reason
• DO NOT write response comments for the user - they need to explain their own code
• DO NOT write the fix for review comments - show the pattern, user writes the code
• DO NOT argue with reviewers through the LLM - if the user disagrees, they should discuss directly and respectfully
• DO NOT batch all fixes into one response - address each comment individually with its own commit if the repo prefers it