以低接触式规划和端口预留运行 Codex
Run Codex with low-touch planning and port reservation
Codex 完成任务。一位评论者让 Codex 在人类规划和审查下主要自行执行，包括一个端口经纪人负责预留和注入端口.
做了什么
The commenter describes a split workflow: ChatGPT handled planning, strategy, versioning discipline, and writing the next prompt, while Codex did the implementation, with outputs manually copied back and forth between the two in extended-thinking chats. In a similar setup, another commenter added a small tool called portbroker so an agent could reserve a free port, log it, and inject it into $PORT automatically, reducing manual environment handling during parallel runs.
效果如何
They say this cut down prompt drift and made multi-agent or parallel local development more stable and predictable. The main drawback mentioned is that the handoff is still manual, with long threads causing slowdown and occasional need to start fresh sessions.
值得注意
The practical catch is the copy-paste loop between ChatGPT and Codex, plus needing to reset when threads get "sludge" from growing too long.
试试这个 Prompt
I've written a plan for [task] in [repo path]: [paste or point to plan file]. Carry it out mostly on your own, and before starting any dev server or test that needs a network port, pick a free port, reserve it, and pass it in through [env var or config setting] so it doesn't clash with other running jobs. When done, summarize what you changed and which checks passed so I can review, and ask me before pushing or deleting anything.