The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides boilerplate code and a zero-dependency template (SKILL.md, zero-dependency-server-template.md) that import exec and spawn from the Node.js child_process module. Although the documentation (Pattern 3 and 6) advises against their use for specific tasks, their presence in the core templates provides a mechanism for executing system commands.\n- [INDIRECT_PROMPT_INJECTION]: The bridge architecture described is designed to ingest data from external LLM clients and route it to an internal agent, creating an injection surface.\n
Ingestion points: Data enters the bridge via process.stdin as JSON-RPC messages (e.g., tools/call) in SKILL.md and zero-dependency-server-template.md.\n
Boundary markers: The templates do not define or implement delimiters or "ignore" instructions for the ingested tool arguments.\n
Capability inventory: The skill patterns include making authenticated HTTP requests to a Gateway API (hermesApi) and writing task state to the file system for persistence.\n
Sanitization: While Pattern 5 suggests using fs.existsSync() to validate project paths, there is no mention of sanitizing or escaping the content of goals or tasks before they are interpolated into prompts for the downstream agent.\n- [DATA_EXFILTRATION]: Pattern 1 in SKILL.md outlines a method for reading an API key from an environment file and transmitting it in the Authorization header of HTTP requests to an external Gateway API. While functional for an API bridge, this represents a pattern for handling and transmitting sensitive credentials over the network.

mcp-bridge-patterns