The Agent Skills Directory

[DYNAMIC_EXECUTION] (MEDIUM): The skill highlights the use of dspy.ProgramOfThought in core/building-modules.md. This module programmatically generates and executes Python code at runtime to solve logic and math problems, which represents a significant execution vector if the model is influenced by malicious input.\n- [DATA_EXFILTRATION] (LOW): Multiple Python files, including dspy_pex.py, dspy_pex_gen.py, and pex_data_loader.py, contain hardcoded absolute file paths (e.g., /Users/mikhail/Cursor/Knowledge-Graph/data/). This exposes the local username and directory structure of the skill author's machine.\n- [INDIRECT_PROMPT_INJECTION] (LOW): The PEXGenerator module in dspy_pex_gen.py processes external medical data and interpolates it into prompts without adequate sanitization or boundary enforcement.\n
Ingestion points: pex_data_loader.py reads from CICM_SAQ.csv and SAQAtomic2.fixed.json.\n
Boundary markers: None; retrieved data is concatenated directly into the LLM context string.\n
Capability inventory: The skill utilizes dspy.Predict, dspy.ChainOfThought, and dspy.ProgramOfThought.\n
Sanitization: No escaping or validation is performed on the ingested medical content before prompt interpolation.\n- [EXTERNAL_DOWNLOADS] (LOW): In applications/building-rag-pipelines.md, a hardcoded IP address (http://20.102.90.50:2017) is provided for a hosted ColBERTv2 retrieval service, creating a dependency on an unauthenticated external network resource.

dspy