Reddit 用户总结 Codex 隔夜运行验收法
一名在 Apple Silicon Mac 上运行 Codex workers 的 Reddit 用户,用可验证结果判断任务是否完成:代码任务需看到测试通过,研究任务则要有可读取的保存文件。该用户会把超时或被中断的检查标为“未验证”,并把工作拆成带独立检查的小任务,同时指出隔夜重试会消耗 ChatGPT 套餐额度。重启或崩溃后如何区分已完成与中途终止的任务,他仍未解决。
这则实践分享给出了核验无人值守 Codex 任务的实用方法,也明确暴露了任务恢复与状态判定仍存在的缺口。
I run Codex workers on my own Apple Silicon Mac, signed in with a ChatGPT plan, and the hard part of leaving them alone was never the work. It was the morning: a quiet screen and no idea what actually finished.
What I've settled on so far:
- Done is something I can open. A code task counts when I see the real test output (one run: 6 tests, OK, exit 0, 55 seconds). A research task counts when there's a saved file I can read, not a message saying it's done.
- A third state between done and failed. Someone in another thread called it unverified, and it stuck. If a check timed out or got cut off, it lands in front of me as unverified instead of quietly retrying.
- Small tasks, each with its own check. A bad turn costs one task, not the whole night.
- The boring limits are the real ones. The computer has to stay on, and every task counts toward my ChatGPT plan's limits, so a loop that keeps retrying overnight is spending plan, not just time.
The line I keep coming back to: silence is not a status.
Honest limits: one Mac, a handful of runs, not a study.
The part I haven't solved: after a restart or a crash, telling a task that finished from one that was cut off halfway. Logs tell me what started, not what ended.
What do you actually let an agent run overnight, and how do you tell finished from cut off the next morning?