The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands to interface with external CLI tools (codex, claude, and tmux).- [COMMAND_EXECUTION]: The skill explicitly instructs the agent to use high-risk flags that bypass core security mechanisms:
The --dangerously-bypass-approvals-and-sandbox flag for the codex exec command removes isolation and user confirmation requirements.
The --dangerously-skip-permissions flag for the claude CLI disables authorization checks for the prompt being processed.- [COMMAND_EXECUTION]: The skill uses tmux send-keys to programmatically inject strings into a terminal session, which can lead to arbitrary command execution if the input is not strictly controlled.- [PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection by acting as a bridge that passes unvalidated prompts from one agent to another.
Ingestion points: The <prompt> placeholder within the command templates in SKILL.md.
Boundary markers: No specific delimiters or "ignore instructions" headers are used to wrap the input data.
Capability inventory: The tools being invoked have the ability to read/write files and execute system commands.
Sanitization: No sanitization or escaping of the prompt string is implemented before execution.

confer