The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the usage of dspy.ProgramOfThought, a core module designed to generate and execute Python code at runtime to handle arithmetic and symbolic logic. While a primary feature of the framework, it involves executing dynamically generated code based on model outputs.
[DYNAMIC_EXECUTION]: Documentation in references/examples.md and references/modules.md includes implementation examples using the eval() function to create calculation tools. Although the examples show a restricted environment using {"__builtins__": {}}, this pattern facilitates the execution of string-based inputs as code.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates how to build RAG and agent systems that retrieve information from untrusted external sources like web searches and Wikipedia.
Ingestion points: Data enters the agent context through search_wikipedia, search_web, and dspy.Retrieve modules.
Boundary markers: The framework uses Signatures to structure prompts, but the documentation does not explicitly detail sanitization or instructions to ignore malicious embedded commands in retrieved data.
Capability inventory: The skill leverages capabilities for network requests (LM providers), local file storage (ChromaDB persistence), and dynamic code execution.
Sanitization: While eval() examples use limited global scope, no general filtering of retrieved external content is shown before interpolation into prompts.
[EXTERNAL_DOWNLOADS]: The installation guide suggests fetching the latest development version of the library from the official Stanford NLP GitHub repository. This is a reference to a well-known academic research organization.

dspy