The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill features the dspy.ProgramOfThought module, which is designed to generate and execute Python code at runtime to solve reasoning and mathematical tasks. This is a core capability of the framework but constitutes a dynamic execution risk if the generated code is influenced by untrusted input.
Evidence: Found in SKILL.md and references/modules.md describing the generation and execution of Python code for calculations.
[DYNAMIC_EXECUTION]: A code example for a calculation tool utilizes the eval() function to process mathematical expressions. Although it attempts to limit the execution environment by restricting builtins, eval() remains a high-risk pattern for code injection.
Evidence: references/examples.md contains a calculate function using eval(expression, {"__builtins__": {}}, {}).
[INDIRECT_PROMPT_INJECTION]: The skill architecture heavily features RAG (Retrieval-Augmented Generation) and multi-agent workflows, which are susceptible to indirect prompt injection if untrusted data retrieved from external sources contains malicious instructions.
Ingestion points: External data is ingested via dspy.Retrieve in SKILL.md and references/examples.md.
Boundary markers: Signatures in SKILL.md define structured inputs but do not include explicit security delimiters to prevent instruction override.
Capability inventory: The skill includes tools for code execution (ProgramOfThought) and tool usage (ReAct) across multiple files.
Sanitization: The examples show basic usage of restricted eval() but lack comprehensive input validation for LLM-generated code blocks.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions to install the framework directly from a public Git repository, which involves downloading external code.
Evidence: SKILL.md includes the installation command pip install git+https://github.com/stanfordnlp/dspy.git.

dspy