performance-proof
Purpose
Proves or refutes a performance regression by measurement, never by reading code.
Source Skill purpose
When to use it
Use performance-proof when a change has a performance budget or a suspected regression that needs measurement. It runs the project's own benchmark on a frozen base and the change in separate worktrees, interleaving repeated samples. The report records environment, warmup, commands and times. A bundled script compares medians, noise and intervals against the configured threshold or a budget fixed by the plan. Noise yields an inconclusive result, and a missing benchmark yields unsupported. An unmeasured pattern remains a judgment, not a measured regression.
The comparison keeps units consistent and lets the statistics script compute dispersion rather than guessing from a chart. Environment load and tool versions are recorded with samples. A possible hot-path pattern without measurement stays judged.
Source Skill purpose
How it works
Source Flow
- Read the budget Inspect task, plan, diff and any stated performance limit.
Source Flow
- Find the benchmark Use the project's own benchmark command or mark measurement unsupported.
Source Flow
- Set thresholds Resolve defaults, project configuration and the plan's overriding budget.
Source Flow
- Prepare worktrees Record separate base and change worktrees before creating them.
Source Flow
- Measure repeatedly Run warmups and interleaved samples under recorded environment conditions.
Source Flow
- Compare statistics Let the comparison script calculate median, dispersion and interval.
Source Flow
- Separate judgments Cite measurements for numerical claims and keep unmeasured patterns judged.
Source Flow
- Validate evidence Run the performance report validator with configured thresholds.
Source Flow
- Clean worktrees Remove only the recorded isolation paths after verifying them.
Source Flow
- Report the result Present measurements, environment, threshold, noise and computed status.
Source Flow
Key rules
- Support every performance claim with a recorded measurement row.
Source Hard rules
- Use only a benchmark command already provided by the project.
Source Hard rules
- Measure the same operation at base and change revisions.
Source Hard rules
- Let the script compute statistics instead of judging them by eye.
Source Hard rules
- Measure without tuning product code or benchmark inputs.
Source Hard rules
Output and gate
Output
The skill definition states no output.
Source Skill definition
Gate
- No blocking measurement row yields a pass with judged items listed separately.
Source Gate rule (team mode)
- A measured regression beyond the threshold yields a blocking failure.
Source Gate rule (team mode)
- Excessive noise yields inconclusive rather than pass or fail.
Source Gate rule (team mode)
Source
Source Skill definition