The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes outputs from multiple subagents (candidates and judges) in Phase D and E, which constitutes an attack surface where a subagent's output could attempt to influence the parent agent's decision-making process.
Ingestion points: Processes subagent artifacts and rationales in SKILL.md Phase D and Phase E; reads model configuration from pstack-models.md.
Boundary markers: Explicitly mandates that child agents receive only the task brief and path to shared grounding, excluding the parent's full conversation context. It also requires the use of a concrete rubric for evaluation rather than holistic feel.
Capability inventory: Utilizes platform tools delegate_task, task_status, and task_cancel. Performs file system operations for creating candidate output locations in /tmp or git worktrees.
Sanitization: Relies on structured evaluation (rubric) and synthesis logic to validate subagent outputs before incorporation.

arena