The Agent Skills Directory

[SAFE]: The skill provides instructions and documentation for a RAG performance benchmarking tool. It relies on local scripts located in a scripts/ directory and standard Python tools like uv and pip for installation. No remote code execution or unauthorized access patterns were found.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data, specifically RAG server responses (SSE streams) and synthetic queries generated by an LLM. While this presents an indirect prompt injection surface common to RAG evaluation tools, there are no specific malicious patterns detected. The skill includes a 'playbook' for summarization which provides a layer of structure for the agent's analysis.
Ingestion points: RAG server SSE streams (citations and metrics) and responses from synthetic LLM endpoints (chat completions).
Boundary markers: None explicitly mentioned in the documentation for the agent to use when processing external data.
Capability inventory: The skill uses uv run and python -m to execute local benchmark scripts, performs file I/O for reports and query storage, and makes network requests to RAG servers and LLM endpoints.
Sanitization: The synthetic query generator performs minimal cleanup of LLM output, such as stripping leading numbering and whitespace.
[COMMAND_EXECUTION]: The documentation provides standard shell commands for managing dependencies and running benchmarks locally using the uv tool. These commands are typical for development environments and do not involve privilege escalation or unauthorized system modifications.

rag-perf