The Agent Skills Directory

[PERSISTENCE]: The skill provides a mechanism to persist behavior policies, memories, and new skills across sessions using the global_ parameter. This is the intended primary purpose of the tool, allowing the agent to refine its 'harness' based on conversation outcomes.
[INDIRECT_PROMPT_INJECTION]: The skill accepts natural language instructions that are used to guide the host-side refinement process. If these instructions are derived from untrusted user data or external sources, they could be used to influence the agent's persistent instructions.
Ingestion points: The instructions argument in the run() function within src/refine/__init__.py and the corresponding interface in SKILL.md.
Boundary markers: None; the skill passes the raw string instructions directly to the host-side bridge without delimiters or sanitization headers.
Capability inventory: The skill can modify system prompts, memories, and available skills, which directly controls the agent's future behavior.
Sanitization: The code includes basic Python type checking (ensuring instructions is a string) but does not validate or sanitize the natural language content of the instructions.

refine