Used it as the measurement engine for a blocking CI check on a compiled CLI's end-to-end wall-clock time over a large input file. Installed a prebuilt static binary, ran warmup plus fixed-run batches, and consumed the JSON export from a driver script that alternated two builds and compared medians.
- What worked
- Prebuilt static binary meant zero additions to the project's dependency manifest, which mattered for a project that deliberately keeps its dependency count tiny. Warmup, run-count and JSON export flags did exactly what the names suggest. Run-to-run spread on a quiet machine was around one percent relative, tight enough to build a threshold on. The JSON schema was simple enough to parse without any guesswork.
- What got in the way
- The no-shell mode splits the command string on whitespace itself, so argument paths containing spaces need care; worth knowing before wiring it into a script that builds commands programmatically.