Test Cleaner

• Categorize findings by severity (P1/P2/P3)
• Propose specific changes for each finding
• Estimate impact on test suite speed and coverage
• Iterate with subagents until the plan is comprehensive

Phase 3: Adversarial Review (CRITICAL QUALITY GATE)

This is THE KEY PHASE -- the most important phase in the entire workflow. NEVER skip, NEVER rush, NEVER settle for "good enough."

Expose the plan to two neutral reviewer subagents:

• Reviewer A (Opus): Independent critique of the optimization plan
• Reviewer B (Sonnet): Independent critique from a different perspective

Both reviewers challenge assumptions, identify risks, and suggest improvements.

Mandatory iteration loop: After each reviewer pass, if ANY issues, concerns, or suggestions remain from EITHER reviewer, revise the plan and re-submit to BOTH reviewers. Repeat until BOTH reviewers explicitly approve with ZERO reservations. Do not proceed to Phase 4 until unanimous, unconditional approval is reached.

This is the critical quality gate. Loop as many times as needed. The test suite must be bulletproof before execution begins.

Phase 4: Execution

Apply the agreed changes:

• Machine-readable output format: ERROR [file:line] type: description
• Include REMEDIATION suggestions and SEE references
• Use pnpm test:segment, pnpm test:random, pnpm test:critical for targeted validation

Phase 5: Verification

• Run full test suite after changes
• Compare before/after metrics (count, duration, coverage)
• Feed findings into compound-agent memory via npx ca learn

Test Scripts Integration

• pnpm test:segment <module> -- Test specific module in isolation
• pnpm test:random <pct> -- Deterministic random subset (seeded per-agent)
• pnpm test:critical -- P1/critical tests only (fast CI feedback)

Memory Integration

• Run npx ca search "test optimization" before starting
• After completion, capture findings via npx ca learn
• Feed patterns into CCT system for future sessions

Common Pitfalls

• Deleting tests without verifying coverage is maintained elsewhere
• Optimizing for speed at the cost of correctness
• Settling for partial approval or cutting the Phase 3 review loop short before BOTH reviewers approve with zero reservations
• Making changes without machine-readable output
• Not feeding results back into compound-agent memory

Quality Criteria

• All 5 phases completed (analysis, planning, review, execution, verification)
• Both adversarial reviewers approved with zero reservations after iterative refinement
• Machine-readable output format used throughout
• Full test suite passes after changes
• Coverage not degraded
• Findings captured in compound-agent memory