askit-evaluate
Purpose
Assess a known, local component or plugin against STANDARD.md. Three modes. conformance (the default) runs the deterministic portable scripts and returns a per-rule report (pass / warn / error), the satisfied tier, and concrete remediation. behavioral runs a skill against its eval-set and judges whether it triggers and behaves as expected, delegating to askit-quality-grader. review forms a qualitative judgment (correctness, altitude, naming, whether a component is warranted), delegating to askit-reviewer. Only conformance is deterministic and gate-safe; behavioral and review are opt-in LLM-judged passes that produce evidence, never a CI gate result (Design Principle 3, ADR 0023).
When to use
When the user asks to evaluate, audit, or check a skill or plugin, asks "what tier is this" or "what is blocking the next tier" (conformance), asks whether a skill actually triggers and behaves correctly (behavioral), or wants a qualitative review (review).
conformance mode (default, deterministic)
- Determine the target path (a plugin root with
library.json, or a single skill directory with SKILL.md).
- Run:
node scripts/evaluate.mjs <path> --json.
- Present the findings grouped by rule, the tier (for a plugin), and the remediation. Lead with errors, then warnings.
- For a shareable, designed report, render the same object:
node scripts/evaluate.mjs <path> --format=html --out report.html (a self-contained page for a non-engineer) or --format=md (the Markdown twin for PR review and agents). It renders the same deterministic object the terminal shows, adds no judgment, and does not change the verdict. See references/report-format.md.
- If there are warnings or errors, point the user at
askit-build-skill in improve mode to fix them.
behavioral mode (opt-in, LLM-judged)
- Locate the target's eval-set under
evals/ (triggering {query, should_trigger} cases and {given, expect} behavior cases). The evals/ convention is forward-looking: most real targets do not ship one. If it is absent, the grader DERIVES a case set instead - should-fire queries and adversarial near-misses from the target's description (read sibling skills' descriptions to make the no-fire cases genuinely competitive), plus behavior cases from the documented workflow - judges by static analysis of the artifact, and says so in the evidence.
- Delegate to
askit-quality-grader: it runs the skill against the cases (on-disk or derived) and judges fire / no-fire and output quality.
- Report the verdict per case with evidence. This is evidence, not a gate result; it never fails CI.
review mode (opt-in, qualitative)
- Delegate to
askit-reviewer: it reviews the component or change for correctness, altitude, naming, and whether it is warranted.
- Report findings by severity with remediation. Like behavioral, this is advisory, not a gate result.
Scope
The conformance core is deterministic and is what CI runs (Design Principle 3). Behavioral and review are opt-in LLM-judged layers beside the gate: the Standard defers the full eval engine to roadmap, and G3 requires only the deterministic baseline (ADR 0023). See references/report-format.md for the report shape. The delegations are permitted in agents/_chain-permitted.yaml (askit-evaluate may invoke askit-evaluator for conformance, askit-quality-grader for behavioral, and askit-reviewer for review).
1---2name: askit-evaluate3description: Evaluates a skill or plugin against the Advanced Skill Library Standard across three modes, producing deterministic conformance findings and a tier, an opt-in behavioral pass, and a qualitative review. Use when you want to audit conformance, judge whether a skill behaves and triggers correctly, get a qualitative review, or see what blocks the next tier.4---56# askit-evaluate78## Purpose9Assess a known, local component or plugin against `STANDARD.md`. Three modes. `conformance` (the default) runs the deterministic portable scripts and returns a per-rule report (pass / warn / error), the satisfied tier, and concrete remediation. `behavioral` runs a skill against its eval-set and judges whether it triggers and behaves as expected, delegating to `askit-quality-grader`. `review` forms a qualitative judgment (correctness, altitude, naming, whether a component is warranted), delegating to `askit-reviewer`. Only `conformance` is deterministic and gate-safe; `behavioral` and `review` are opt-in LLM-judged passes that produce evidence, never a CI gate result (Design Principle 3, ADR 0023).1011## When to use12When the user asks to evaluate, audit, or check a skill or plugin, asks "what tier is this" or "what is blocking the next tier" (conformance), asks whether a skill actually triggers and behaves correctly (behavioral), or wants a qualitative review (review).1314## conformance mode (default, deterministic)151. Determine the target path (a plugin root with `library.json`, or a single skill directory with `SKILL.md`).162. Run: `node scripts/evaluate.mjs <path> --json`.173. Present the findings grouped by rule, the tier (for a plugin), and the remediation. Lead with errors, then warnings.184. For a shareable, designed report, render the same object: `node scripts/evaluate.mjs <path> --format=html --out report.html` (a self-contained page for a non-engineer) or `--format=md` (the Markdown twin for PR review and agents). It renders the same deterministic object the terminal shows, adds no judgment, and does not change the verdict. See [references/report-format.md](references/report-format.md).195. If there are warnings or errors, point the user at `askit-build-skill` in `improve` mode to fix them.2021## behavioral mode (opt-in, LLM-judged)221. Locate the target's eval-set under `evals/` (triggering `{query, should_trigger}` cases and `{given, expect}` behavior cases). The `evals/` convention is forward-looking: most real targets do not ship one. If it is absent, the grader DERIVES a case set instead - should-fire queries and adversarial near-misses from the target's description (read sibling skills' descriptions to make the no-fire cases genuinely competitive), plus behavior cases from the documented workflow - judges by static analysis of the artifact, and says so in the evidence.232. Delegate to `askit-quality-grader`: it runs the skill against the cases (on-disk or derived) and judges fire / no-fire and output quality.243. Report the verdict per case with evidence. This is evidence, not a gate result; it never fails CI.2526## review mode (opt-in, qualitative)271. Delegate to `askit-reviewer`: it reviews the component or change for correctness, altitude, naming, and whether it is warranted.282. Report findings by severity with remediation. Like behavioral, this is advisory, not a gate result.2930## Scope31The conformance core is deterministic and is what CI runs (Design Principle 3). Behavioral and review are opt-in LLM-judged layers beside the gate: the Standard defers the full eval engine to roadmap, and G3 requires only the deterministic baseline (ADR 0023). See [references/report-format.md](references/report-format.md) for the report shape. The delegations are permitted in `agents/_chain-permitted.yaml` (`askit-evaluate` may invoke `askit-evaluator` for conformance, `askit-quality-grader` for behavioral, and `askit-reviewer` for review).
product-on-purpose