multi-agent-collaboration
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to use standard development tools for workflow management and isolation.\n
- Employs git worktree to isolate file system changes when multiple agents work in parallel.\n
- Executes local Python scripts (scripts/workflow.py) and testing modules (unittest) for task validation and integration.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process and integrate data from multiple sub-agents, which creates an inherent surface for indirect instructions.\n
- Ingestion points: Processes worker results (summarized in worker_result_v1), delegation plans (agent-delegation-plan.json), and shared handoff capsules.\n
- Boundary markers: Utilizes neutral briefs, handoff capsules, and predefined stable contracts to maintain boundaries between agent contexts.\n
- Capability inventory: The lead agent manages execution via Git commands, local script calls, and platform-specific execution runners (codex exec, claude -p).\n
- Sanitization: Implements path normalization and alias validation to prevent scope bypass; mandates result verification against gate policies before integration.
Audit Metadata