The Agent Skills Directory

[SAFE]: Heuristic detections of destructive system commands (such as rm -rf / and chmod 777) are identified as false positives. These strings appear within the skill's security checklists and verification rules for detection purposes, rather than as commands to be executed.\n- [COMMAND_EXECUTION]: The skill invokes local tools including awk, git, find, and a project-specific script (node tools/pattern-alignment.js) to perform structural and freshness checks on other skills. These operations are scoped to auditing tasks and governed by explicit boundary rules prohibiting modification or execution of audited files.\n- [EXTERNAL_DOWNLOADS]: Uses curl and the GitHub CLI (gh) to verify the status of external references and repository metadata. These network operations target well-known services for the purpose of verifying link integrity and skill freshness.\n- [PROMPT_INJECTION]: The skill processes content from other (potentially untrusted) skills, creating a surface for indirect prompt injection.\n
Ingestion points: Reads SKILL.md and referenced files within audited skill directories.\n
Boundary markers: Includes 'Boundary Rules' mandating read-only static analysis and strictly forbidding execution of audited skills.\n
Capability inventory: Shell execution (awk, git, find, node), file system reads, and network metadata checks (curl, gh).\n
Sanitization: No explicit content sanitization is mentioned before audited content is included in health reports.

bs-skill-health