The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates an architectural pattern where a controller agent processes and integrates output ("handoffs") from worker subagents. This creates a surface for indirect prompt injection if subagents are tasked with analyzing untrusted data (such as monorepo mapping or reconnaissance) and their findings are subsequently incorporated into the controller's decision-making flow.
Ingestion points: The integration of worker summaries and artifacts into the controller's context occurs during the "After handoff" phase described in SKILL.md and the templates in references/handoff.md.
Boundary markers: The skill provides structured Markdown templates (Status, Summary, Artifacts) in references/handoff.md to delimit worker responses, which acts as a partial mitigation but does not provide complete isolation from malicious content in the worker's report.
Capability inventory: The controller agent has access to tool execution and file system modification (e.g., implementing bug fixes or running development tools), making it a potential target for instructions smuggled through worker output.
Sanitization: The instructions lack specific guidance on validating or sanitizing worker-provided content before it is incorporated into the main session's goal or plan.

using-subagent