agent-orchestrate
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill manages the execution of external agent CLIs (Claude, Codex, Cursor) and user-defined verification commands. To mitigate risks from agent-generated code, the orchestrator implements mandatory sandboxing for verification and integration checks using
sandbox-execon macOS andbwrapon Linux. These sandboxes restrict file system access and disable common mechanisms for process escape. - [DYNAMIC_EXECUTION]: The orchestrator spawns subprocesses for sub-agents and tests. It uses a robust process management system that verifies process identity (PID, start time, and boot ID) before sending signals, preventing accidental signaling of unrelated processes due to PID reuse.
- [EXTERNAL_DOWNLOADS]: The skill resolves and executes official packages from Anthropic and OpenAI using
pnpm dlxas a fallback. These packages originate from well-known technology vendors and are explicitly hardcoded within the adapter modules, minimizing supply chain risks. - [INDIRECT_PROMPT_INJECTION]: The skill processes data that may contain instructions, such as repository files or sub-agent outputs. It mitigates injection risks by using structured result protocols, JSON Schema validation, and isolating agent execution in dedicated git worktrees with strict permission controls.
Audit Metadata