The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a pattern for ingesting data from external tool handlers into the conversation history. While this creates a surface for indirect prompt injection (where tool outputs could contain malicious instructions), the skill provides standard implementation details for agentic loops.
Ingestion points: The dispatch function processes toolCall.input and returns output from registered handlers in SKILL.md.
Boundary markers: The implementation does not include specific boundary delimiters or sanitization filters for the tool output beyond basic stringification.
Capability inventory: The loop is designed to call arbitrary handlers based on model tool_use blocks.
Sanitization: The skill recommends using JSON.stringify for result injection and includes logic to handle and report errors back to the model using is_error: true.
[SAFE]: The skill does not perform any sensitive file operations, network requests, or credential handling. It is primarily educational content providing code templates for developers.

tool-execution-loop