orchestration
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill defines a framework where sub-agents are spawned to execute system commands, tools, and CLI utilities according to the task decomposition. It mandates a least-privilege approach, restricting tool access for each phase sub-agent to only what is strictly necessary.
- [INDIRECT_PROMPT_INJECTION]: The orchestration pattern involves processing user-supplied instructions and data from external phases, creating an entry point for indirect prompt injection. The skill implements multiple defensive layers, including path-based isolation (sub-agents read specific files rather than receiving content in prompts), context-resetting role preambles, and mandatory user-judged review halts before acting on findings.
- Ingestion points: User task strings and data documents generated by intermediate phases stored in the run directory (e.g., .agents/orchestration//investigate/).
- Boundary markers: Explicit role-boundary preambles defined in references/orchestration_template.md that reset agent state and establish ownership of work.
- Capability inventory: File system access (read/write), shell command execution, and sub-agent spawning/management.
- Sanitization: Mechanical validation of phase outputs against defined schemas and a required human-in-the-loop review step for critical findings.
Audit Metadata