The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to utilize specific flags designed to suppress security guardrails. Specifically, it recommends using claude --permission-mode bypassPermissions and codex --yolo. These configurations disable user-in-the-loop confirmation for sensitive operations like file system changes and arbitrary shell command execution.
[PROMPT_INJECTION]: The skill documents a workflow that is highly susceptible to indirect prompt injection via external file ingestion.
Ingestion points: The instructions recommend writing long tasks to external files and then commanding a sub-agent to execute them (e.g., read /tmp/task.md and follow it).
Boundary markers: There are no provided instructions or best practices for using delimiters or 'ignore' instructions to prevent the delegated agent from obeying malicious commands embedded in the task file.
Capability inventory: The skill grants access to powerful execution environments (Codex CLI, Claude Code, and cmux) with the ability to perform persistent shell operations and remote VPS access.
Sanitization: There is no recommendation for validating, sanitizing, or escaping the content of the task files before they are processed by the delegated agent.
[COMMAND_EXECUTION]: The orchestration logic involves piping strings directly into terminal sessions via cmux send. The skill mentions manual handling of escaping characters (e.g., avoiding \" because it causes 'unexpected EOF' in bash), which suggests that raw, potentially unvalidated strings are being executed in a shell environment, presenting a command injection risk.

delegating-to-agents