| name | decisions-harness-audit |
| description | Use when auditing DecisionsAI, Codex, Claude, Cursor, Gemini, Pi, Cline, RTK, workflow pre_chain skills, hooks, projected commands, skill routing, or harness setup for drift, duplication, dead references, slow hooks, bloated context, or stale config. |
Decisions Harness Audit
Use this for read-only harness health checks. It audits what is actually installed and projected, then produces verified findings with benchmark commands.
Workflow
- Map the surface: Decisions bundled skills, registry rows, workflow
pre_chain usage, projected project skills, local agent configs, RTK, and available CLI backends.
- Inventory by domain: hooks/config, instruction surface, skill registry, projected commands, workflow chains, session/history friction, state files, and secrets exposure.
- Verify each finding against disk before reporting it. Rejected findings stay out of recommendations.
- Write findings as delete/update/enhance/new recommendations with evidence, confidence, effort, and benchmark commands.
- For Decisions workflows, return artifact paths and a PASS/MISS/SKIP summary. Do not apply fixes.
References
- Read
references/harness-adapters.md for harness-specific paths and command equivalents.
- Use
scripts/harness_inventory.py for a deterministic first-pass inventory.
Rules
- This skill is read-only.
- Do not remove, rewrite, or move files.
- Do not recommend loosening permissions as an optimization.
- Treat duplicate skills as merge candidates only when the local Decisions skill is not already canonical.
- AgentManager is reference material only; do not suggest replacing DecisionsAI board orchestration with it.