The Agent Skills Directory

[COMMAND_EXECUTION]:
The framework is designed to execute Python, SQL, and YAML code blocks extracted from AI agent responses on Databricks clusters and warehouses to verify their functionality.
This logic is implemented via MCP tools such as mcp__databricks__execute_databricks_command and mcp__databricks__execute_sql in src/skill_test/grp/executor.py.
Utility scripts like scripts/add_example.py use system commands (pbpaste, xclip) to interact with the system clipboard for developer convenience.
[EXTERNAL_DOWNLOADS]:
The optimization and evaluation components (optimize.py, mlflow_eval.py) communicate with external LLM provider endpoints, including Databricks Model Serving, OpenAI, and Anthropic.
It also interacts with MLflow tracking servers to log and retrieve evaluation metrics and session traces.
[PROMPT_INJECTION]:
The skill exhibits an indirect prompt injection surface because it ingests responses generated by other skills and executes the embedded code blocks during verification.
Ingestion points: The interactive function in src/skill_test/cli/commands.py receives a response string containing markdown code blocks.
Boundary markers: Relies on standard markdown triple-backtick delimiters (e.g., ```python) to identify code segments.
Capability inventory: Includes high-privilege operations such as arbitrary code execution on compute resources and file uploads to Unity Catalog volumes.
Sanitization: Performs basic syntax validation but does not sanitize the logical content of the generated code before execution.

skill-test