正在打开每日关注
正在打开每日关注
通过独立 worktree、心跳与循环,让 Codex 夜间并行处理、审查并修复多个 PR。
My new workflow for working while I sleep: 1. Codex can create threads in worktrees, set up heartbeats to check in, and run loops. 2. Ask GPT-5.5 Pro to help you redefine and improve the prompt (huge game changer). 3. Ask Codex to review PRs and issues, or build new features in parallel, each in its own dedicated worktree. 4. Codex-bot: use Codex-bot on GitHub to flag issues and run a loop that fixes them all until you're clear. The prompt isn't just the request anymore It's the whole workflow. Ran it overnight, woke up to merged-ready PRs. Game changer.
I've discovered a new type type of workflow that lets me have work done in the morning, while I sleep: Loops! This is my first type of video made this way, where I bring you guys into how I do my work and how I build things! Hope you guys like it! https://t.co/8jizHlzfe7
是什么 这是一条“无人值守研发”实践总结:不是一次性让模型写代码,而是把任务编排成可持续运行的工作流。先用 GPT-5.5 Pro 重写目标与约束,再让 Codex 为每个 PR 建立独立 worktree/thread,分别开发、审查和推送;GitHub 上的 Codex-bot 继续发现问题并循环修复,直到 PR 清晰、可合并。配图中的四个 PR、隔离规则与“Goal complete”状态,和叙述相互印证。
核心设计 价值在隔离、反馈和闭环:每个 PR 绑定自己的分支、文件、上下文和线程,减少并行代理互相污染;心跳与循环把请求变成长任务;变更数、分支、推送和总体状态则提供早晨验收所需的可观测性。这里的 prompt 已经是目标、边界、权限、检查频率和完成条件的组合。
关键洞察 效率提升主要来自任务编排,而不只是模型能力。它适合依赖明确、可验证、可拆分的 PR;并行度越高,隔离规则、日志、失败恢复和合并门禁越重要。
边界与风险 截图只是成功案例,不能证明所有任务都能安全自动合并。复杂架构改动、共享配置、隐含产品判断和高风险发布仍需人工审查;循环修复也可能放大错误。应把测试、停止条件和人工审批写进流程。