The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes Python scripts (new_workflow.py, collect_results.py, verify_workflow.py) for scaffolding and managing workflow artifacts. These scripts operate locally on files within a .workflow/ directory. The analysis shows these are utility scripts intended for project structure management and do not execute arbitrary shell commands or external code.
[PRIVILEGE_ESCALATION]: The skill explicitly warns against and requires approval for destructive or high-privilege operations. It includes comprehensive 'Risk Gates' documentation (references/risk-gates.md) that instructs the agent to ask for permission before performing sensitive actions like deleting files, modifying external systems, or accessing credentials.
[DATA_EXFILTRATION]: The skill operates entirely on local repository files. Network operations are not used in the provided scripts. The integration and reporting features only aggregate local markdown files into a summary report within the same project directory.
[PROMPT_INJECTION]: The skill uses clear trigger boundaries and decision rules to prevent misuse. It explicitly defines negative triggers (e.g., ordinary code reviews or small bug fixes) where the orchestration logic should not be invoked. It also includes instructions for the agent to maintain its original objective when spawning subtasks.

codex-dynamic-workflows