bmad-eval-runner

Skill trigger evaluations verify whether a skill's description actually causes it to fire on target queries. They are necessary when you need to validate behavior or debug why a skill fires unexpectedly, ensuring the description accurately maps to the intended trigger conditions.

You capture and grade artifacts by executing an artifact eval against a prompt, which saves generated files to disk and grades each expectation against real outputs. The runner preserves these durable artifacts and produces an aggregate HTML report for inspection.

Yes, you can benchmark skill behavior changes using either local or Docker isolation. The runner executes checks in a fresh, clean-room environment to validate behavior and capture transcripts, keeping all run artifacts on disk for reproducible comparison regardless of the isolation mode.

The best way to generate a reproducible report for skill evaluations is running the evals in isolation mode. This process captures run artifacts, grades expectations against real outputs, and produces an aggregate HTML report for durable, shareable review of the evidence trail.

Your skill fires on the wrong queries when its description does not accurately map to target trigger conditions. You can debug why a skill fires by running trigger evals in isolated mode to verify behavior and inspect the captured transcripts for the exact match logic.