Critique swarm

Critique Swarm: Parallel Adversarial Review

Launches 4 parallel Agent invocations, each with a distinct adversarial mandate, then a 5th synthesis agent that deduplicates findings and ranks them.

When to Use

• Before approving a complex implementation plan
• Before merging a feature branch with cross-cutting changes
• When you want independent multi-angle critique beyond what /multi-review provides
• After a port/transplant operation to catch fidelity loss

When NOT to Use

• Single-file trivial edits → use /validate
• Code-level quality review → use /multi-review
• Ready to ship → use /ship (which includes /session-critique)

Procedure

Step 1: Identify the review target

If $ARGUMENTS contains a file path, review that file. Otherwise, review the current session's changes via git diff HEAD.

Step 2: Launch 4 parallel critique agents

Spawn all 4 agents in a single message (parallel launch) using the Agent tool. Each agent receives the review target and its specific mandate.

Agent 1 — Security & Correctness:

Review the changes for security vulnerabilities, correctness regressions, broken invariants, and unsafe patterns. Check for: command injection, path traversal, missing input validation at system boundaries, race conditions, and any behavior change that could break existing consumers. Report findings with file:line references.

Agent 2 — Scope Creep & Over-Engineering:

Review the changes for scope creep and over-engineering. Check for: unnecessary abstractions, features not requested in the spec, extra deliverables beyond what was asked, verbose preambles, premature generalization, and any "while we're here" cleanup that wasn't part of the task. Flag anything that could be removed without affecting the stated goal.

Agent 3 — Fidelity Loss:

Review the changes for fidelity loss during porting or transplanting content. Check for: dropped sections, missing taglines or sub-groups, altered meaning during rewording, references that point to moved/renamed files, and any content that was in the source but is absent from the output. Compare source and destination line by line where applicable.

Agent 4 — Test & Validation Gaps:

Review the changes for missing tests and validation gaps. Check for: untested code paths, missing edge case coverage, assertions that don't verify the actual behavior, test files that exist but don't run, and validation steps mentioned in the plan but not implemented. Flag any change that modifies behavior without a corresponding test.

Step 3: Synthesize findings

After all 4 agents complete, launch a 5th synthesis agent:

You have received findings from 4 independent critique agents reviewing the same changes. Your job: (1) deduplicate findings that appear in multiple reports, (2) rank all unique findings by severity (HIGH/MEDIUM/LOW), (3) mark any false positives with reasoning for why they're false, (4) produce a single ordered list of recommended fixes — minimal set only. Do not implement any fixes. Output format:

=== CRITIQUE SWARM VERDICT ===
Reviewed: [target]
Agents: 4 (security, scope, fidelity, tests)

## Findings (ranked by severity)
| # | Severity | Category | File:Line | Issue | Fix |
|---|----------|----------|-----------|-------|-----|

## False Positives Dismissed
| # | Category | Reason |

## Recommended Action
[Fix N issues before proceeding / Clean — no blocking issues found]

Step 4: Present verdict to user

Display the synthesis report. Do NOT implement any fixes automatically. The user decides which fixes to apply.

Relationship to Other Skills

Skill	Reviews	Use Together?
/multi-review	Code quality (style, correctness, security, performance)	Yes — complementary
/session-critique	Session work — decision-aware (the session's in-conversation decisions + regressions, scope, dangling refs)	Either/or — use /session-critique when you need decision-adherence; critique-swarm for cold scope review
/validate	Deterministic + rule-based checks	Yes — different layer