The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a multi-agent orchestration pattern that ingests external data (task requirements and implementation reports) into subagent contexts. Ingestion points include the requirements, context, and patch fields in the implementer and reviewer templates (references/implementor-prompt.md and references/batch-reviewer-prompt.md). While this creates a potential surface for instructions hidden in data, the skill mandates robust structural mitigations: restricting subagents to a predefined list of owned files, capturing path-scoped patches to reject unauthorized changes outside the assigned scope, and utilizing a separate read-only agent for a final quality and compliance review. Although explicit sanitization is not detailed, the capability inventory for subagents is strictly constrained to specific files and validation commands, and boundary markers are implemented via template sections, significantly limiting the impact of malicious requirements data.

code-subagents