📄 中文摘要
构建一个具有自主 AI 代理的代码 → 评审 → 测试管道是一个复杂的任务，尤其是需要确保流程的确定性。在经过两个月的探索，包括使用 Copilot 代理会话、构建自己的包装器 Protoagent、评估 Ralph Orchestrator 以及深入研究 OpenClaw 的内部机制后，发现 Lobster（OpenClaw 的工作流引擎）是一个理想的基础。然而，Lobster 缺乏循环支持，因此贡献了具有循环支持的子工作流步骤，使得完全确定性的多智能体管道成为可能。
如何在 OpenClaw 内部构建确定性多智能体开发管道（并为 Lobster 贡献缺失的部分）
📄 English Summary
How I Built a Deterministic Multi-Agent Dev Pipeline Inside OpenClaw (and Contributed a Missing Piece to Lobster)
Building a code → review → test pipeline with autonomous AI agents requires a deterministic orchestration, avoiding reliance on LLMs for flow decisions. After two months of exploring Copilot agent sessions, developing a custom wrapper called Protoagent, evaluating Ralph Orchestrator, and delving into the internals of OpenClaw, Lobster, OpenClaw's workflow engine, emerged as the right foundation. However, it lacked loop support. A contribution was made to Lobster by adding sub-workflow steps with loop capabilities, enabling fully deterministic multi-agent pipelines.