The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines an agentic workflow that processes external training data and user-provided questions which can influence the agent's logic and tool usage.\n
Ingestion points: The trainset parameter in the optimizer.compile method and the question argument in the ResearchAgent.forward method in SKILL.md.\n
Boundary markers: The implementation lacks explicit delimiters or system instructions to prevent the agent from following malicious instructions embedded within the processed data.\n
Capability inventory: The agent is equipped with a search tool (network access via ColBERTv2) and a calculate tool (Python execution capability) as defined in SKILL.md.\n
Sanitization: No input validation or sanitization is performed on user-provided questions or training examples before they are processed by the agent.\n- [DYNAMIC_EXECUTION]: The skill implements a tool for performing mathematical calculations by executing Python code dynamically at runtime.\n
Evidence: The calculate function in Phase 2 of SKILL.md utilizes dspy.PythonInterpreter() to evaluate strings provided by the agent.\n- [DATA_EXFILTRATION]: The skill performs network operations to a remote service for knowledge retrieval.\n
Evidence: The search tool in SKILL.md connects to an external IP address (http://20.102.90.50:2017/wiki17_abstracts) to query a ColBERTv2 knowledge base.

dspy-gepa-reflective