The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to interact with Git and a platform-specific CLI tool called codex. These operations are central to its purpose of managing code changes and executing automated reviews. The scripts/git_isolation.py script specifically implements checks to ensure only task-related paths are committed, providing a layer of protection against accidental data modification.
[DYNAMIC_EXECUTION]: The scripts/run_vision_review.py script facilitates independent code reviews by spawning a new process via the codex exec command. This process is configured with an ephemeral, read-only sandbox, which is a security control that limits the sub-agent's ability to make persistent or unauthorized changes.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to process external CSV and Markdown files (approved specs), it is naturally exposed to indirect prompt injection. However, the instructions include extensive 'Claim-Evidence Alignment' checks and 'Outcome Contracts' specifically designed to prevent the agent from being misled by untrusted data or hallucinating completion results without actual evidence.

mission-csv-execute