| name | yao-doctor-skill |
| description | Audit local and OpenClaw skill libraries for privacy theft, credential theft, stealthy exfiltration, unsafe execution chains, deceptive instructions, and persistence behavior, then generate a visual HTML security report. Use when Codex needs to security-review one skill or a full skill library before install, enablement, or execution. Do not use for generic cleanup, usage estimation, or app-level security review outside skill packages. |
Yao Doctor Skill
Boundary
Own this recurring job: security-review skill packages as untrusted code and untrusted instructions, separate broad capability risk from actual unsafe behavior, and produce evidence-backed recommendations before the skills are installed or executed.
Do not route here for:
- generic code security reviews outside skill folders
- skill cleanup, archival, deduplication, or stale-usage analysis
- privacy policy writing without scanning files
- creating a brand-new skill with no security audit need
Default Workflow
- Run
scripts/run_yao_doctor_skill.py [root] [more-roots].
If no roots are provided, auto-discover the current workspace, OpenClaw skill roots, Codex and Claude local skill roots, Codex plugin-cache skills, supported local workbench configuration surfaces, and project-level workbench surfaces such as repo AGENTS.md, CLAUDE.md, or selected .claude/.codex files.
- Scan every discovered skill directory containing
SKILL.md, plus recognized workbench surfaces such as local Codex and Claude configuration bundles and project-local agent configuration areas.
- For each skill, score two separate layers:
- capability risk: what the skill can access or execute
- unsafe behavior: what the skill appears to do with sensitive data or hidden actions
- Reuse the incremental cache when possible, and surface per-target diffs plus evidence type and confidence in the final report.
