skill-eval

# zuvo:skill-eval — Behavioral skill evaluator Measure whether a skill's own instructions lead a competent agent to the correct behavior. For each eval case in the skill's corpus, a fresh **executor** sub-agent runs the skill against the case prompt; a fresh **grader** sub-agent then scores that run's transcript against the case's assertions — judging only real tool calls, never prose or marker text. The output is a per-assertion pass/fail report with verbatim evidence, plus an optional old-vs-new score diff. This skill is **dev-only**: it reads the repo's `evals/` corpus and (for `--compare`) `.git` history, so it is deliberately NOT distributed into installed plugin caches. --- ## Argument Parsing | Token | Meaning | Default | |-------|---------|---------| | `[skill-name]` | The skill to evaluate; resolves its corpus at `evals/<skill-name>.evals.json`. | required unless `--all-evals` | | `--compare <ref>` | Also score the OLD version of the skill materialized from git `<ref>` and report an old-vs-new diff. | off (grade current version only) | | `--all-evals` | Evaluate every `evals/*.evals.json` corpus in the repo, not just one skill. | off | | `--dry-run` | Resolve + validate the corpus and print the plan (cases × assertions) without dispatching executors/graders. | off | Parse `$ARGUMENTS`: the first non-flag token is `[skill-name]`; `--compare` consumes the following token as `<ref>`. If neither `[skill-name]` nor `--all-evals` is given, stop and ask which skill to...