The Agent Skills Directory

[DATA_EXPOSURE]: The skill documents a pattern for context injection in references/6.context.md that reads local file content based on paths provided at runtime. While intended for loading session documents, this demonstrates a capability for file system access that requires careful restriction in production environments.
[COMMAND_EXECUTION]: In references/7.mcp_server.md, the tutorial uses MultiServerMCPClient to launch local Model Context Protocol (MCP) servers by executing shell commands (e.g., python -m ...). This is a standard integration method but involves subprocess execution.
[INDIRECT_PROMPT_INJECTION]: The skill outlines patterns for Retrieval-Augmented Generation (RAG) and Web Search in references/10.rag.md and references/11.web_search.md. These components ingest untrusted data from external URLs and search providers. The skill provides remediation examples in references/3.middleware.md, showing how to implement content filtering and PII detection to mitigate these risks.
[EXTERNAL_DOWNLOADS]: The guide references and utilizes official APIs and resources from established services including Alibaba DashScope, Tavily, and the author's official documentation site (luochang212.github.io). These are documented neutrally as intended functional dependencies.
[DYNAMIC_EXECUTION]: A scientific calculation tool in scripts/tools/tool_math.py implements mathematical expression evaluation using Python's ast module. The implementation follows security best practices by utilizing a strict whitelist of allowed operators and functions and a custom NodeVisitor to prevent arbitrary code execution.

dive-into-langgraph