The Agent Skills Directory

EXTERNAL_DOWNLOADS (LOW): The skill utilizes npx memberstack-cli to execute commands. While this involves downloading the package from the NPM registry, it is the standard method for using the tool and is managed via a controlled wrapper script.
INDIRECT_PROMPT_INJECTION (LOW): The skill identifies that data returned from the Memberstack API (like member names or record content) is untrusted user-generated content. It mitigates this via a wrapper script (run_memberstack.py) that implements:
Ingestion points: CLI output from memberstack-cli and reference content from search_references.py.
Boundary markers: Output is wrapped in --- BEGIN MEMBERSTACK CLI OUTPUT --- markers.
Capability inventory: The agent can execute Memberstack CLI commands via subprocess.run (without shell=True).
Sanitization: A regex-based sanitizer (_INJECTION_RE) checks for and removes common prompt injection patterns (e.g., 'ignore previous instructions') from the output and documentation before the agent processes it.
CREDENTIALS_UNSAFE (SAFE): The skill proactively blocks attempts to access local authentication files (e.g., ~/.memberstack/auth.json) by checking command arguments in the wrapper script and explicitly forbidding the agent from reading these files in the SKILL.md instructions.
COMMAND_EXECUTION (SAFE): Command execution is restricted to the specific Memberstack CLI utility through a Python wrapper that validates arguments against sensitive file paths and enforces a confirmation step for destructive commands.

memberstack-cli