The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script validate_api_claims.py implements a validation harness that extracts Python code blocks from Markdown documentation files (such as references/verified-skeleton.md) and executes them using the exec() function. The script compiles documentation strings at runtime and runs them within a controlled environment to verify API signatures and return shapes.
[DYNAMIC_EXECUTION]: The validate_api_claims.py script utilizes importlib.import_module() to dynamically load decomposition worker modules at runtime based on entries in a registry.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process user-provided DataFrames across multiple workflows (Econ, Epi, and ML-Causal). It also incorporates LLM-assisted features like sp.llm_dag_propose for causal structure discovery, which creates a surface where malicious data content or untrusted LLM outputs could influence the agent's behavior.
Ingestion points: DataFrames are loaded and processed in pipeline-econ.md, pipeline-epi.md, and pipeline-ml-causal.md.
Boundary markers: No explicit boundary markers or 'ignore' instructions for data ingestion are documented.
Capability inventory: The skill possesses capabilities for local file writing (Word, Excel, LaTeX), shell command execution (via the validation script), and external LLM interactions.
Sanitization: There is no evidence of explicit sanitization or filtering of input data before it is processed by the statistical estimators or passed to LLM-based discovery tools.

statspai-analysis