| name | bench |
| description | Run and interpret pvf headless performance diagnostics. Use for generating benchmark PDF fixtures, running `cargo bench --bench perf`, writing JSON reports, comparing results across PDFs or code changes, and explaining scenario metrics or regressions. |
Bench
Use this skill to run, compare, or explain repo-local performance diagnostics.
Before changing benchmark behavior, report shape, scenario metadata, or
performance-diagnostics policy, read the relevant docs/reference.md
Performance Diagnostics section and docs/testing.md guidance on performance
diagnostics versus correctness tests.
For a straightforward benchmark rerun, inspect only the bench command, fixture,
and report context needed to answer the request.
Use benches/perf.rs as the bench entry point, benches/fixtures/ as the fixture source, and target/bench/ for generated fixtures and reports.
Agent Workflow
Infer the measurement goal from the conversation, then choose the smallest run that answers it.
Run benchmark commands sequentially, not in parallel.
For comparisons, keep variables unrelated to the comparison axis stable.
Report only the metrics that explain the user's question, and include enough context to reproduce the result: command, fixture, scenarios, warmup, iterations, output path, and comparison axis.