The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell scripts and binary commands to inventory local skills and manage filesystem state.
bin/skill-scan.sh performs recursive directory walks and metadata extraction from other installed skills.
lib/cleanup-macos.mjs uses spawnSync to invoke git, ditto, and cp for transactional moves and state recovery.
lib/runtime-adapters.mjs executes agent binaries such as codex, claude, and cursor-agent with diagnostic arguments to verify their runtime configuration.
[DYNAMIC_EXECUTION]: The skill implements a sophisticated runtime compilation and execution pattern on macOS.
lib/cleanup-macos.mjs includes logic to locate the system's C compiler (clang) via xcrun and compile a bundled helper source file (native/cleanup-macos-helper.c) into a temporary binary.
This compiled helper is used to perform atomic, race-free filesystem operations (e.g., renameatx_np) and security metadata verification.
This behavior is consistent with the skill's purpose as a system-level governance utility and is applied exclusively to the skill's own internal components.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a high-privilege ingestion point for data from other skills.
Ingestion points: bin/skill-scan.sh reads and parses SKILL.md files from any directory on the system recognized as an agent skill path.
Boundary markers: The scanner uses jq to strictly serialize extracted metadata into JSON, reducing the risk of schema confusion or downstream injection.
Capability inventory: The skill has broad capabilities including file deletion (via quarantine), directory creation, and network probing (of local agent binaries).
Sanitization: The scanner redacts potential secrets and restricts promoted source claims to well-known domains like GitHub.

skill-hygiene