The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly aim to bypass safety filters and security constraints in downstream agents.
Evidence: Recommends using claude --permission-mode bypassPermissions and codex --yolo to ensure autonomous operation by removing built-in safety prompts.
[COMMAND_EXECUTION]: The skill provides patterns for executing and automating shell commands that bypass interactive user confirmation.
Evidence: The instructions for using cmux send followed by cmux send-key enter are designed to automate command input in a way that avoids manual review.
Evidence: The use of the --yolo flag with the Codex CLI is a documented method to suppress confirmation prompts for generated code execution.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by delegating instructions across agent boundaries without mandatory sanitization.
Ingestion points: Processes instructions passed via cmux send to target agent surfaces.
Boundary markers: Recommends using double quotes to wrap prompts but lacks explicit instruction to ignore embedded malicious content in the relayed payload.
Capability inventory: Possesses full shell execution capability through the orchestrated agents (Codex, Claude, Pi, Hermes).
Sanitization: Lacks sanitization beyond simple shell-escaping recommendations.

delegating-to-agents