git-pr-autofix

/git-pr-autofix

Automated PR review-fix loop. Requests bot reviews, polls for responses, analyzes comments, fixes code, runs internal review, pushes, and repeats until no critical issues remain.

Usage

/git-pr-autofix <PR#>
/git-pr-autofix 1 --max-rounds 5
/git-pr-autofix 1 --reviewers codex,gemini-code-assist

Instructions

When the user invokes this skill:

1. Parse Arguments

• PR# (required): Pull request number
• --max-rounds N: Maximum review-fix iterations (default: 5)
• --reviewers: Comma-separated bot reviewer names (default: codex,gemini-code-assist)
• Detect repo from gh repo view --json nameWithOwner

2. Pre-flight Checks

Run these checks before entering the loop:

# Verify we're on the PR branch
gh pr view <PR#> --json headRefName,state
# Verify clean working tree
git status --porcelain
# Verify tests pass
npm test (or equivalent)

If any check fails, report and ask user how to proceed.

3. Enter the Review-Fix Loop

iteration = 0
while iteration < max_rounds:
    iteration += 1
    
    ## Step A: Request Reviews
    Record the current timestamp and latest commit SHA:
      REVIEW_REQUEST_TS=$(date -u +%Y-%m-%dT%H:%M:%SZ)
      LATEST_SHA=$(git rev-parse --short HEAD)
    Post a PR comment mentioning all reviewers with commit context:
      gh pr comment <PR#> --body "@reviewer1 @reviewer2 please review (round N, commit ${LATEST_SHA})"
    
    ## Step B: Poll for Reviews (background agent)
    Launch the `review-collector` agent in the BACKGROUND with:
      - repo, pr_number, reviewers list
      - after_timestamp: REVIEW_REQUEST_TS from Step A
      - stabilization_wait: 60 (seconds to wait after first detection)
      - timeout: 600 (10 minutes)
    
    The agent will:
      - Poll `gh api` every 30s using jq fromdateiso8601 for exact timestamp filtering
      - Wait for ALL reviewers to respond
      - After first detection, wait 60s for reviewers to finish posting inline comments
      - On timeout: collect whatever exists and return with timeout=true
    
    While waiting, report: "Waiting for reviewers... (polling every 30s)"
    
    If the agent times out or returns incomplete results:
      - Ask user: "N명 중 M명만 응답. 계속 대기 / 수동 수집 / 건너뛰기?"
      - If user chooses manual collection, run gh api directly and proceed
    
    ## Step C: Analyze Comments (agent)
    Automatically pipe the review-collector output to the `comment-analyzer` agent:
      Input: review-collector's full output (reviews + inline comments)
      Also provide: PR scope context (what this PR implements, relevant AC)
      Output: categorized list:
        - CRITICAL: Must fix now (bugs, security, data corruption)
        - WARN: Should fix (performance, reliability, consistency)
        - INFO: Nice to have (style, optimization suggestions)
        - DEFERRED: Out of scope (different Seed, Phase 2, design change)
      
      For each item: { severity, file, line, summary, fix_description, deferred_reason? }
    
    ## Step D: Present to User
    Show analysis as a table:
    
    | # | Severity | File | Summary | Action |
    |---|----------|------|---------|--------|
    | 1 | CRITICAL | cache.ts:82 | Tag escape vulnerability | Fix |
    | 2 | WARN | route.ts:150 | Missing inArray filter | Fix |
    | 3 | DEFERRED | echoes.ts:311 | Tree pagination redesign | Seed 6 |
    
    Then ask: "N개 수정, M개 보류 예정입니다. 진행할까요?"
    
    - If user says no or wants to adjust: let them modify the plan
    - If user approves: proceed to Step E
    
    ## Step E: Fix Code
    For each approved fix (CRITICAL + WARN):
      1. Read the file at the specified location
      2. Apply the fix based on the analyzer's fix_description
      3. Verify the fix doesn't break anything nearby
    
    After all fixes:
      - Run tests: `npm test` (or project test command)
      - Run lint: `npx eslint src/` (or project lint command)
      - If tests/lint fail, fix before proceeding
    
    ## Step F: Internal Review (before push)
    Run TWO internal reviews in parallel:
    
    1. **Codex review** (optional, requires codex plugin):
       Use `/codex:rescue` skill to review the changes
       Focus: bugs, edge cases, security gaps
       If not available, skip this step and rely on self-review only
    
    2. **Self-review checklist** (Opus 4.6 max effort):
       Check all modified files against:
       - [ ] Query performance (indexes, full scans)
       - [ ] Serverless safety (no fire-and-forget)
       - [ ] Cursor/orderBy consistency
       - [ ] Atomicity (transactions where needed)
       - [ ] Key design (composite keys, deduplication)
       - [ ] Edge cases (empty string, null, empty array)
       - [ ] UI↔API state consistency
    
    If internal review finds issues:
      - Fix them immediately
      - Do NOT push code with known issues
      - Re-run tests after fixing
    
    ## Step G: Commit + Push
    ```bash
    git add <modified files>
    git commit -m "fix: address round N review — X issues (summary)"
    git push
    ```
    
    ## Step H: Check Loop Condition
    If this round had 0 CRITICAL items → BREAK (done!)
    If iteration >= max_rounds → BREAK (warn user)
    Otherwise → continue loop (back to Step A)

4. Finalize

After loop exits:

1. Post a final PR comment summarizing ALL deferred items:

   ## Deferred Items (out of scope for this PR)
   | Item | Rationale | When |
   |------|-----------|------|
   | ... | ... | ... |
   

2. Report to user:

   PR Review Loop Complete
   =======================
   Rounds: N
   Fixed: X issues
   Deferred: Y items
   Tests: all passing
   

5. Error Handling

• Review timeout: If reviewers don't respond within 10 min:
  1. Run gh api directly to collect any partial reviews
  2. Ask user: "N명 중 M명 응답. 수집된 리뷰로 진행 / 계속 대기 / 취소?"
  3. If partial reviews exist, offer to proceed with what's available
• Test failure after fix: Report which test failed, attempt auto-fix, if can't fix ask user
• Max rounds reached: Warn user, show remaining issues, suggest manual review or scope reduction

Configuration

Create .claude/git-pr-autofix.local.md to customize:

reviewers:
  - codex
  - gemini-code-assist
max_rounds: 5
poll_interval_seconds: 30
poll_timeout_seconds: 600
test_command: "npm test"
lint_command: "npx eslint src/"