The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.Popen to launch the local grok executable and its own background worker processes to manage ACP (Agent Client Protocol) sessions. Evidence is found in scripts/acp_client.py and scripts/agent_task_runtime.py where subprocesses are created to handle the external agent's lifecycle.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input through task prompts and handles responses from the external Grok agent, creating a surface for potential instruction injection. 1. Ingestion points: Prompts provided via --prompt, --prompt-file, or stdin in scripts/grok.py. 2. Boundary markers: Prompts are encapsulated within JSON-RPC messages sent to the Grok agent. 3. Capability inventory: Execution of subprocesses and local file system access for logging and state management. 4. Sanitization: Standard JSON encoding is used for message transport between processes.
[DYNAMIC_EXECUTION]: The skill utilizes ctypes to interface with the Windows kernel32 library to manage process groups and ensure reliable cleanup of the background agent processes. Evidence is found in scripts/agent_task_runtime.py where ctypes.WinDLL is used to call CreateJobObjectW and AssignProcessToJobObject.

grok-cli