The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a pattern for extracting entities and facts from user messages using an LLM, which creates a surface for indirect prompt injection.
Ingestion points: User messages are directly interpolated into extraction prompts within the EntityMemory.extractAndStore method.
Boundary markers: The code examples use simple double quotes ("${message.content}") to wrap untrusted content, lacking robust delimiters or explicit instructions to the LLM to ignore embedded commands within the processed text.
Capability inventory: The skill utilizes llm.complete for extraction and context construction. It does not perform file-system or network writes directly based on this input.
Sanitization: The skill's body includes architectural warnings to treat extracted text as "untrusted evidence" and avoid promoting generated claims into facts, though the provided code snippets themselves do not implement input sanitization.

conversation-memory