The Agent Skills Directory

Indirect Prompt Injection (LOW): The orchestration logic relies on a main agent reading data directly from task descriptions modified by subtask agents, which is an untrusted data source.\n
Ingestion points: The get_task tool retrieves the description field in the polling loop defined in SKILL.md.\n
Boundary markers: Absent. Instructions are appended via simple string concatenation and the agent checks for status using string matching (e.g., ## Status in result.description) without delimiters.\n
Capability inventory: The skill invokes create_task, update_task, and start_workspace_session which can manage and execute logic across multiple repositories.\n
Sanitization: No sanitization or escaping of the subtask-provided results is performed before the main agent processes the input.

subtask-orchestration