The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructs the agent to execute tools with flags designed to bypass security guardrails and user confirmation prompts. Specifically, it recommends using codex --yolo and claude --permission-mode bypassPermissions. These configurations remove the human-in-the-loop requirement, allowing the AI to execute arbitrary code or system modifications without review.
[COMMAND_EXECUTION]: The instructions provide specific patterns for driving terminal multiplexers (cmux) to send commands to other agents. This includes methods to simulate keyboard input (send-key ... enter) and bypass terminal constraints, which grants the orchestrating agent direct, unmediated control over the host environment and other agent sessions.
[INDIRECT_PROMPT_INJECTION]: The orchestration workflow involves relaying instructions between different LLM-based agents (Pi, Codex, Claude, Hermes). This creates a significant attack surface where untrusted data processed by one agent can be interpolated into commands or prompts sent to another, potentially leading to a chain-reaction exploit across the agent cluster.
Ingestion points: Data received from the user or tool outputs intended for relay (SKILL.md).
Boundary markers: None specified in the instructions for separating relay data from agent instructions.
Capability inventory: Direct shell access via cmux, codex, pi, and claude (SKILL.md).
Sanitization: None provided; instructions focus on ensuring commands are not "broken" by shell escaping rather than sanitizing content.

delegating-to-agents