| name | review |
| description | 4-agent code review. Correctness, security, performance, style. Auto-fixes NITs. |
Activate When
/godmode:review, "review this", "check my code"
Workflow
1. Gather Diff
git diff main...HEAD
Read every changed file. >500 lines -> split into
per-directory reviews.
2. Multi-Agent Review (4 agents, parallel)
- Correctness: logic errors, off-by-ones, null,
unhandled promise rejections, missing returns
- Security: injection (SQL/XSS/cmd), auth bypass,
secrets in code, insecure defaults
- Performance: N+1 queries, O(n^2) in hot paths,
missing pagination, memory leaks
- Style: naming, dead code, inconsistent patterns,
missing types/returns
Each agent outputs:
SEVERITY | FILE:LINE | DESCRIPTION | SUGGESTED FIX
Severities: MUST-FIX, SHOULD-FIX, NIT.
IF no findings: state {category}: No issues found.
3. Merge & Deduplicate
Combine findings. Same file:line = merge, keep highest
severity, first agent's fix. Sort: MUST-FIX first.
4. Auto-Fix NITs
SAFE: imports, formatting, whitespace, comment typos.
UNSAFE (human): logic, API, security, test changes.
Each auto-fix = separate commit:
"review-autofix: {description}".
5. Verdict
Score: 0-10 (median of all agents).
- 10: 0 MUST, 0 SHOULD -> APPROVE
- 8-9: 0 MUST, 1-3 SHOULD -> APPROVE
- 5-7: 1+ MUST -> REQUEST CHANGES
- 0-4: 2+ MUST or critical security -> REJECT
IF score < 8: author responds to every MUST + SHOULD.
Re-review only modified parts. Max 3 iterations.
Quality Targets
- Target: <30min average review turnaround
- Target: 0 critical issues missed per review
- Max PR size: <400 lines changed for effective review
- Target: 0 MUST-FIX findings remaining at merge time
- Target: >8/10 verdict score before approval
- Target: <3 re-review iterations to reach approval
- Target: <500 line diffs per review (split if larger)
- False positive rate: <10% of all findings
- Auto-fix success rate: >95% of NIT fixes pass tests
Hard Rules
- Every finding: file:line + suggested fix (code).
- MUST-FIX blocks merge. NIT = auto-fix if safe.
- Review against spec + tests, not personal style.
- Score = median of 4 agents.
- Deduplicate same file:line across agents.
TSV Logging
Append .godmode/review-log.tsv:
timestamp scope category severity file_line description status
Keep/Discard
KEEP if: finding has file:line + severity + fix.
DISCARD if: vague, lacks evidence, or duplicate.
Failure Classification
On finding DISCARD: classify and append to .godmode/review-failures.tsv.
False positive classes: stale_context, framework_idiom, test_only_code, intentional_pattern, duplicate_finding.
Cross-Model Review (Optional)
For higher-confidence reviews, dispatch agents with different models:
- Agent 1 (primary model): Correctness + Performance review
- Agent 2 (secondary model): Security + Style review
Findings caught by only one model are flagged as "cross-model finding" —
these are the blind spots that self-play misses.
Enable with: --cross-model flag or when reviewing security-critical code.
Stop Conditions
STOP when FIRST of:
- All 4 agents completed and findings merged
- Max 3 re-review iterations
- Re-review found 0 new issues
Autonomous Operation
On failure: git reset --hard HEAD~1. Never pause.
Error Recovery
| Failure | Action |
|---|
| Diff too large (>500) | Split per-directory reviews |
| Vague findings | Discard, re-run with explicit instruction |
| Auto-fix breaks tests | Revert, reclassify as SHOULD-FIX |
| Conflicting findings | Escalate to highest severity |
