The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/headless_fanout.py to execute the claude binary. This is the primary functionality of the skill, used to launch subagents. The implementation uses argument lists (rather than a raw shell string), which prevents common shell injection vulnerabilities.
[DATA_EXFILTRATION]: The skill includes proactive measures to prevent sensitive data exposure. The headless_fanout.py script identifies and removes environment variables that might contain credentials (e.g., tokens, secrets, passwords) before launching child processes.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator for subagents, processing task lists that contain prompts. It provides security boundaries such as tool restrictions (--permission-mode dontAsk) and session isolation (--no-session-persistence) to minimize the risk of malicious instructions in the task data affecting the parent session or other subagents.

dispatching-subagents