bmad-eval-runner

Run artifact and trigger evaluations in a fresh Docker or local isolation environment to prevent leaking host memory or ancestor Claude configuration. This preserves resulting artifacts, transcripts, and HTML reports reproducibly.

Trigger evaluation measures whether a skill description actually causes the skill to fire when expected. It validates trigger updates by testing skill activation in a controlled, isolated environment without host interference.

Benchmark a skill by running artifact and trigger evaluations in isolated mode to generate an HTML report. The process captures transcripts, applies grading, and returns the run folder plus the report file.

Yes, Docker supports isolated skill regression checks. The runner applies Docker or local isolation modes to compare skill versions reproducibly without host configuration leaking into benchmark results.

Running skill evals in isolation prevents host configuration, memory, and ancestor Claude instructions from affecting benchmarking results. This satisfies reproducible execution and accurate grading across regression checks.

Compare two skill versions by running clean-room artifact and trigger evaluations in isolated mode. This generates reproducible HTML reports containing captured transcripts, grading metrics, and preserved output files for analysis.