The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/mcp-search.py executes a local binary (defaulting to ms) to perform skill searches. It correctly avoids shell execution by passing arguments as a list and uses shlex.split to safely parse the binary path from the MS_BIN environment variable. It also manages the lifecycle of the search process correctly using process groups (start_new_session=True) to ensure clean termination.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a search engine for other AI skills, creating a surface where malicious instructions embedded in the indexed content or crafted queries could potentially influence the agent.
Ingestion points: User-provided search queries and the contents of retrieved SKILL.md files processed via the search helper.
Boundary markers: Search results are encapsulated in structured JSON objects, separating metadata from full guidance content.
Capability inventory: The skill can execute the local ms search binary via subprocess.
Sanitization: Search queries are JSON-encoded to prevent protocol-level injection, and retrieved data is parsed through standard JSON libraries.

ms