The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill functions as a coordinator that distributes a user-defined task to multiple parallel runners using the delegate_task tool.
Ingestion points: The brief constructed in Phase B (Fan out) serves as the primary injection vector, as it includes the task description and grounding data which may contain untrusted content. (Found in SKILL.md)
Boundary markers: The skill notes that sub-agents receive only the standalone brief, but it does not specify explicit delimiting or 'ignore' instructions to prevent the sub-agents from following commands embedded in the task data. (Found in SKILL.md)
Capability inventory: The orchestration process utilizes delegate_task for agent spawning and t3_thread_launch for environment isolation (git worktrees or /tmp paths). (Found in SKILL.md)
Sanitization: The process includes a mandatory 'Verify' phase (Phase F) where the synthesized output is scrutinized to ensure it meets success criteria, providing a structural check against unintended sub-agent behaviors. (Found in SKILL.md)

arena