The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes heavy use of terminal commands to drive the cmux CLI, including subcommands like cmux workspace create, cmux send, and cmux hooks setup. It also explicitly instructs the agent to launch other agents (Codex, Claude, Pi) with flags that bypass standard security permissions and sandboxes (e.g., --dangerously-bypass-approvals-and-sandbox, --dangerously-skip-permissions).
[CREDENTIALS_UNSAFE]: The skill facilitates the injection of environment variables from .env files into workspaces using the --env-file flag. While it includes instructions to mask these values and avoid reading them directly, the mechanism itself handles sensitive API keys (e.g., ANTHROPIC_API_KEY, OPENROUTER_API_KEY).
[PROMPT_INJECTION]: The skill processes task descriptions and orchestrates multiple agents. It is susceptible to indirect prompt injection (Category 8) where data processed by one agent inside a cmux surface could contain malicious instructions that influence the orchestrating agent's next actions.
Ingestion points: Data is read from terminal surfaces using cmux read-screen (SKILL.md).
Boundary markers: None explicitly defined for isolating agent output from orchestrator instructions.
Capability inventory: Full shell access via cmux, capability to spawn/kill processes, and network-enabled agents.
Sanitization: None described; the skill relies on the agent to 'read back to verify'.

cmux