The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The compact.run() function accepts an optional instructions string that directs the summarization process. While this represents a potential surface where untrusted data could be interpolated into host-side instructions, the risk is minimal as it is a core feature for context management within the agent's controlled environment.
Ingestion points: instructions argument in src/compact/__init__.py.
Boundary markers: None explicitly implemented in the skill code.
Capability inventory: Uses the rlm.host_request bridge for platform communication.
Sanitization: Validates that input is a string or None, preventing primitive type-based injection.

compact