The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local helper scripts using subprocess.run for scaffolding tasks. These operations are restricted to internal files within the repository and use validated, normalized inputs, preventing arbitrary execution.- [SAFE]: Prompt injection patterns identified in the evaluation files are defensive test cases designed to verify the agent's safety filters, not active exploits.- [SAFE]: Robust path validation logic in plugin_builder.py prevents directory traversal attacks by enforcing relative paths and verifying they remain within the intended plugin root.- [SAFE]: The skill processes external plugin sources using an 'assume-malicious' posture, requiring contract validation and manual review before any files are incorporated into the environment.

plugin-builder