The Agent Skills Directory

[DYNAMIC_EXECUTION]: The evaluation script evals/verify-application-proposals.py uses exec() and compile() to execute Python code blocks extracted from the references/application-proposals.md file at runtime. This allows for arbitrary code execution derived from non-executable markdown files.
[INDIRECT_PROMPT_INJECTION]: The skill describes a pattern for "Typed proposals" in references/application-proposals.md that ingests untrusted JSON data into the agent's instructions. This creates a surface for indirect prompt injection where malicious content in the data could influence agent behavior.
Ingestion points: The context function in references/application-proposals.md injects ctx.deps.scoped_json into the agent's instructions.
Boundary markers: The instructions explicitly warn that "Request notes are untrusted data" and recommend returning only drafts to maintain application-level control.
Capability inventory: The skill defines an Agent capable of tool execution and model requests, although the specific recommendation for this pattern is to omit command tools.
Sanitization: The pattern relies on Pydantic BaseModel output validation to ensure the response adheres to a specific schema, but does not sanitize the input content beyond type validation.

pydanticai