# Performance Testing

English | 한국어 | 简体中文

`performance-testing` is a coding-agent skill for measuring, comparing, and explaining runtime performance with controlled experiments and reproducible logs. Its core loop is **Define → Baseline → Candidate → Confirm → Analyze → Decide → Re-measure → Record**. ## Use this skill for - Benchmarking latency, throughput, CPU, memory, allocation, or resource cost - Comparing a baseline with a candidate implementation or configuration - Diagnosing performance regressions and mixed tradeoffs - Designing microbenchmarks, load tests, stress tests, or soak tests - Profiling an observed bottleneck and testing cause hypotheses - Implementing a requested optimization and proving the before-and-after result ## Key behavior - Uses equivalent conditions and repeated measurements - Records absolute values, relative differences, variation, and confidence - Writes a project-owned performance log when measurements are run - Preserves failed attempts and rejected hypotheses - Investigates a slower candidate before considering alternatives or rollback - Treats emergency rollback as containment, not as root-cause analysis - Adds a permanent CI gate only when the benchmark is representative and stable ## Do not use it for - Correctness-only failures; use `bugfix` - New functionality without a performance objective; use `feature-dev` - Behavior-preserving cleanup; use `refactoring` - General performance definitions that can be answered directly - Production or third-party load generation without explicit authorization ## Install ```bash skill-forge install performance-testing --lang en --agent codex skill-forge install performance-testing --lang ko --agent claude skill-forge install performance-testing --lang zh --agent cursor ``` ## Included resources - `assets/performance-log-template.md`: reproducible experiment and cause-analysis log - `references/measurement-quality.md`: workload, metric, variability, threshold, and CI-gate guidance - `references/regression-analysis.md`: regression confirmation, competing hypotheses, causal confidence, and response selection - `examples/`: trigger and non-trigger boundary cases