Experimental Claude Code skills for optimizing the Datadog Agent.

This repository contains custom skills designed to assist with performance analysis, profiling, and optimization work on the Datadog Agent codebase.

Environment validation checklist.

Autonomous, optimization hunter.

Systematically explores the codebase looking for memory optimization opportunities. When optimizations are found, invokes review-optimization. When bugs are discovered, invokes validate-correctness.

Reviews memory optimization patches using a 5-persona peer review system (Duplicate Hunter, Skeptic, Conservative, Go Expert, Greybeard). Requires unanimous approval backed by benchstat data showing >=5% improvement with p<0.05. Rejects patches without statistical evidence. If bugs are found during review, invokes validate-correctness.

Validates discovered bugs with reproducing tests. Called by other skills when bugs are found during optimization work. Creates unit tests that fail before the fix and pass after, plus fuzz tests when appropriate. Ensures no bug fix is merged without a test that would have caught it.

Salvages optimization work that was done without proper benchmarks. Audits changes, generates missing benchmark evidence, validates claims. Keeps only changes with proven improvements. If bugs are discovered during rescue, invokes validate-correctness.

Useful when claude loses the plot.

                    ┌─────────────┐
                    │  preflight  │
                    └──────┬──────┘
                           │
           ┌───────────────┼───────────────┐
           │               │               │
           ▼               ▼               ▼
   ┌───────────────┐ ┌───────────┐ ┌───────────────┐
   │     hunt-     │ │  review-  │ │    rescue-    │
   │ optimization  │ │optimization│ │ optimization  │
   └───────┬───────┘ └─────┬─────┘ └───────┬───────┘
           │               │               │
           │  optimization │               │
           │───────────────▶               │
           │                               │
           │         bug found             │
           └───────────────┬───────────────┘
                           │
                           ▼
                 ┌─────────────────┐
                 │    validate-    │
                 │   correctness   │
                 └─────────────────┘

All skills record outcomes to YAML databases in .claude/skills/*/. The goal is to give claude a memory. It sort of works.