doctor

madbench doctor

Three checks, three commands, output printed as it came. No agent is dispatched, because this command's whole value is being a mechanical answer a human can trust. An agent summarising a gate is a gate you cannot cite.

Run all three

Run each from the project root, one command per call, and paste its complete stdout and stderr into the reply. Do not trim, do not paraphrase, do not stop at the first failure — a reader wants all three verdicts side by side.

1. Skills — do the plugin's skills match the installed madbench?

if command -v madbench >/dev/null; then madbench skills --check --plain --dir "${CLAUDE_PLUGIN_ROOT}/skills"; else echo "skipped — madbench is not on PATH, so there is no release to compare the skills with"; fi

madbench's own compare: its first line is skills <W> · madbench <I> · <verdict>, and the verdict is current, stale, pinned, disagree or unmeasured. Its exit codes are madbench's, not this command's — see the table below. Whether the skills' example benches load is madbench's own release test, so it is not repeated here.

2. Layout — does the bench tree follow the standard?

if [ -n "$(find . -maxdepth 3 \( -name madbench.yaml -o -name '*.madbench.yaml' -o -name '*.experiment.yaml' \) -not -path '*/node_modules/*' -print -quit)" ]; then bun "${CLAUDE_PLUGIN_ROOT}/scripts/check-bench-layout.ts"; else echo "skipped — this project has no bench file yet, so there is no layout to check"; fi

One bench root; every bench directory carries a Bench or Experiment file and a README.md with the required frontmatter; no loose TypeScript at a bench root; no alias key in any bench file.

The probe is the point here too. A project that has not written a bench yet has no layout to hold, and the checker would honestly report exit 2, could not measure. That is not a state the reader can act on, so the probe skips and says so. Once one bench file exists, the checker runs, and a missing bench root is then a real red: it means a bench file sits somewhere the standard does not allow.