The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external tools for document processing, specifically textutil on macOS and the mineru-open-api utility for PDF OCR. These invocations in src/parser.py use secure argument lists rather than raw shell strings, which prevents shell injection.
[COMMAND_EXECUTION]: In src/cli.py, the setx command is used on Windows to persist API keys in the user environment. This is a standard functionality for CLI tools to maintain state across sessions.
[INDIRECT_PROMPT_INJECTION]: The skill processes external documents (.pdf, .docx, .md) which are chunked and stored in a vector database. This creates an indirect prompt injection surface common to RAG systems, as the agent will eventually process the retrieved text which may contain adversarial instructions. The skill does not implement specific sanitization for this content, but this is consistent with its primary purpose and carries low risk.
[SAFE]: The detection of eval() in src/reranker.py by static analysis tools is a false positive; the code calls the PyTorch .eval() method to set a machine learning model to evaluation mode, which is distinct from the dangerous Python eval() built-in function.

local-rag