---

name: review-optimization description: Reviews memory optimization patches using a 5-persona peer review system. Requires unanimous approval backed by benchmarks. Bugs discovered during review are valuable - invoke /validate-correctness to validate them.

Memory Optimization Patch Review

A rigorous 5-persona peer review system for memory optimization patches. Requires unanimous approval backed by concrete benchmark data. Duplicate Hunter persona prevents redundant work.

CRITICAL: Recording is MANDATORY

EVERY review outcome MUST be recorded in reviews.yaml.

Valuable Outcomes

Finding bugs during optimization review is SUCCESS, not failure.

---

Phase 0: Pre-flight

Run /preflight first. Then check for duplicate work:

Copycat ~/.claude/skills/review-optimization/reviews.yaml
cat ~/.claude/skills/hunt-optimization/hunts.yaml

Check for:

• Same branch name already reviewed → REJECT as "DUPLICATE"
• Same file + technique already approved → REJECT as "DUPLICATE"

---

Phase 1: Measurement

Copy# Baseline
git checkout origin/main
go test -bench=<Name> -benchmem -count=10 ./pkg/... > /tmp/old.txt

# Optimized
git checkout -
go test -bench=<Name> -benchmem -count=10 ./pkg/... > /tmp/new.txt

# Compare
benchstat /tmp/old.txt /tmp/new.txt

Statistical Requirements

• Minimum 10 runs (-count=10)
• P-value < 0.05 for significance
• B/op improvement >= 5%

NO EXCEPTIONS

If benchstat output is missing → REJECT. Period.

• "Test dependencies don't work" → REJECT. Fix deps first.
• "Theoretically better" → REJECT. Prove it with numbers.
• "Obviously an improvement" → REJECT. Obvious is not measured.
• "Will benchmark later" → REJECT. Benchmark now.

The only path to APPROVE for optimizations is benchstat showing p<0.05 and >=5% improvement.

---

Phase 2: Five-Persona Review

Duplicate Hunter (Checks for Redundant Work)

• Branch not already in reviews.yaml
• File + technique combo not already approved
• No substantially similar optimization exists
• If duplicate found → REJECT with "DUPLICATE: see "

Skeptic (Demands Proof)

• Hot path verified (>1000 calls/sec or top 10% profile)
• Benchmark improvement >5% with p<0.05
• Statistical significance confirmed

Conservative (Guards Correctness)

• All tests pass
• No semantic changes
• No bugs introduced (if bug found → /validate-correctness)

Go Expert (Compiler Knowledge)

• Escape analysis checked
• No new heap escapes

Greybeard (Simplicity)

• Code still readable
• Complexity justified by improvement

---

Phase 3: Decision

When Bug Is Found

If review discovers a bug instead of an optimization:

Copy/validate-correctness

The validate-correctness skill will:

1. Create a test that reproduces the bug
2. Verify the fix works
3. Record in validations.yaml

Then return here to record the finding as a SUCCESS (bug found).

---

Phase 4: MANDATORY Recording

Copy# Append to reviews.yaml
  - branch: <name>
    date: YYYY-MM-DD
    verdict: APPROVE|REJECT|BUG_FOUND
    reason: "<summary>"
    lesson: "<pattern>"
    validated_by: "/validate-correctness"  # if bug

---

Usage

Copy/review-optimization