The Agent Skills Directory

[SAFE]: The skill documents legitimate framework capabilities for the AG2 library, focusing on cross-cutting concerns like logging, retries, and history management.
[PROMPT_INJECTION]: The middleware hooks provide an interface for processing data from external sources, which is a standard surface for indirect prompt injection.
Ingestion points: Data enters the agent loop via the on_turn, on_llm_call, on_tool_execution, and on_human_input hooks.
Boundary markers: No specific delimiters or boundary markers are defined in the provided BaseMiddleware interface; security depends on custom implementation logic.
Capability inventory: Middleware hooks can perform prompt mutation, history truncation, tool result modification, and observability tasks through standard Python code.
Sanitization: The documentation suggests using the on_tool_execution hook for redacting tool results, providing a pattern for sanitizing potentially sensitive data.

ag2-middleware