The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents methods for ingesting application data into the agent's context through useAgentContext and useFrontendTool. While this creates a surface for indirect prompt injection from untrusted application data, it is a primary function of the library and includes standard developer documentation for implementation.
Ingestion points: useAgentContext (shares JSON-serializable values), useFrontendTool (parameters passed to handlers).
Capability inventory: Subprocess calls/file access depends on the specific tools implemented by the developer using the library; the library itself provides the transport layer.
Boundary markers: Documentation refers to Zod schemas for parameter validation which acts as a data-level boundary.
Sanitization: Not explicitly covered in snippets, though the use of Zod for parameter validation is recommended in the examples.
[COMMAND_EXECUTION]: The evaluation configuration (eval.yaml) contains standard setup commands for a Docker environment, including apt-get update and the installation of git and jq. These are standard tools and do not represent a security risk within the context of the skill's testing framework.

copilotkit-develop