Resolve PR Reviews

Use this skill to process review comments on a PR. Pass the PR number as an argument (e.g. /resolve-pr-reviews 1041).

Workflow

Step 1: Fetch unresolved review comments

Use the GitHub API to get all review comments that haven't been resolved:

# Get inline review comments (pull request review comments)
gh api repos/{owner}/{repo}/pulls/{pr}/comments --jq '.[] | select(.position != null or .line != null)'

# Get issue-level comments (general PR comments from reviewers, not bots)
gh api repos/{owner}/{repo}/issues/{pr}/comments

Filter out:

• Deployment/CI bots (Vercel deploy previews, CI status checks)
• Already-resolved threads
• Your own previous replies

Keep:

• AI review comments (Claude review bot, CodeRabbit inline suggestions) — these are actionable review feedback
• Human reviewer comments

Step 2: Analyze each comment

For each unresolved comment:

1. Read the relevant code being commented on
2. Understand the reviewer's concern
3. Propose a concrete fix OR explain why it should be skipped
4. Categorize severity: must fix, good idea, or skip (with reasoning)

Step 3: Present to user

IMPORTANT: You MUST ask the user this question BEFORE showing any analysis. Do NOT skip this step or present comments prematurely.

After fetching and analyzing comments internally, use the AskUserQuestion tool to present a selection prompt:

Question: "Found N review comments on PR #XXXX. How would you like to go through them?"
Options:
  - "One by one" — Present each comment individually, user decides fix/skip before next
  - "All at once" — Present all comments together, user reviews full list

Wait for the user's selection before proceeding. Then:

• "one by one" (default): Present each comment individually with your analysis and proposed solution. Wait for user to decide "fix" or "skip" before moving to the next one.
• "all at once": Present all comments together with your analysis. User reviews the full list, then says which to fix and which to skip.

For each comment, show:

• The reviewer's comment (abbreviated)
• Your own independent analysis — don't just parrot the reviewer. Verify if the concern is actually valid, check the relevant code/dependencies, and explain what's really happening. If the reviewer is wrong or partially wrong, say so.
• Your proposed fix (code diff) or skip reasoning
• Your recommendation

Step 4: Apply fixes

Apply all approved fixes to the codebase.

Step 5: Commit and push

Run the /commit skill to format, lint, typecheck, stage, and commit. Then push:

git push

Step 6: Reply to comments on GitHub

Reply to each comment using the correct GitHub API endpoints:

For inline review comments (pull request review comments):

# Reply to an inline review comment (creates a reply in the same thread)
gh api repos/{owner}/{repo}/pulls/{pr}/comments/{comment_id}/replies \
  --method POST \
  -f body="<reply text>"

For issue-level comments (general PR comments):

gh api repos/{owner}/{repo}/issues/{pr}/comments \
  --method POST \
  -f body="<reply text>"

Reply content:

• If fixed: "Fixed in <commit_sha>." (keep it short)
• If skipped: Brief explanation of why (1-2 sentences)
• Tag the reviewer with @username when replying to top-level comments

Important Notes

• Never guess comment IDs — always fetch them from the API first
• Test the reply endpoint — pulls/{pr}/comments/{id}/replies is for inline review comment threads. issues/{pr}/comments is for general PR comments.
• Don't reply to deploy/CI bots — skip Vercel deploy previews, CI status comments. DO reply to AI review bots (Claude, CodeRabbit inline suggestions).
• Keep replies concise — reviewers don't want essays