Evaluates review findings in the preceding conversation, critically assesses each comment, and fixes only justified issues. Use when review findings need to be addressed, when PR review comments need resolution, or when asked to "fix review issues", "address review findings", "レビュー対応して"
Evaluates review findings from the preceding conversation (severity-classified reviews, PR comments, free-form feedback — any format), critically assesses each one, fixes justified issues, and produces an updated summary with a new verdict.
Not for: when no review findings exist, running the review itself (that is a separate skill)
This skill evaluates, addresses, and re-summarizes review findings. It does not run reviews or merge branches.
Locate review findings in the preceding conversation. Accept any of these formats:
[SEVERITY] Title + File: + Issue: + Fix:For each finding, record the reviewer name (agent name, GitHub username, etc.) and the target file/line.
If no review output is found in the conversation, ask the user to provide it.
For each extracted issue:
Critical principle: Do not blindly accept all review feedback. Reviewers can be wrong. Evaluate each issue on its own merits against the actual code.
Process in priority order: CRITICAL → HIGH → MEDIUM → LOW.
For each issue classified as FIX:
For DOWNGRADE issues: only fix if the corrected severity is still CRITICAL or HIGH.
Output the results in this exact format:
## Address-Reviews Summary
### Addressed Issues
#### [CRITICAL] Issue Title (reviewer-name)
- **File**: path/file.ts:line
- **Problem**: brief description of what was wrong
- **Fix applied**: brief description of what was changed
#### [HIGH] Issue Title (reviewer-name)
- **File**: path/file.ts:line
- **Problem**: brief description
- **Fix applied**: brief description
### Not Addressed
#### [MEDIUM] Issue Title (reviewer-name)
- **File**: path/file.ts:line
- **Reason**: explanation of why this was skipped (false positive / severity downgraded / stylistic preference / etc.)
#### [LOW] Issue Title (reviewer-name)
- **File**: path/file.ts:line
- **Reason**: explanation
### Updated Review Summary
| Severity | Original | Fixed | Remaining |
|----------|----------|-------|-----------|
| CRITICAL | N | N | N |
| HIGH | N | N | N |
| MEDIUM | N | N | N |
| LOW | N | N | N |
Verdict: APPROVE/WARNING/BLOCK — explanation based on remaining issues
Re-evaluate the verdict based on remaining (unfixed) issues only:
If all CRITICAL and HIGH issues were fixed or validly downgraded/skipped, the verdict should change to APPROVE.
| Mistake | Fix |
|---|---|
| Blindly fixing every reported issue | Evaluate each issue independently; skip false positives with justification |
| Not reading actual code before fixing | Always read the referenced file and line before deciding |
| Changing verdict without re-counting | Re-count remaining issues by severity after all fixes |
| Fixing MEDIUM/LOW when CRITICAL exists | Prioritize CRITICAL → HIGH → MEDIUM → LOW |
| Applying reviewer's fix verbatim when a better fix exists | Use the suggestion as a starting point; improve when appropriate |
| No review output in context | Ask user to provide review results |
| Fix introduces new issue | Verify surrounding code still works after each change |
| Findings without severity left untriaged | Assign severity yourself before triaging |