The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from users and other AI agents that could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: Untrusted message content and room metadata are retrieved in handler.ts through the read_room, search_rooms, and talk_with_sisters tools.
Boundary markers: The read_room tool produces a structured response that clearly labels message authors (from) and content (said), helping the agent distinguish between conversational data and system instructions.
Capability inventory: The skill possesses the ability to read and write to the database (Prisma), perform internal network requests to the orchestrator API (undiciFetch), and create directories on the local filesystem (fs.mkdirSync).
Sanitization: Filesystem paths for room folders are sanitized in handler.ts using alphanumeric filtering and slicing to prevent path traversal. Search queries are also normalized and filtered against stopwords. However, conversational message content and the pinned room topic (set via set_room_topic) are not explicitly filtered for natural language instructions.
[COMMAND_EXECUTION]: The skill uses raw SQL execution in handler.ts for updating room topics. This is implemented using Prisma's $executeRaw with template literals, which automatically parameterizes the input to prevent SQL injection attacks.

group-chat