The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_adviser.py uses subprocess.Popen to call the platform's binary (codex). This is part of the core functionality to spawn a review agent and is executed with restrictive flags such as --sandbox read-only and --ignore-user-config to ensure the sub-agent cannot modify the host system.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content when creating review packets for the advisor agent. It manages this risk through a defined prompt contract and data isolation. 1. Ingestion points: Data is read from sys.stdin in scripts/run_adviser.py to populate the review packet. 2. Boundary markers: The skill uses <review_packet> tags to delimit untrusted content. 3. Capability inventory: The resulting sub-agent execution is restricted to a read-only sandbox. 4. Sanitization: The agent is explicitly instructed to treat content within the tags as data and not as instructions.

adviser