| name | fix-review |
| description | Analyze and apply unresolved PR review comment fixes. Use when user provides a GitHub PR URL with review comments to address, or asks to fix PR feedback. |
Fix PR Review Comments
Process unresolved review comments from $1, categorize validity, apply approved fixes.
Arguments
PR-URL (required) — GitHub PR URL, e.g. https://github.com/owner/repo/pull/123--auto — non-interactive: auto-apply valid fixes, skip questionable, auto-commit
Phase 1: Fetch & Analyze
In interactive mode, start with EnterPlanMode and remain in plan mode through Phase 2.
1. Extract PR Info
OWNER=$(sed 's|.*github.com/\([^/]*\)/.*|\1|' <<< "$1")
REPO=$(sed 's|.*github.com/[^/]*/\([^/]*\)/.*|\1|' <<< "$1")
PR=$(sed 's|.*/pull/\([0-9]*\).*|\1|' <<< "$1")
2. Fetch Unresolved Review Threads
Read references/templates.md for the GraphQL query and response filtering.
3. Categorize Each Comment
Read references/examples.md in full before categorizing — contains criteria definitions, validation gates, and few-shot examples.
For each comment, reason explicitly:
1. What does reviewer request? → [extract specific request]
2. What code does it reference? → [verify in diff via Read]
3. Is suggestion technically correct? → [validate against language/framework]
4. Does similar code exist? → [Grep codebase]
5. Would change break functionality? → [trace execution path]
6. Is defensive code already present? → [check guards/validation]
Conclusion: Valid/Invalid/Questionable because [specific reason]
Language context: Go → Effective Go, error handling, concurrency; Python → PEP 8, type hints; Rust → API Guidelines, ownership; TypeScript → type safety; Bash → Google Shell Style Guide.
Phase 2: Present Findings
Read the presentation format in references/templates.md and use it to group findings by category.
Phase 3: Get User Decisions (interactive only)
Skip this phase when --auto is passed.
Call ExitPlanMode, then for each questionable comment use AskUserQuestion with these option types:
- Style choices: a) Apply reviewer's style b) Keep current style c) Standardize across file
- Ambiguous requests: a) Apply literal interpretation b) Apply inferred intent c) Skip and reply to reviewer
- Trade-off decisions: a) Apply change (accept trade-off) b) Skip (keep current behavior)
Add to valid or skip based on response.
Phase 4: Apply Fixes
For each approved fix, verify against validation gates in references/examples.md, then:
- Use Edit to apply the change
- Continue to next fix — do NOT stage or commit yet
After all edits:
git status
git diff
npm test 2>/dev/null || pytest 2>/dev/null || go test ./... 2>/dev/null || cargo test 2>/dev/null || true
Auto mode only — commit after valid fixes applied:
git add .
git commit -s -S -m "fix: address PR review comments"
Phase 5: Summary
Summary: Fixed N/M unresolved review comments
Applied: ✓ N valid fixes across K files
Skipped: ? X questionable ✗ Y invalid
Changes: [staged and committed | NOT staged — manual review required]
Next steps (interactive):
1. git diff — review changes
2. git add -p — stage selectively
3. git commit -s -S — commit
4. Reply to invalid comments on GitHub
Rules
- Search codebase for patterns before applying any fix (Grep)
- Follow linter rules from
~/.claude/CLAUDE.md — fix root cause, never use skip/disable directives
- Never auto-commit in interactive mode
- Never mark review threads as resolved
- Never assume on ambiguous comments — ask (interactive) or skip (auto)
Input: /fix-review https://github.com/owner/repo/pull/42
Behavior: EnterPlanMode → fetch threads → categorize (read examples.md) →
present findings → ExitPlanMode → ask about questionable → apply valid
fixes with Edit → run tests → print summary (no auto-commit)
Input: /fix-review https://github.com/owner/repo/pull/42 --auto
Behavior: fetch threads → categorize → apply valid fixes immediately →
skip questionable → git commit -s -S → print summary (no user interaction)
Input: PR comment "Add type hints" on def handle_request(req: Request) -> Response:
Phase 1 output: ✗ [N/A] api.py:89 — INVALID, type hints already present
Phase 4: no Edit applied; reply drafted for reviewer
