The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess module to execute codex and claude CLI tools.
Evidence: Found in scripts/adapters.py within the _run_cli function. It executes commands using a list of arguments (avoiding shell injection) and restricts the environment variables passed to the child process to a whitelist (PATH, HOME, USER, etc.).
[EXTERNAL_DOWNLOADS]: The skill makes network requests to the OpenRouter API to fetch model responses.
Evidence: Found in scripts/adapters.py using urllib.request to connect to https://openrouter.ai/api/v1/chat/completions. OpenRouter is a well-known service for LLM inference.
[DATA_EXPOSURE]: The skill reads API credentials from a local configuration file.
Evidence: Found in scripts/adapters.py where it attempts to read OPENROUTER_API_KEY from ~/.claude/auth/ai-ml-services.env if the environment variable is not set. This is a standard practice for local credential management.
[PROMPT_INJECTION]: The skill processes data from external models, creating a surface for indirect prompt injection (Category 8).
Ingestion points: Output from OpenRouter, Codex, or Claude CLI tools entering the agent context via the dispatch function in scripts/adapters.py.
Boundary markers: Responses are wrapped in Markdown headers (e.g., ### Reference N) and the skill provides explicit instructions to the aggregator agent (_SYNTHESIS_HINT in scripts/moa.py) to treat these references as external viewpoints rather than direct instructions.
Capability inventory: The skill can execute subprocesses (codex, claude) and perform network requests via urllib.
Sanitization: Includes a sanitize_error function in scripts/adapters.py that uses regex to mask API keys and bearer tokens in error messages before they are displayed.

moa