skills/d3lm/ai-lab/agent-orchestrate/Gen Agent Trust Hub

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-exec on macOS and bwrap on 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 dlx as 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
Risk Level
SAFE
Analyzed
Nov 17, 2026, 07:11 PM