Dev Review Pr

Check GitHub Actions status, review feedback from automated reviewers (Code Rabbit, Claude), implement or respond to feedback, and merge the PR when all feedback is resolved.

MANDATORY: Run /dev-local-review before every commit and push. This applies to step 5 (commit and push fixes) and any other commit made during this workflow. Do NOT push code that has not passed a local review. Skipping this step wastes a GitHub review round on issues that could have been caught locally for free.

The user provides:

• PR number: the pull request number to review (optional - infers from current branch)

Workflow

1. Check GitHub Actions status

Use gh pr checks <pr-number> to see if automated reviews are complete.

Important: This command returns exit code 8 if any checks are pending, and outputs them to stderr. Handle both stdout and stderr to capture all check statuses. Output format: check-name status duration url

If any checks are still running (pending):

• Report which checks are pending
• Exit with message: "GitHub Actions still running. Run this skill again when checks complete."
• DO NOT sleep or wait — let the user run the skill again later

If checks failed:

• Report which checks failed
• Show the failure URL from the output
• If markdown linting failed, offer to run it locally and fix issues (see step 1.5)
• Otherwise exit and ask user to investigate

If all checks passed:

• Proceed to step 2 (fetch review comments)

1.5. Fix markdown linting failures (if needed)

If the "Markdown Lint" check failed:

1. Run locally to see errors:

   npx markdownlint-cli2 "**/*.md"
   

2. Attempt auto-fix:

   npx markdownlint-cli2 --fix "**/*.md"
   

3. Manually fix remaining errors (especially MD040 - missing language tags)

4. Verify all errors resolved:

   npx markdownlint-cli2 "**/*.md"
   

5. Commit and push fixes:

   git add <fixed-files>
   git commit -m "Fix markdown linting errors
   
   Co-Authored-By: Claude <noreply@anthropic.com>"
   git push
   

6. Exit and wait for checks to re-run — user should invoke /dev-review-pr again after checks pass

1.7. Check CodeRabbit review status

CodeRabbit is configured to auto-pause after the first reviewed commit (reviews.auto_review.auto_pause_after_reviewed_commits: 1 in .coderabbit.yaml). This means after every push, CodeRabbit reviews once and then pauses. The "Reviews paused" banner in the PR comment is normal and expected — it does NOT mean something went wrong or that the review is missing.

Key insight: CodeRabbit completes its review before pausing. The review comments and review body are already available.

How to check:

Look at the CodeRabbit check status from step 1:

• If CodeRabbit shows pass — the review is done. Proceed directly to step 2 to fetch the review comments. Ignore the "Reviews paused" banner — it just means future commits won't be auto-reviewed.
• If CodeRabbit shows pending or no CodeRabbit check yet — the review hasn't started or is still running. Exit and ask the user to re-run later.

After implementing feedback and pushing fixes (step 5):

Since auto-pause is on, CodeRabbit will NOT automatically review the new commit. Post a single comment to request a fresh review:

gh pr comment <pr-number> --body "@coderabbitai review"

Do NOT post @coderabbitai resume — we want it to stay paused after each review to avoid review thrashing from rapid commits.

If CodeRabbit shows "Currently processing new changes":

• The review is actively running. Exit and tell the user to re-run later.
• Do NOT fetch review comments yet — they may be incomplete.

1.7a. Detect clean re-review (no actionable comments)

After CodeRabbit re-reviews a fix commit, it sometimes finds nothing new to flag. When this happens, CodeRabbit does not post an APPROVED review. Instead, it posts an issue comment (not a review comment) containing:

No actionable comments were generated in the recent review.

The review state from step 6 may still show CHANGES_REQUESTED or COMMENTED from a previous round — that's stale and misleading.

How to detect:

# Get the latest "no actionable comments" timestamp
NO_ACTION_TS=$(gh api repos/{owner}/{repo}/issues/{pr-number}/comments --paginate \
  | jq -r -s '(add // []) | map(select(.user.login == "coderabbitai[bot]"
    and (.body | contains("No actionable comments"))))
    | sort_by(.created_at) | last | .created_at // empty')

# Get the latest CHANGES_REQUESTED or COMMENTED review timestamp
LAST_REVIEW_TS=$(gh api repos/{owner}/{repo}/pulls/{pr-number}/reviews --paginate \
  | jq -r -s '(add // []) | map(select(.user.login == "coderabbitai[bot]"
    and (.state == "CHANGES_REQUESTED" or .state == "COMMENTED")))
    | sort_by(.submitted_at) | last | .submitted_at // empty')

# Compare timestamps (lexicographic works for ISO 8601 dates)
if [[ -z "$NO_ACTION_TS" ]]; then
  echo "No 'no actionable comments' signal found"
elif [[ -z "$LAST_REVIEW_TS" || "$NO_ACTION_TS" > "$LAST_REVIEW_TS" ]]; then
  echo "Clean re-review detected — proceeding to step 6b"
else
  echo "Review feedback still pending"
fi

If the "no actionable comments" comment is newer than the last review with feedback, CodeRabbit is satisfied.

When detected, skip straight to step 6b (pre-merge branch update and resolve). Do not loop back to step 2 looking for feedback that does not exist.

2. Fetch review comments

Important: Inline review comments (the "tasks" users see in the GitHub UI) are NOT returned by gh pr view --json comments. You must use the API directly:

gh api repos/{owner}/{repo}/pulls/{pr-number}/comments

This returns an array of review comment objects with:

• path — file path
• line / start_line — line numbers
• body — comment body (markdown, may include severity, suggestions)
• html_url — link to view the conversation
• user.login — reviewer (e.g., "coderabbitai[bot]", "claude-code[bot]")

CodeRabbit comment format:

• Body starts with severity: _⚠️ Potential issue_ | _🟠 Major_, _🟡 Minor_, or other markers
• Includes suggested fixes in <details><summary>Suggested fix</summary> blocks
• May include committable suggestions between <!-- suggestion_start --> markers
• Important: CodeRabbit may leave nitpick/style comments alongside major issues—fetch ALL comments, not just the first few

Parse and categorize feedback by:

• Severity (🟠 Major, 🟡 Minor, nitpick/style)
• File and line number
• Reviewer (CodeRabbit, Claude, human)

Present ALL comments to the user sorted by severity (Major → Minor → Nitpick) so security/critical issues are addressed first.

2a. Correctly identify which comments belong to the latest review round

CRITICAL — this is the #1 cause of missed feedback. Do NOT filter comments by original_commit_id to find "new" comments. CodeRabbit posts comments against various commit SHAs depending on diff positioning, and a comment on commit A may actually be new feedback triggered by commit B's review.

The correct approach: use the review ID, not the commit ID.

1. First, fetch all reviews and find the latest CodeRabbit review:

   REVIEW_ID=$(gh api --paginate repos/{owner}/{repo}/pulls/{pr-number}/reviews \
     | jq -s 'add | map(select(.user.login == "coderabbitai[bot]"))
              | sort_by(.submitted_at) | last | .id')
   echo "Latest CodeRabbit review: $REVIEW_ID"
   

2. Then, fetch comments and match them to reviews by pull_request_review_id:

   gh api --paginate repos/{owner}/{repo}/pulls/{pr-number}/comments \
     | jq -s --argjson rid "$REVIEW_ID" \
       'add | map(select(.pull_request_review_id == $rid))
            | map({id, path, line, body: .body[:200]})'
   

3. Comments from the latest review are the new comments for this round. Comments from earlier reviews that were NOT auto-resolved are carried forward and still need action.

Why original_commit_id filtering is wrong:

• CodeRabbit may post a comment against commit X even when reviewing commit Y (if the code at that location hasn't changed between commits)
• A single review round can produce comments with different original_commit_id values
• Filtering by the latest commit SHA misses comments CodeRabbit posted against older commit positions — this causes the agent to report "only 1 new comment" when there are actually 7

Alternative quick approach: If you don't want to join on review IDs, simply fetch ALL CodeRabbit comments, exclude those that have a human reply dismissing them (user said "out of scope"), and present the rest grouped by file. Let the user decide which are new vs. already handled.

2b. Check review bodies for duplicate and nitpick comments

Critical: CodeRabbit embeds additional feedback directly in review bodies — not as separate inline comment threads. These appear in collapsible sections titled "Duplicate comments" and "Nitpick comments" within the review. They do NOT create their own threads, so they will be missed if you only fetch inline comments.

Fetch review bodies from CodeRabbit:

gh api --paginate repos/{owner}/{repo}/pulls/{pr-number}/reviews \
  | jq -s 'add
    | map(select(.user.login == "coderabbitai[bot]"))
    | sort_by(.submitted_at)
    | map({id: .id, state: .state, body: .body, date: .submitted_at})'

Look for these sections in the review body:

• ♻️ Duplicate comments — Issues raised previously that CodeRabbit still considers unresolved after re-review. These are NOT resolved just because the original thread was addressed — CodeRabbit is saying the fix was incomplete or the issue persists. Treat these as active feedback.
• 🧹 Nitpick comments — Low-severity suggestions that didn't warrant a blocking review thread. Still present them to the user.

For each duplicate/nitpick comment found:

• Extract the file path, line numbers, severity, and description
• Include the suggested patch if present
• Add it to the feedback summary alongside inline comments
• Do not dismiss duplicate comments — they indicate CodeRabbit re-reviewed and still found the issue. The fix from the previous round was likely incomplete.

When a review body contains duplicate comments but no new inline threads:

• The review state will be COMMENTED (not CHANGES_REQUESTED)
• The inline comment threads from the previous round may be auto-resolved
• But the duplicate section means there is still work to do — do NOT skip to the merge step just because all threads show as resolved

3. Present feedback summary and classify into themes

Sort comments by severity: Major → Minor → Nitpick/Style, so critical issues (especially security) are addressed first.

Show the user the raw feedback list, then group comments into themes. A theme is a class of issue that may have multiple instances across the codebase. Examples:

## PR Review Feedback Summary

### Pending conversations: X (grouped into Y themes)

**Theme 1: [description] (N comments, severity)**
- file.c:123 — [specific instance]
- file.c:456 — [specific instance]
- README.md:78 — [specific instance]

**Theme 2: [description] (N comments, severity)**
- ...

### Resolved conversations: Z

If no pending conversations, skip to step 6.

4. Build a verification plan (CRITICAL — do not skip)

This is the step that prevents 10+ round feedback loops. Before touching any code, build a comprehensive plan for each theme. The goal: fix every instance in one pass so CodeRabbit sees zero duplicates on re-review.

4a. For each theme, run an impact analysis

For each theme, before writing any fix, spawn an analysis agent (subagent_type: "Explore") to answer:

1. Where else does this pattern appear? Search all files changed in this PR for the same class of issue — not just the line CodeRabbit flagged. Use grep/glob to find every instance.

   Examples of what "same pattern" means:

   • CodeRabbit says "docs say circles" → grep for "circle" in all .md, .h, and .c files touched by this PR
   • CodeRabbit says "unchecked fclose" → grep for every fclose call in the file and check if any are unchecked
   • CodeRabbit says "test only checks return value" → check ALL similar tests for the same weakness, not just the one cited
   • CodeRabbit says "division by zero if width is 0" → find ALL divisions by width, height, or any user-derived value in the function
   • CodeRabbit says "fabsf() should be SDL_fabsf()" → grep for ALL bare C stdlib calls (fabsf, sinf, cosf, sqrtf, memset, memcpy, strcmp, strlen, malloc, free) across every .c and .h file in the PR — this is a critical cross-platform portability issue

2. What documentation describes this code? Identify every README, API doc, header comment, and inline comment that references the behavior being changed. These MUST be updated when the code changes.

3. What tests cover this code? Identify existing tests. If the fix changes behavior, those tests must be updated. If no test exists, one must be written.

4. Why did we get this wrong? Understanding the root cause prevents adjacent bugs:

   • Copy-paste without adaptation → check all copies
   • Misunderstanding an API → check all uses of that API
   • Missing a case in a switch/if chain → check the full chain
   • Generated code from a template → check all generated instances

4b. Write the verification plan

For each theme, produce a checklist:

Theme: "Radio button shape — docs say circles, code draws rectangles"
Root cause: Copy-pasted checkbox description without adapting for radio buttons

Code fixes:
  [ ] common/ui/forge_ui.h:234 — change "circle" to "rectangle" in header doc
  [ ] common/ui/README.md:89 — update radio button description
  [ ] lessons/ui/15-editable-controls/README.md:156 — fix shape description
  [ ] OR: actually make radio buttons round in forge_ui_ctx_radio()

Doc sync:
  [ ] Verify all 3 doc locations match after fix
  [ ] Check if any diagram shows radio buttons (update if so)

Tests:
  [ ] Add test asserting radio button geometry is rectangular (or round)
  [ ] Verify existing radio button tests still pass

Grep verification:
  [ ] grep -r "circle" across all files in this PR — zero false matches remain

Present the full plan to the user before executing. The user may spot things the analysis missed or may prefer a different approach (e.g. "just make it round" vs "fix all the docs").

4c. Execute the plan with a team

For each theme that the user approves, spawn agents in parallel to handle the three concerns simultaneously:

1. Code agent (subagent_type: "coder", run_in_background: true) — Fix the code issue across ALL instances found in the impact analysis. Not just the line CodeRabbit pointed at — every instance of the pattern.

2. Test agent (subagent_type: "tester", run_in_background: true) — Write or update tests for the fix. For library changes in common/, this is mandatory. The test should:

   • Verify the fix works (positive case)
   • Verify edge cases (zero values, NULL, overflow)
   • Verify the old bug does not regress

3. Doc agent (subagent_type: "coder", run_in_background: true) — Update ALL documentation that describes the changed behavior:

   • README sections ("What you'll learn", "Key concepts", code examples)
   • API reference docs in common/*/README.md
   • Header comments and inline comments in .h files
   • Diagram descriptions if applicable

All three agents work from the same verification plan. Give each agent the full plan so they understand the scope, but assign them their specific section (code, tests, docs).

After all agents complete, review their changes together. Check:

• Do the doc changes match the code changes?
• Do the tests actually test the fix?
• Did any agent's changes conflict with another's?

4d. Cross-verify before committing

After all theme fixes are applied, run a final verification:

1. Grep check: For each theme, re-run the grep queries from the impact analysis. Every instance should be resolved. If any remain, fix them before proceeding.

2. Build check: Build the project to catch compile errors.

   cmake --build build 2>&1 | tail -20
   

3. Test check: Run relevant tests to catch regressions.

   ctest --test-dir build -R <relevant-test>    # C tests
   pytest tests/pipeline/ -v                     # Python tests (if applicable)
   

4. Doc consistency check: For each file pair (code + doc), verify the doc accurately describes the current code behavior. Read both files and confirm they agree.

5. Markdown lint:

   npx markdownlint-cli2 "**/*.md"
   

6. Python lint (if Python files changed):

   ruff check && ruff format --check
   

If any check fails, fix the issue before proceeding. Do NOT push code that fails tests or lint.

4e. Run local review before committing (MANDATORY — DO NOT SKIP)

STOP. You MUST run /dev-local-review here. This is not optional. If you skip this step, the push will contain avoidable issues that waste a GitHub review round. No exceptions.

After all cross-verification checks pass, run /dev-local-review to catch any issues the theme-based fixes may have introduced. This is a mandatory gate — do NOT proceed to step 5 (commit and push) until the local review completes with zero findings.

This prevents burning a GitHub review round on issues that could have been caught locally for free. The local review loop follows its own strict rules (see /dev-local-review skill) — it must run until zero findings or until the user explicitly agrees to stop.

5. Commit, push, and request re-review

After ALL themes are resolved and ALL checks pass:

1. Stage all changed files by explicit name (never git add -A):

   git add <file1> <file2> ...
   

2. Create a single commit summarizing all fixes:

   git commit -m "$(cat <<'EOF'
   Address PR feedback: [summary]
   
   Themes addressed:
   - [Theme 1]: [what was fixed, how many instances]
   - [Theme 2]: [what was fixed, how many instances]
   
   Verification:
   - All instances found via grep — none remaining
   - Tests added/updated for each fix
   - Documentation synced with code changes
   - Build and tests pass
   
   Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
   EOF
   )"
   

3. Push once to the PR branch:

   git push
   

4. Request CodeRabbit re-review:

   gh pr comment <pr-number> --body "@coderabbitai review"
   

   Do NOT use @coderabbitai resume — we want auto-pause to stay active.

5. Exit with: "Changes pushed. Run this skill again after checks complete."

On next run after CodeRabbit re-reviews:

• Check if comments are auto-resolved (CodeRabbit detects implemented fixes)
• If duplicates appear AGAIN, the fix was still incomplete. Go back to step 4a and do a deeper analysis — do not just patch the specific line again.
• If still unresolved after implementing the suggested fix:
  • Mention CodeRabbit in a comment: "@coderabbitai I've implemented your suggestion in commit ABC123. Can you verify and resolve this conversation?"
  • OR if CodeRabbit's check is complete (not currently running): "@coderabbitai Why hasn't this been auto-resolved? Is there an issue with my implementation?"
  • If CodeRabbit is still checking: wait for it to finish before asking

5.5. Escalation: 3+ review rounds on the same theme

If the same theme has appeared as a duplicate comment in 3 or more review rounds, the point-fix approach has failed. Escalate:

1. Stop fixing symptoms. Read the entire function or module that keeps getting flagged. Understand the design, not just the flagged line.

2. Identify the root cause. Common patterns:

   • A wrapper function and its inner function disagree on semantics
   • A layout/drawing split where both sides compute geometry independently
   • Copy-pasted code that was adapted incompletely
   • A variable declared with the wrong type, causing casts everywhere

3. Propose a design fix to the user. This may be a small refactor — extract a shared constant, change a type at its declaration, unify two code paths. Present it as: "This has been flagged N times because [root cause]. I recommend [design fix] instead of patching individual lines."

4. Implement the design fix with full test coverage and doc updates. This replaces the point-fix cycle with a single structural change.

6. Check if ready to merge

Check two things: the PR-level review decision AND individual reviews.

# PR-level decision (may lag behind individual reviews)
gh pr view <pr-number> --json reviewDecision,statusCheckRollup

# Individual reviews — check the LATEST review from each reviewer
gh api --paginate repos/{owner}/{repo}/pulls/{pr-number}/reviews \
  | jq -s 'add | [.[] | {user: .user.login, state: .state, date: .submitted_at}]
        | group_by(.user)
        | map(sort_by(.date) | last)'

Important: reviewDecision can show CHANGES_REQUESTED even after a reviewer has submitted a newer APPROVED review. This happens because GitHub tracks the earliest unresolved review, not the latest. Always check the latest review from each reviewer — if their most recent review is APPROVED, the PR is approved regardless of what reviewDecision says.

Verify:

• All status checks pass
• The latest review from each reviewer is APPROVED
• No unresolved conversations that need action

If the latest review is CHANGES_REQUESTED:

• There are blocking issues. Go back to step 2.

If the latest review is COMMENTED:

• COMMENTED is not an approval — do NOT proceed to merge.
• The reviewer left feedback (nitpicks, suggestions, or observations) that needs to be acknowledged before merging. Follow step 6a below.

If the latest review is APPROVED:

• All feedback is resolved. Proceed to step 7.

6a. Handle COMMENTED review state

When the latest review from a reviewer is COMMENTED, it means they left non-blocking feedback but did not approve. This typically happens when:

• All major issues are resolved but nitpicks remain
• CodeRabbit found minor suggestions after a re-review
• A reviewer left observations without requesting changes

Workflow:

1. Read every comment in the COMMENTED review (inline threads, review body nitpicks, and duplicate comments). Present each to the user with context.

2. For each comment, ask the user: implement the suggestion, or skip it?

3. If the user wants to implement: Fix the issue, commit, push, and request re-review (@coderabbitai review). Exit and wait for the next review round.

4. If the user wants to skip: Request that CodeRabbit resolve the conversation:

   gh pr comment <pr-number> --body "@coderabbitai resolve"
   

   This tells CodeRabbit the feedback was acknowledged and intentionally skipped. CodeRabbit responds with an APPROVED review.

5. CRITICAL: No commits after resolve. Any new commit on the branch — your own code, a merge from main, a rebase — dismisses the approval that @coderabbitai resolve grants. This forces you to start the cycle over.

   Also do NOT request another review (@coderabbitai review) — that triggers a new feedback round.

6. Update the branch before requesting resolve. Any commit after resolve — including a merge from main — dismisses the approval. So update the branch first, then resolve, then merge the PR.

   Practical workflow:

   a. Push all local changes first (if any). Wait for checks to pass.

   b. Check if the branch is behind main:

   gh pr view <pr-number> --json mergeStateStatus \
     --jq '.mergeStateStatus'
   

   c. If BEHIND, merge main and push before requesting resolve:

   git fetch origin main
   git merge origin/main --no-edit
   git push
   

   Wait for checks to pass on the updated branch. CodeRabbit's auto-pause prevents the main merge from triggering new review feedback.

   d. Once the branch is up to date and checks pass, request resolve:

   gh pr comment <pr-number> --body "@coderabbitai resolve"
   

   e. After CodeRabbit posts the APPROVED review, merge immediately:

   gh pr merge <pr-number> --squash --delete-branch
   

   No --admin needed because the branch is already up to date and no commits were made after the resolve.

   Fallback — if you already resolved before updating the branch:

   If the approval was dismissed by a main merge, use --admin to merge:

   gh pr merge <pr-number> --squash --delete-branch --admin
   

   Only use --admin when: the nitpick was intentionally skipped with user confirmation, all checks are green, and the only reason for the missing approval is the main merge dismissing it. Present this to the user for confirmation.

6b. Pre-merge branch update and resolve

When all feedback is addressed (either via fixes or the "no actionable comments" signal from step 1.7a), prepare the branch for merge:

1. Check if the branch is behind main:

   gh pr view <pr-number> --json mergeStateStatus \
     --jq '.mergeStateStatus'
   

2. If BEHIND, update the branch first:

   git fetch origin main
   git merge origin/main --no-edit
   git push
   

   Wait for all checks to pass on the updated branch. CodeRabbit's auto-pause prevents the main merge from triggering new review feedback.

3. Once the branch is up to date and checks pass, request resolve:

   gh pr comment <pr-number> --body "@coderabbitai resolve"
   

4. After CodeRabbit posts the APPROVED review, proceed to step 7.

Why this order matters: Any commit after @coderabbitai resolve (including a merge from main) dismisses the approval. By merging main before resolve, the approval sticks and you can merge the PR normally without --admin.

7. Merge the PR

Ask user for confirmation:

✅ All feedback resolved, all checks passing, PR approved.

Ready to merge PR #X: "Title"
Branch: lesson-03-uniforms-and-motion → main

Merge method:
1. Squash and merge (recommended for lessons)
2. Merge commit
3. Rebase and merge

After user confirmation:

gh pr merge <pr-number> --squash --delete-branch

Show success message with merged commit SHA.

GitHub CLI commands reference

# Check PR status (exit code 8 if any checks are pending)
gh pr checks <pr-number>

# Get inline review comments (the "tasks" shown in GitHub UI)
# IMPORTANT: This is the correct way to fetch review feedback
gh api repos/{owner}/{repo}/pulls/{pr-number}/comments

# View PR details (does NOT include inline review comments)
gh pr view <pr-number> --json reviewDecision,statusCheckRollup

# View specific workflow run
gh run view <run-id>

# Reply to a specific review comment thread (CORRECT - creates nested reply)
# Use this to respond to CodeRabbit/Claude feedback on specific lines
# IMPORTANT: Must include PR number in path - repos/{owner}/{repo}/pulls/{pr-number}/comments/{comment-id}/replies
# Example: gh api repos/RosyGameStudio/forge-gpu/pulls/1/comments/2807683534/replies
gh api repos/{owner}/{repo}/pulls/{pr-number}/comments/{comment-id}/replies \
  -f body="response text"

# Post a general PR comment (AVOID - doesn't thread properly with review feedback)
# This creates a standalone comment, not a reply to a review thread
gh pr comment <pr-number> --body "response text"

# Ask CodeRabbit to verify/resolve after implementing a fix (reply to the comment thread)
# IMPORTANT: Must include PR number in the path (not just comment ID)
gh api repos/{owner}/{repo}/pulls/{pr-number}/comments/{comment-id}/replies \
  -f body="@coderabbitai I've implemented your suggestion in commit ABC123. Can you verify and resolve this conversation?"

# Resolve a review thread manually
# Note: GitHub CLI doesn't directly support this - may need gh api or manual UI resolution
# Usually let CodeRabbit auto-resolve when it re-reviews

# Merge PR
gh pr merge <pr-number> --squash --delete-branch

The anti-pattern this skill prevents

Fix-the-line-not-the-pattern is the #1 cause of 10+ round review cycles. It looks like this:

1. CodeRabbit says "docs say circles but code draws rectangles" on file A
2. Agent fixes file A. Pushes.
3. CodeRabbit says "docs say circles" on file B (same issue, different file)
4. Agent fixes file B. Pushes.
5. CodeRabbit says "docs say circles" on file C. Duplicate comment.
6. Repeat for 14 rounds.

The correct response to step 1 is: grep for "circle" across every file in the PR, fix all of them, update all docs, add a test, and push once.

Every piece of feedback is a signal about a class of issue, not just a single line. Treat it that way.

Implementation notes

• Never sleep or wait — if actions are running, exit and tell user to re-run
• Can be run multiple times — each run picks up where it left off
• Commit messages for feedback changes should reference the PR number
• If Code Rabbit or Claude give conflicting advice, prioritize project conventions from CLAUDE.md
• Always show the user what will be changed before making code modifications
• Use the same commit message format as publish-lesson (with Co-Authored-By line)
• The "X out of Y pending tasks" shown in GitHub UI are the unresolved review comment threads—fetch them via gh api repos/{owner}/{repo}/pulls/{pr-number}/comments
• CodeRabbit auto-pause is intentional: CodeRabbit is configured to auto-pause after reviewing each commit (reviews.auto_review.auto_pause_after_reviewed_commits: 1). The "Reviews paused" banner is normal — it means CodeRabbit already completed its review and is waiting. Do NOT post @coderabbitai resume — we want it paused to prevent review thrashing. After pushing fixes, request a single fresh review with @coderabbitai review. If CodeRabbit shows "Currently processing new changes", wait for it to finish before fetching comments.
• CodeRabbit auto-resolution: CodeRabbit automatically resolves conversations when it detects the suggested fix was implemented in a new commit. Let it auto-resolve; only ask it to resolve manually if it doesn't detect your fix after re-review.
• CodeRabbit duplicate comments: When CodeRabbit re-reviews after a fix, it may move previously unresolved issues into a "Duplicate comments" section in the review body instead of creating new inline threads. These are NOT resolved — they indicate the fix was incomplete. Always check the latest review body for ♻️ Duplicate comments and 🧹 Nitpick comments sections and treat them as active feedback requiring action.
• Reply to comment threads: Always use gh api repos/{owner}/{repo}/pulls/{pr-number}/comments/{comment-id}/replies to reply to specific review comments. This keeps conversations threaded. Do NOT use gh pr comment for replies—it creates unthreaded general comments.
  • Common mistake: Forgetting to include the PR number in the path (e.g., repos/{owner}/{repo}/pulls/comments/{id}/replies won't work—must be pulls/{pr-number}/comments/{id}/replies)
• Team execution is mandatory for non-trivial feedback. If a theme touches both code and documentation, or if a theme involves library code in common/, you MUST spawn separate agents for code, tests, and docs. A single agent making all changes will miss the cross-cutting concerns that cause duplicate comments.
• Impact analysis before code changes. Never start editing code until the grep/search phase is complete. The 30 seconds spent searching saves hours of review round-trips.
• Tests catch what CodeRabbit cannot. CodeRabbit does static analysis. A test that exercises the actual code path will find runtime bugs (NULL derefs, off-by-one, division by zero) that no review tool can see. For library changes in common/, tests are non-negotiable.
• "No actionable comments" is CodeRabbit's clean signal. When CodeRabbit re-reviews and finds nothing, it posts an issue comment (not a review) containing "No actionable comments were generated in the recent review." The review state stays COMMENTED or CHANGES_REQUESTED from the previous round — do not wait for an APPROVED review that will never come. Detect this comment (step 1.7a) and proceed to step 6b.
• Merge main before resolve, not after. Any commit after @coderabbitai resolve dismisses the approval. Always update the branch with main before requesting resolve (step 6b). This avoids the catch-22 of needing --admin to merge.

Error handling

• If gh CLI is not installed or not authenticated, provide clear setup instructions
• If PR doesn't exist, report error
• If branch has conflicts, report and exit (user must resolve manually)
• If merge is blocked by branch protection rules, report the blocking rules

Example interaction

Running dev-review-pr for PR #224...

✓ All GitHub Actions checks passed
✓ Found 6 pending conversations

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Grouped into 3 themes:

Theme 1: Doc/code mismatch — radio button shape (3 comments, Minor)
  - common/ui/forge_ui.h:234 — header says "circle"
  - common/ui/README.md:89 — API docs say "circular"
  - lessons/ui/15-editable-controls/README.md:156 — lesson says "round"

Theme 2: Unchecked division by zero (2 comments, Major)
  - forge_ui.h:1200 — sv_rect.w can be 0
  - forge_ui.h:1240 — hue_rect.w can be 0

Theme 3: Weak test assertions (1 comment, Minor)
  - test_ui_controls.c:89 — only checks return, not side effects

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Impact analysis for Theme 1:

  grep found 5 additional instances of "circle/circular/round"
  in radio button context across the PR files:
    - forge_ui.h line 198 (function doc)
    - forge_ui.h line 892 (inline comment)
    - README.md line 34 (widget table)
    - README.md line 201 (API section)
    - SKILL.md line 45 (skill description)

  Root cause: checkbox description was copy-pasted for radio
  buttons without adapting the shape description.

  Verification plan:
    Code: fix all 8 instances (3 flagged + 5 found)
    Docs: verify all docs agree after fix
    Tests: add assertion for radio button draw data
    Grep: re-run "circle" search — expect 0 matches

[User approves plan]
[Spawns code + test + doc agents in parallel]
[All agents complete]
[Cross-verification: grep confirms 0 remaining, tests pass]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
All 3 themes resolved. Building... ✓ Tests... ✓ Lint... ✓

✓ Committed: "Address PR feedback: fix radio shape docs (8 instances),
  add zero-guards to color picker (3 divisions), strengthen test assertions
  (6 widget tests)"
✓ Pushed (single push)
✓ Requested CodeRabbit re-review

Run this skill again after checks complete.