sdsrss/agentsmd/agentsmd-safety-coverage-audit

agentsmd-safety-coverage-audit

Use the selected SKILL.md absolute path from the live skills list; never infer it from the process cwd. Define and call the adjacent launcher in the same shell. The launcher preserves selected bundle → manifest-owned standalone → versioned CLI identity checks and exports plugin context only for a verified selected bundle.

SKILL_MD="<selected SKILL.md absolute path from the live skills list>"
agentsmd_skill_run() {
  node "$(dirname "$SKILL_MD")/scripts/agentsmd-run.js" "$SKILL_MD" "$@"
}

Header comments and deny/advisory strings are documentation, not proof. This audit cross-references the hook layer against its own claims and the manifest, four ways:

agentsmd_skill_run      # human report
agentsmd_skill_run --json
agentsmd_skill_run --hook=pre-bash-safety-check.sh

• Arrow-claim sweep — every → claim (header block or deny/advisory string) is split on →/; and each clause keyword-grepped against the hook's code body (header stripped). A clause with zero hits = a partial-impl candidate: the header promises a link the code never implements (the failure this audit exists to catch).
• Manifest cross-ref — every enforcement: hook|both rule whose rule_hits_section is live must be emitted by some hook. A live section with no emitter is an unimplemented gap; a hook-enforced rule whose section is NOT live reads as hook-planned (its hook isn't built yet — expected, not a gap).
• Bypass-token coverage — a documented [allow-*] escape hatch must appear on a code line (a real guard), not just in a comment.
• Orphan emission — a §-section literal a hook emits that no manifest rule declares (telemetry the governance layer can't see).

TOTAL GAPS: 0 + exit 0 = clean; exit 3 = at least one gap (the summary lists each). Wired into npm test as a coherence gate, so a hook that documents more than it implements fails CI. From the repo instead of an install: node scripts/safety-coverage-audit.js.