The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a --parallel-tests flag that allows the agent to execute shell commands to verify fixes during the review process. This functionality is heavily protected by running the commands in an isolated temporary root with a sanitized environment that strips potentially dangerous variables (e.g., NODE_OPTIONS, LD_PRELOAD, DYLD_INSERT_LIBRARIES).
[DATA_EXFILTRATION]: The tool contains an extensive secret detection engine that identifies hardcoded credentials, JWTs, and private keys in the code being reviewed. It automatically blocks review bundles if sensitive files (e.g., .env, .ssh/id_rsa) or secret-like patterns are detected, preventing accidental exposure of sensitive data to external AI providers.
[REMOTE_CODE_EXECUTION]: The skill invokes external AI engines via CLI tools but enforces strict isolation for each. It explicitly refuses engines that lack sufficient isolation (e.g., Cursor, Copilot) and ensures that supported engines (Codex, Claude, Pi) run in empty, temporary workspaces without access to the host filesystem tools or ignored repository metadata.
[SAFE]: The 'malicious' patterns identified in the test fixtures (e.g., shell injection examples in test-review-harness.py) are used solely for security calibration and unit testing to ensure the tool can correctly identify and block real vulnerabilities.

autoreview