The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs pytest and pytest-xdist using package managers (uv or pip) if they are missing in the environment. These are standard, well-known testing frameworks for Python development.
[COMMAND_EXECUTION]: The skill executes shell commands to interact with the MLflow CLI (mlflow traces get) and the pytest runner (pytest). These commands are necessary for the skill's workflow of diagnosing and verifying agent behavior.
[DATA_EXFILTRATION]: The skill retrieves trace data from an external MLflow tracking server via MLFLOW_TRACKING_URI. This communication is central to the skill's purpose of analyzing and fixing agent failures recorded in MLflow.
[PROMPT_INJECTION]: The skill processes trace data which includes untrusted content from previous LLM interactions (ingestion point: mlflow traces get in SKILL.md). While the skill possesses capabilities including command execution (pytest) and package installation, it uses this data for diagnostic purposes. As an indirect prompt injection surface, this represents a low risk factor inherent to agent debugging tools.
[REMOTE_CODE_EXECUTION]: The skill dynamically generates Python test scripts containing mlflow.genai.evaluate assertions and executes them using the pytest runner. This behavior is the primary mechanism of the skill to ensure fixes are verified through a regression test suite.

fix-agent-issue