The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads assignment data from an external file (phase-2-assignments.md) and passes this content directly to subagents using the Agent tool. If these assignments contain malicious instructions, they could influence the behavior of the recipient agents. The skill provides a partial mitigation by instructing subagents to ignore conversation history and focus only on the provided prompt, but no technical sanitization or strict boundary delimiters are used.
Ingestion points: The skill reads phase-2-assignments.md (via the Read tool) to facilitate both the pingpong-detector check and the final dispatch.
Boundary markers: The instructions in Step 3 explicitly tell subagents to read only the provided prompt and avoid implicit conversation context, which acts as a logical instruction-based boundary but lacks technical enforcement.
Capability inventory: The skill utilizes Write and Edit for log management, Bash for directory preparation (returns/), and the Agent tool to invoke subagents.
Sanitization: No validation, escaping, or filtering of the content from phase-2-assignments.md is performed before it is sent to the Agent tool.

coliseum-dispatch