skill-auditor

/skill-auditor — Three-pass skill quality audit

Validates a skill's SKILL.md against the unified AgentOps template. Pass 1 wraps heal-skill for structural hygiene; Pass 2 adds 8 content-discipline checks not covered by heal; Pass 3 folds the 10-category Skill Quality Rubric (docs/reference/skill-quality-rubric.md) into the report as a deterministic 0-30 productization score (advisory). The report also includes an advisory Context Density Rule block for intent, boundary, evidence, decision, constraint, and next action coverage.

⚠️ Critical Constraints

• Auditor is read-only. Reports findings; never modifies the target. Why: PR-002 (external validation) — the auditor must remain a separate gate from the implementer. To repair findings: use heal-skill --fix for Pass-1 issues, hand-edit for Pass-2 issues.
• Pass 1 delegates, never reimplements. The auditor calls /heal-skill --check <target> and parses its output. Why: PR-006 (cross-layer consistency) — heal-skill's checks are the source of truth for structural hygiene; reimplementation creates drift.
• Pass 2 must accept AgentOps' existing conventions. Specifically description-has-triggers accepts THREE valid forms (YAML | block scalar OR Triggers:/Use when: markers OR metadata.triggers array with 3+ items). Why: finding f-2026-05-06-auditor-checks-must-fit-host-conventions — auditor checks must validate against the host substrate's existing valid artifacts before promotion to required gate.
• Verdict aggregation rule: any check returns fail → FAIL; otherwise any returns warn → WARN; otherwise PASS. Why: prevents silent severity downgrade.
• Density coverage is advisory-only. Missing density fields never changes

the PASS/WARN/FAIL verdict and does not satisfy packet-boundary enforcement in soc-2c1p.1. Why: the hard Context Density Rule belongs at execution packet boundaries; this skill only helps reviewers find low-signal prose.

• Pass 3 rubric is advisory-only. The 0-30 rubric score never changes the

PASS/WARN/FAIL verdict. Why: Pass 1+2 gate template conformance (does this ship); the rubric measures market-facing maturity (is this product-grade) — a low rubric score on a structurally-clean skill is a productization backlog signal, not a ship blocker (soc-ads5v).

• Pass 3 scoring is deterministic and rubric-sourced. The 10 categories

come verbatim from docs/reference/skill-quality-rubric.md; each gets a 0-3 score plus an explainable reason derived only from the skill directory contents. Why: an explainable, reproducible score is auditable; an LLM-graded one is not.

What It Detects

Pass 1 — delegated to heal-skill

| heal.sh code | Check | |--------------|-------| | MISSING_NAME | frontmatter name present | | MISSING_DESC | frontmatter description present | | NAME_MISMATCH | frontmatter name matches directory | | UNLINKED_REF | references/*.md linked from SKILL.md | | DEAD_REF | linked references actually exist | | SCRIPT_REF_MISSING | scripts referenced exist | | CATALOG_MISSING | user-invocable skills in using-agentops/ catalog |

Pass 2 — 8 NEW checks

| # | Check id | Severity | |---|----------|----------| | 1 | description-has-triggers | WARN (downgraded from FAIL after pilot) | | 2 | constraints-frontloaded | WARN | | 3 | rationale-present | WARN | | 4 | verification-checkpoints | WARN | | 5 | output-spec-explicit | FAIL | | 6 | quality-rubric | WARN | | 7 | references-modularization | WARN (conditional, only if SKILL.md > 400 lines) | | 8 | trigger-clarity | WARN (downgraded from FAIL after pilot) |

Full check definitions and accepted forms in [references/audit-checks.md](references/audit-checks.md).

Advisory density report

The JSON report includes a separate density block with six report-only fields: intent, boundary, evidence, decision, constraint, and next_action. Read [references/context-density-checks.md](references/context-density-checks.md) for detection rules, limits, and false-positive handling.

Pass 3 — rubric scoring (10 categories, advisory)

audit.sh runs scripts/score_agentops_skill.py --audit-block and folds the result into audit-report.json under a rubric key. The 10 categories come verbatim from [docs/reference/skill-quality-rubric.md](../../docs/reference/skill-quality-rubric.md) (read it for the per-score 0/1/2/3 definitions): trigger_quality, kernel_clarity, progressive_disclosure, helper_scripts, validation, self_test, assets_templates, subagents_roles, safety_boundaries, packaging.

Each category scores 0-3 (0 missing/unsafe → 3 product-grade and mechanically validated) with an explainable reason. Total 0-30 maps to a rating band: C (0-10), B (11-20), A (21-26), S (27-30). The score is advisory — it never changes the PASS/WARN/FAIL verdict.

Standalone (markdown) for picking the smallest productization patch:

python3 skills/skill-auditor/scripts/score_agentops_skill.py skills/<name> --markdown

Use it to pick the smallest patch (SELF-TEST.md, linked references, helper scripts, assets, subagents, safety boundaries, or validation), then re-run this auditor and heal-skill.

Execution Steps

Step 1: Pass 1 (heal-skill delegation)

bash skills/heal-skill/scripts/heal.sh --check <target>

Capture stdout. Each line [CODE] <path>: <msg> becomes one Pass-1 finding.

Step 2: Pass 2 (8 NEW checks)

For each check_* function in scripts/audit.sh, run against <target>/SKILL.md. Each emits pass, warn, or fail to stdout.

Checkpoint: Pass 2 must run independently of Pass 1 (no shared state); a heal.sh failure does NOT short-circuit Pass 2.

Step 3: Pass 3 (rubric scoring)

python3 scripts/score_agentops_skill.py <target> --audit-block

audit.sh calls this and embeds the result under the report's rubric key. Each of the 10 rubric categories gets a 0-3 score + reason; total 0-30, rating band C/B/A/S. If python3 or the scorer is unavailable, rubric is emitted as null (fail-open) and the JSON stays valid.

Checkpoint: Pass 3 is advisory — its score is computed but NOT counted in the verdict.

Step 4: Aggregate verdict

fails > 0  → FAIL
warns > 0  → WARN
otherwise  → PASS

Density coverage and the Pass-3 rubric are computed before emission but are NOT counted in the verdict.

Step 5: Emit report

JSON conforming to schemas/audit-report.json to stdout (or to file with --json <path>); markdown summary (including the Pass-3 rubric line) to stderr.

Output Specification

Format: JSON conforming to schemas/audit-report.json (default) plus markdown text summary. Filename: typically .agents/audits/<skill-name>-audit.json when --json <path> is supplied; otherwise stdout. Exit code: 0 for PASS or WARN; 1 for FAIL; 2 for usage error or missing target.

Density advisory: JSON includes density.status, density.fields[], and density.summary. Treat missing fields as review prompts, not gates.

Rubric (Pass 3): JSON includes rubric.total_score, rubric.max_score, rubric.rating, rubric.advisory (always true), and rubric.categories[] (10 entries, each {category, score, reason}). Emitted as null if the scorer is unavailable. Treat the score as a productization backlog signal, not a gate.

Quality Rubric

• [ ] Auditor never modifies target SKILL.md or any other file
• [ ] Pass 1 invokes heal.sh with --check (NOT --fix)
• [ ] All 8 Pass-2 checks emit one of: pass, warn, fail, n/a
• [ ] description-has-triggers accepts all three valid forms (verified by running auditor against AgentOps' existing single-line-description skills like forge, heal-skill, council)
• [ ] Aggregate verdict applies max-severity rule (no silent downgrade)
• [ ] Density advisory reports all six fields without changing the aggregate verdi