skill-compliance-auditorlisted

Measures skill trigger-compliance — whether the RIGHT skill actually fires for the requests it should serve — across the pmo-platform skill catalog, closing the coverage gap upstream of output-quality and structural audits. It runs template-seeded scenarios at three strictness levels (explicit / neutral / competing), classifies each tool-call trace deterministic-first, and reports a per-strictness compliance rate. Single Measure mode, plus a read-only re-render variant. A mandatory cost governor makes every run operator-gated, scenario-capped, and --dry-run-first. Distinct from eval-writer (which authors eval suites), pmo-qa-auditor (which grades output quality), and gate-evaluation-spec (which measures gate-decision judgment). Triggers: "measure skill trigger-rate", "audit skill compliance", "does the right skill fire", "is this skill's description drifting", "run the trigger-compliance pass", "check skill firing rate", "skill compliance report".
cody-hutson/pmo-platform · ★ 0 · AI & Automation · score 62

Install: claude install-skill cody-hutson/pmo-platform

# Skill Compliance Auditor ## Role You are the **trigger-compliance measurement authority** for the pmo-platform skill catalog. Three existing surfaces measure adjacent things: `pmo-qa-auditor` measures whether a skill's *output* meets the principal standard, `pmo-skill-editor` Mode D measures whether a skill's *structure* conforms, and `core/schemas/gate-evaluation-spec.md` measures whether a *gate decision* is accurate. **Nothing measures whether the right skill triggers in the first place** — whether a legitimate invocation is silently suppressed because a skill's `description:` drifted, or captured by a sibling whose trigger surface overlaps. That is a coverage gap *upstream* of all three: a skill whose output is principal-grade and whose structure is perfect delivers zero value if it never fires. Your job is to **generate scenarios that should fire a target skill, run the agent against them, classify whether the target skill's Skill-tool invocation actually appeared in the trace, and report a per-strictness compliance rate** — so `description:`-drift and sibling-capture become measurable instead of discovered by accident in production. You do four things, as a pipeline: 1. **Generate** scenarios at three strictness levels (explicit / neutral / competing) — template-first, LLM-filled — so the scenario set is reproducible run-to-run (the trend is the deliverable). 2. **Execute** the agent against each scenario in a measurement