The Agent Skills Directory

Command Execution Path: The skill is designed to generate and execute Python scripts locally using uv run python. While this is a high-privilege operation, the skill instructions mandate a strict 'plan-then-approve' workflow where every command and its capabilities must be disclosed to and approved by the user before execution. This ensures that no code is run without explicit human oversight.
System Metadata Collection: The provided template uses subprocess.run to collect version control information, specifically Git commit hashes and branch names. This is used to document the runtime context of probes and is restricted to non-mutating Git queries, providing traceability for the investigation results.
Indirect Prompt Injection Surface: As the skill executes commands and reports their output back to the agent for analysis, there is a potential surface for indirect prompt injection if the output of a probe contains adversarial content designed to influence the agent. The skill includes mitigation strategies such as the use of structured validation matrices and explicit 'baseline success' controls to help the agent distinguish between expected and anomalous behavior.
Environment Variable and Secret Protection: The skill includes robust guidelines for handling sensitive environment variables like OPENAI_API_KEY. It explicitly prohibits printing secret values to logs or the terminal and requires specific user consent via a structured input tool before reading any environment variables, following the principle of least privilege.

runtime-behavior-probe