The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines an agentic loop that processes user-supplied input from an external queue or stdin, which is a standard pattern for interactive agents.
Ingestion points: Untrusted data enters the agent context through Queue.Queue<string> in the conversation function and process.stdin in the readStdinInto helper (SKILL.md).
Boundary markers: User input is wrapped in Items.userText objects, providing structural role separation in the history passed to the model.
Capability inventory: The skill uses Toolkit.executeAll (SKILL.md) to process tool calls generated by the model, enabling autonomous actions based on the provided tools.
Sanitization: Input is passed directly to the model as text without specific sanitization or filtering, relying on the model's own safety guardrails.

effect-uai-agentic-loop