This skill should be used when the user asks to "check this code", "validate this", "verify this implementation", "is this correct", "review this code", "check for errors", "multi-agent verification", or mentions production-critical code, financial calculations, security implementations, or high-stakes operations. Provides comprehensive multi-agent verification workflow with specialized critic agents.
Executes multi-agent verification with specialized critics for production-critical code validation.
/plugin marketplace add reggiechan74/cc-plugins/plugin install code-coherence@cc-pluginsThis skill inherits all available tools. When active, it can use any tool Claude has access to.
examples/financial-calculation.mdreferences/research-paper.mdreferences/swiss-cheese-model.mdExecute comprehensive multi-agent verification for production-critical code using the "Team of Rivals" architecture. This skill orchestrates specialized critic agents with opposing incentives to catch errors before they reach production, achieving 92%+ reliability versus 60% single-agent baseline.
Activate this skill for high-stakes code where errors have significant consequences:
Do not use for exploratory coding, prototypes, or low-stakes experiments.
Before any code changes, create an execution plan with pre-declared acceptance criteria:
Acceptance criteria must be:
Example plan structure:
EXECUTION PLAN
Scope: Refactor authentication to JWT tokens
Files: src/auth/login.ts, src/auth/middleware.ts
SUCCESS CRITERIA (Pre-Declared):
✓ All existing auth tests pass
✓ No OWASP Top 10 vulnerabilities introduced
✓ Token storage uses httpOnly cookies (not localStorage)
✓ Token refresh logic handles 401 responses
✓ Code coverage maintained at 85%+
CRITICS ASSIGNED:
- Code Critic: Test pass rate, coverage, logic correctness
- Security Critic: OWASP compliance, token storage, auth flow
- Domain Critic: Session management business rules
RETRY BUDGET: 6 iterations before escalation
ESTIMATED COST: +40% tokens (~$0.15 additional)
VETO AUTHORITY: ANY critic can reject (unanimous approval required)
Execute changes while maintaining strict context boundaries:
Context isolation rules:
Run specialized critics in parallel (default) or sequential (optional):
Code Critic evaluation:
EVALUATION: Code Changes
✓ Syntax: All files parse correctly
✓ Logic: JWT token generation handles expiration
✓ Performance: No N+1 queries introduced
❌ VETO: Missing error handling in token refresh
Reason: 401 response from refresh endpoint not caught
Impact: Users experience silent logout without feedback
Fix: Add try-catch around fetch(), show re-login modal
Security Critic evaluation:
EVALUATION: Security Review
✓ OWASP: No SQL injection vectors
✓ Auth: JWT signature validation correct
✓ Data: No sensitive data in logs
❌ VETO: XSS vulnerability in error messages
Reason: User input echoed in error modal without escaping
Impact: Reflected XSS if attacker controls username
Fix: Use textContent instead of innerHTML for error display
.claude/rules/Domain Critic evaluation:
EVALUATION: Domain Rules (Financial)
✓ Precision: Using Decimal type, not float
✓ Rounding: Banker's rounding applied
❌ VETO: Missing audit trail for token generation
Reason: Financial regulations require logging all auth events
Impact: Cannot prove compliance during audit
Fix: Log token generation with user_id, timestamp, IP to audit table
Process critic verdicts following hierarchical veto authority:
Retry loop example:
Iteration 1: Security Critic vetoes (XSS vulnerability)
→ Executor fixes: Use textContent instead of innerHTML
Iteration 2: Domain Critic vetoes (missing audit trail)
→ Executor fixes: Add logging to audit table
Iteration 3: All Critics approve ✓
→ Advance to user review
Retry budget: 6 iterations (default), configurable per project Escalation: Require human review, downgrade to single-agent, or fail with report
Present final results with complete decision history:
Summary report format:
✓ COHERENCE CHECK COMPLETE
Status: All Critics Approved
Iterations: 3 (budget: 6 remaining)
Time: 4.2 minutes (verification: 1.8min, execution: 2.4min)
Cost: +38.6% tokens ($0.15 additional)
CRITIC VERDICTS:
✓ Code Critic: Approved (iteration 1)
✓ Security Critic: Approved (iteration 2, after XSS fix)
✓ Domain Critic: Approved (iteration 3, after audit trail added)
FILES CHANGED:
- src/auth/login.ts (+42, -15)
- src/auth/middleware.ts (+28, -8)
ACCEPTANCE CRITERIA MET:
✓ All existing auth tests pass (18/18)
✓ No OWASP Top 10 vulnerabilities
✓ Token storage uses httpOnly cookies
✓ Token refresh handles 401 responses
✓ Code coverage maintained (87%, was 85%)
View detailed audit trail: /audit-trail show session-abc123
Load settings from .claude/code-coherence.local.md:
Define file patterns requiring automatic verification:
highStakesPatterns:
- "src/auth/**"
- "src/payment/**"
- "src/financial/**"
- "database/migrations/**"
When user requests changes to files matching these patterns, proactively suggest coherence check.
Enable/disable critics and select models:
critics:
code:
enabled: true
model: opus
vetoThreshold: strict # or critical-only
security:
enabled: true
model: opus
domain:
enabled: true
model: opus
specialization: financial # or healthcare, legal, custom
retryBudget: 6
consensusMode: unanimous # unanimous, majority, weighted
parallelExecution: true # parallel (faster) or sequential (cheaper)
autoVerify: false # false = ask user, true = auto-run on high-stakes
acceptableErrorRate: 0.079 # 7.9% residual per research paper
costVisibility:
showTokens: true # Display token counts per critic
showTime: true # Show time breakdown
estimateCost: true # Estimate before running
After creating execution plan, automatically invoke:
/plan-review
This validates the plan itself before execution, checking for completeness, clarity, and measurability.
For domain-specific criteria not covered by pre-built templates:
/acceptance-criteria define for this authentication refactoring
All decisions automatically logged. Retrieve with:
/audit-trail show session-abc123
Or search:
/audit-trail search security critic rejections
After running coherence check, optionally validate critics have orthogonal failure modes:
/swiss-cheese-validation verify independence
Investment: 38.6% computational overhead (research-validated) Return: 80% reduction in user-facing errors
When justified:
When not justified:
Three independent critics with misaligned failure modes:
Code Critic catches 87.8% of errors:
Security Critic catches what Code Critic misses:
Domain Critic catches what both miss:
Result: Errors that slip through one layer encounter another. With orthogonal failure modes, 92.1% of errors caught before user exposure.
Problem: Multiple critics veto for conflicting reasons Solution: Escalate to human review with detailed rationale from each critic
Problem: 6 iterations completed, critics still rejecting Solution: Three escalation options (user-configurable):
Problem: Token usage higher than estimated Solution:
Problem: 7.9% residual error rate (expected per research) Cause: Errors requiring external context (requirement ambiguity, subjective preferences, domain edge cases) Solution: Refine acceptance criteria with more specificity, add custom domain critic
For detailed patterns, advanced techniques, and implementation guides:
references/research-paper.md - Original "Team of Rivals" research paper summaryreferences/swiss-cheese-model.md - Error prevention through layered validationreferences/organizational-intelligence.md - How organizational principles apply to AI systemsreferences/critic-patterns.md - Common critic evaluation patterns and heuristicsreferences/cost-optimization.md - Strategies for reducing overhead while maintaining reliabilityWorking examples demonstrating coherence check in action:
examples/financial-calculation.md - Multi-agent verification for compound interest calculatorexamples/auth-refactoring.md - JWT implementation with security and domain criticsexamples/data-migration.md - Database schema change with rollback validationPre-built acceptance criteria for common scenarios:
templates/financial.yaml - Financial calculation standards (precision, rounding, audit)templates/security.yaml - OWASP Top 10 checklist, auth patternstemplates/performance.yaml - Latency SLAs, memory limits, query optimizationTriggering: Broad patterns for accessibility - users don't need to know exact command Default behavior: Interactive (ask user for approval), option for automation Output format: Summary by default, detailed report on request Critic execution: Parallel by default (faster), sequential option (cheaper) Consensus: Unanimous approval required (any veto blocks), majority/weighted as options
Integration points:
.claude/code-coherence.local.md.claude/coherence-audit/ with git commit referencesApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.
Creating algorithmic art using p5.js with seeded randomness and interactive parameter exploration. Use this when users request creating art using code, generative art, algorithmic art, flow fields, or particle systems. Create original algorithmic art rather than copying existing artists' work to avoid copyright violations.
Create beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.