ai-review-responder

name	ai-review-responder
description	Automate responding to AI code review comments (Gemini, CodeRabbit, Cubic, Sentry) on PRs. Use when asked to "check comments", "check new suggestions", or "respond to AI reviews".

AI Review Responder Skill

Purpose

Automate the full cycle of responding to AI code review comments on GitHub PRs: Fetch → Evaluate → Fix/Decline → Resolve → Re-review

When to Use

Trigger phrases:

• "check comments", "check new suggestions", "check new comments"
• "respond to AI reviews", "handle review feedback"
• "check PR N for suggestions", "review PR comments"

Workflow

Step 1: Identify the PR

• If the user specifies a PR number, use it
• Otherwise, detect from the current branch: gh pr view --json number --jq .number
• Get repo owner/name: gh repo view --json owner,name --jq '"\(.owner.login)/\(.name)"'

Step 2: Fetch Unresolved Threads

IMPORTANT: Always set export GH_PAGER=cat before any gh api call to prevent pager issues.

export GH_PAGER=cat
gh api graphql -f query='
{
  repository(owner: "OWNER", name: "REPO") {
    pullRequest(number: PR_NUMBER) {
      reviewThreads(first: 50) {
        nodes {
          id
          isResolved
          comments(first: 1) {
            nodes {
              createdAt
              author { login }
              path
              line
              body
            }
          }
        }
      }
    }
  }
}' --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false) | {id: .id, author: .comments.nodes[0].author.login, path: .comments.nodes[0].path, line: .comments.nodes[0].line, created: .comments.nodes[0].createdAt, body: .comments.nodes[0].body[0:300]}'

If no unresolved threads, report "No new comments" and stop.

Step 3: Evaluate Each Suggestion

For each unresolved thread, use the code-review-evaluation skill's decision tree:

Category	Default	Action
Bug/error fix	AGREE ✅	Fix it
Missing docs/fields	AGREE ✅	Fix it
Unused code/imports	AGREE ✅	Fix it
Extract constant / reduce duplication	AGREE ✅	Fix it
Edge case handling (null check, mkdir, try/catch)	AGREE ✅	Fix it
Performance improvement (batch API calls)	AGREE ✅	Fix it
Pattern/architecture change	CHECK SKILLS	Cross-reference project skills
Cosmetic refactor (rename, reorder working code)	DECLINE ❌	Resolve with rationale
Repeat of previously declined suggestion	DECLINE ❌	Resolve, reference prior decision
Breaks functionality or project conventions	DECLINE ❌	Resolve with explanation

For pattern/architecture suggestions: Cross-reference the relevant project skill from .github/skills/ before deciding. See the code-review-evaluation skill for the full decision tree.

Step 4: Present Summary Table

Before making changes, present a summary table:

| # | Source | Line | Issue | Verdict |
|---|--------|------|-------|---------|
| 1 | Gemini L594 | Batch API calls | ✅ AGREE — reduces N calls to 1 |
| 2 | Cubic L88 | Don't exit on error | ❌ DECLINE — rationale here |

Step 5: Apply Fixes

For agreed suggestions:

1. Read the relevant code sections (parallel reads for efficiency)
2. Make the minimal fix
3. Validate (syntax check, lint, typecheck as appropriate for the language)

Batch all fixes into a single commit:

git add <files>
git commit -m "fix: address AI review feedback (round N)

- [description of fix 1] (Source)
- [description of fix 2] (Source)

Co-authored-by: Copilot <[e-mail]>"

If push is rejected (non-fast-forward), use git pull --rebase && git push.

Step 6: Resolve All Threads

Resolve ALL threads (both fixed and declined) using GraphQL mutation:

for TID in "THREAD_ID_1" "THREAD_ID_2"; do
  gh api graphql \
    -F tid="$TID" \
    -f query='mutation($tid: ID!) {
      resolveReviewThread(input: {threadId: $tid}) {
        thread { isResolved }
      }
    }' --jq '.data.resolveReviewThread.thread.isResolved'
done

Step 7: Request Re-review

Post a summary comment and trigger review bots:

cat <<'EOF' > /tmp/review-comment.md
Round N: [summary of fixes]. Declined: [summary with rationale].

@coderabbitai review
@gemini review
@cubic-dev-ai review
EOF
gh pr comment PR_NUMBER --body-file /tmp/review-comment.md

Bot trigger commands:

• CodeRabbit: @coderabbitai review
• Gemini: @gemini review
• Cubic: @cubic-dev-ai review

Step 8: Report Results

Present a final summary table showing all actions taken.

Evaluation Rules

Always Agree (Most Cases)

AI reviewers are usually right about:

• Missing null checks, error handling, edge cases
• Unused imports/variables
• Documentation inconsistencies
• Type mismatches
• Extracting repeated values as constants
• Reducing API call count (batching)
• File encoding, trailing newlines
• Security vulnerabilities (generic patterns)

Check Project Skills Before Deciding

Only when suggestion involves project-specific patterns. See code-review-evaluation skill for full list:

• Import patterns (next/link vs @i18n/routing)
• Type organization (/types directory governance)
• Styling (semantic tokens vs arbitrary Tailwind)
• API patterns (three-layer proxy, fetchWithHmac)
• URL/routing (searchParams restrictions)
• Environment variables (4-location rule)

Common Decline Patterns

Pattern	Typical Rationale
Cosmetic refactor of working code	Adds risk without functional benefit
Consolidate queries with different filters	Different queries need different parameters
Replace string ops with stdlib for controlled inputs	Input is a constant we control
Split small scripts into modules	Acceptable for scripts under ~1500 lines
Don't exit on missing required config in CI	CI scripts can't produce useful output without config
Broader exception handling	Specific catches are intentional
Repeated suggestion (same as prior round)	Already evaluated and declined with rationale

Efficiency Tips

• Parallel reads: Read all flagged code sections in one tool call
• Batch edits: Apply all fixes before committing
• Single commit: Group all fixes into one commit per round
• Single resolve loop: Resolve all thread IDs in one bash call
• Count rounds: Track which round this is (check PR comment history)
• Handle push conflicts: Always git pull --rebase if push fails

Example Full Run

User: "check new comments"

Agent:
1. Detects PR from current branch
2. Fetches 4 unresolved threads
3. Presents table:
   | # | Source | Issue | Verdict |
   |---|--------|-------|---------|
   | 1 | Cubic L200 | Add mkdir before write | ✅ Fix |
   | 2 | Cubic L50 | Don't exit on error | ❌ Decline — can't run without config |
   | 3 | Gemini L100 | Consolidate queries | ❌ Decline — different filters |
   | 4 | Gemini L300 | Use urlparse | ❌ Decline — input is constant |

4. Fixes #1, validates syntax
5. Commits and pushes
6. Resolves all 4 threads
7. Posts summary comment + triggers re-review bots
8. Reports: "Round N: 1 fixed, 3 declined. All resolved."

Checklist

• Set export GH_PAGER=cat before GraphQL calls
• Fetch ALL unresolved threads (not just new ones)
• Cross-reference code-review-evaluation skill for pattern/architecture suggestions
• Present summary table before making changes
• Validate changes after fixes (syntax, lint, typecheck)
• Handle git push rejection with --rebase
• Resolve ALL threads (fixed AND declined)
• Include round number in commit and comment
• Trigger review bots using correct commands (@gemini review, not /gemini review)