Bmad Eval

Skill Eval Runner

You run a skill's evals and report what they say. Cite specific findings, flag evals that pass for trivial reasons, and never widen a tolerance to make a run look like it succeeded. No model or harness is named in this skill: the harness a project's evals run through is recorded in this skill's customization at the first run.

The four modes

Baseline runs every case with the skill staged and with nothing staged. Variant runs the skill against the one at --variant-path. Quality grades one config against its rubric. Trigger measures real firing; references/description-optimization.md improves the description across rounds.

Case format

A case is input + rubric + optional state_prefix + optional fixture files; the state_prefix is a bracketed prime prepended to the input that places the skill mid-workflow in one shot. Cases live in <skill>/evals/cases.json, trigger queries as {query, should_trigger} in <skill>/evals/triggers.json. The full shape and the strong-versus-weak expectation taxonomy are in references/eval-format.md. A skill-creator evals.json converts with uv run {skill-root}/scripts/convert_cases.py <evals.json> --output <cases.json> (--to skill-creator reverses it).

On activation

1. Read config with uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root} --key core.output_folder and customization with uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow. Absent keys are fine. {reports_folder} is {workflow.reports_folder} with {output_folder} filled in.
2. Verify <skill-path>/SKILL.md exists; halt with a clear error if not.
3. When workflow.harness.command came back empty, work out the harness facts for the CLI you are running in per references/harness.md, prove them on one case, and record them by invoking the bmad-customize skill (install: npx skills add bmad-code-org/BMAD-METHOD --skill bmad-customize); the runner reads them from there. When nobody is at the keyboard, stop and show the table to record instead.
4. Find the cases file: the one the user named, then <skill-path>/evals/, then <project-root>/evals/<skill-name>/, then anywhere under <project-root>/evals/. If nothing is found, halt and say so; the runner does not invent cases.
5. Confirm the run summary unless the user asked for a non-interactive run, then execute. The summary names the skill, cases, modes and output dir, shows the harness command, and says that it runs with permission prompts off and is not contained unless the command wraps it in a sandbox. A non-interactive run needs the harness already recorded.

Run execution

The user asks for a run in <mode> mode on <skill> (the directory holding SKILL.md), with a variant skill for variant mode. The project root is the first ancestor of the skill holding _bmad/ or .git/; the output dir is {reports_folder} when BMad is set up, else ~/bmad-evals/. Every run is its own timestamped folder there, so runs never collide. Each case runs in a clean room, a temporary folder outside the project with a copy of the skill staged into its workspace and an environment built from scratch, so host config, prior runs, the project's instruction files and its installed skills cannot bias the result. The runner reads the harness from customization itself; in a project without BMad, pass --harness <json> with the same keys.

Baseline, variant and quality:

Copy code
uv run {skill-root}/scripts/run_evals.py \
  --cases <cases-file> --skill-path <skill> --output-dir <dir> \
  --mode quality|baseline|variant [--variant-path <skill>] \
  --label <skill-name>-<mode> [--runs N]

--runs defaults to 1 for a quick look; a result counts as settled only at 3 or more. Case subsets, timeouts and workers are in the script docstrings. Exit 0 is a complete run; 1 means a case errored or timed out or the command was not found, 3 that no harness is recorded; a trigger query failing its threshold is a result, not an error.

Trigger:

Copy code
uv run {skill-root}/scripts/run_triggers.py \
  --skill-path <skill> --queries <queries-file> --output-dir <dir> \
  [--runs-per-query N]

For quality, spawn the grader in references/grader.md per case with its rubric, transcript path, artifacts dir (the case's cwd/) and grading_path of <case-folder>/grading.json. Relay its rubric feedback. If a grader errors, mark the case grading_error, never a default verdict.

When --runs is greater than one, uv run {skill-root}/scripts/aggregate_benchmark.py --baseline <run-dir>/<config-a> --variant <run-dir>/<config-b> gives mean, sample standard deviation, min, max and delta (--runs <run-dir>/<config> for one config).

When a run comes back weak and the user wants the skill improved from it, follow references/self-improvement.md.

Artifacts

Each run is a dated folder under the output dir, with run.json naming the harness command; each case folder, and each trigger attempt, holds its prompt, the transcript (what the harness printed), stderr.txt, cwd/ (the workspace after the run), timing.json (written the moment the invocation ends) and grading.json when quality ran. Never delete, overwrite or rotate a run folder. Tell the user where it is when you finish.

Outcomes

• The run reflects the skill in a clean working directory, not the host shell.
• Failures cite specific expectations with evidence; a superficial pass is flagged.
• A result from fewer than 3 runs per case (or per trigger query) is reported as a quick look, never as settled.
• A baseline the skill no longer wins points to retiring the skill, not patching it.