把整本书压成一句:Codex 的问题从来不是「它能不能写这段代码」,是「有多少活可以不守着做完 —— 而你还能确认它做对了」。聊天框是它最小的样子;四个面、云端沙箱、exec 与 CI、验收工作流把它撑成一个人指挥的异步车间。剩下的,是知道那条边界画在哪 —— 哪些活陪跑、哪些活上云、哪一步切走 —— 然后让它在你不看着的地方把活干完,而你只做机器替不了的两件事:决定派什么,以及确认它真做成了。The whole book in one line: the question was never 'can Codex write this code' — it's 'how much work gets finished without you watching, while you can still confirm it was done right.' The chat box is its smallest form; four fronts, cloud sandboxes, exec and CI, and a review workflow stretch it into an async workshop directed by one person. What remains is knowing where the boundary sits — which work to pair on, which to send up, where to switch away — then letting it finish where you aren't watching, while you do the two things a machine can't: decide what to send, and confirm it truly got done.
7 分钟 · 初稿 2026.077 Min · Drafted 2026.07
把十章压成一句:Codex 的问题从来不是「它能不能写这段代码」,是「有多少活可以不守着做完 —— 而你还能确认它做对了」。ChatGPT 里那个会写代码的模式,是它最小的样子;你付费买到的,是一个四面开工、能把一批活派上云再收回 PR 的异步编码 agent。Ten chapters in one line: the question was never 'can Codex write this code' — it's 'how much work gets finished without you watching, while you can still confirm it got it right.' The code mode inside ChatGPT is its smallest form; what you paid for is an async coding agent on four fronts that sends a batch to the cloud and collects it back as PRs.
— I
回扣:四面与一台引擎The Recap: Four Fronts, One Engine.
回扣一遍这台机器怎么拼起来的。四个面各有主场 —— IDE 陪跑、CLI 进脚本、应用切本地与云、云端批量并行(第 2 章);底下是同一台引擎,AGENTS.md、config、skills、MCP 走到哪跟到哪(第 5、6 章)。1注 1Note 1OpenAI Codex 文档 · Quickstart / Codex cloud —— Codex 的四个入口(ChatGPT 桌面应用与网页、CLI、IDE 扩展、云端)连着同一台引擎,AGENTS.md、config、skills、MCP 跨入口通用;云端把活委派给隔离环境、跑完交回可开 PR 的 diff。截至 2026-07-17。OpenAI Codex Docs · Quickstart / Codex cloud — Codex's four entrances (the ChatGPT desktop app and web, the CLI, the IDE extension, the cloud) are wired to one engine, with AGENTS.md, config, skills, and MCP carrying across; the cloud delegates work to isolated environments and returns a diff you can open as a PR. As of 2026-07-17.第一个循环在 CLI 里跑通:说清目标、看计划、放它动手、看 diff(第 3 章)。放权靠两个旋钮加规则,再加一个替你把关的 reviewer(第 4 章)。差异点在云端:活离开你的机器、进隔离容器、回来是 PR,还能一次派一批(第 7 章);无人值守把这个推到极限 —— exec、SDK、CI、定时任务(第 8 章)。而托住这一切的是验收:检查、复审、回归,凭据在手才算完(第 9 章)。选对节奏,比选贵的档更省(第 10 章)。1注 1Note 1OpenAI Codex 文档 · Quickstart / Codex cloud —— Codex 的四个入口(ChatGPT 桌面应用与网页、CLI、IDE 扩展、云端)连着同一台引擎,AGENTS.md、config、skills、MCP 跨入口通用;云端把活委派给隔离环境、跑完交回可开 PR 的 diff。截至 2026-07-17。OpenAI Codex Docs · Quickstart / Codex cloud — Codex's four entrances (the ChatGPT desktop app and web, the CLI, the IDE extension, the cloud) are wired to one engine, with AGENTS.md, config, skills, and MCP carrying across; the cloud delegates work to isolated environments and returns a diff you can open as a PR. As of 2026-07-17.Recap how the machine assembles. Four fronts, each with its turf — the IDE pairs, the CLI goes into scripts, the app switches local and cloud, the cloud runs batches in parallel (chapter 2); underneath is one engine, with AGENTS.md, config, skills, and MCP following you everywhere (chapters 5, 6).1注 1Note 1OpenAI Codex 文档 · Quickstart / Codex cloud —— Codex 的四个入口(ChatGPT 桌面应用与网页、CLI、IDE 扩展、云端)连着同一台引擎,AGENTS.md、config、skills、MCP 跨入口通用;云端把活委派给隔离环境、跑完交回可开 PR 的 diff。截至 2026-07-17。OpenAI Codex Docs · Quickstart / Codex cloud — Codex's four entrances (the ChatGPT desktop app and web, the CLI, the IDE extension, the cloud) are wired to one engine, with AGENTS.md, config, skills, and MCP carrying across; the cloud delegates work to isolated environments and returns a diff you can open as a PR. As of 2026-07-17. The first loop closes in the CLI: state the goal, see the plan, let it act, read the diff (chapter 3). Permissions run on two dials plus rules, plus a reviewer that gates for you (chapter 4). The differentiator is the cloud: work leaves your machine, enters an isolated container, comes back a PR, and a whole batch at once (chapter 7); unattended pushes it to the limit — exec, the SDK, CI, scheduled tasks (chapter 8). And holding all of it up is review: checks, re-review, regression, done only with evidence in hand (chapter 9). Picking the right cadence saves more than the pricier tier (chapter 10).1注 1Note 1OpenAI Codex 文档 · Quickstart / Codex cloud —— Codex 的四个入口(ChatGPT 桌面应用与网页、CLI、IDE 扩展、云端)连着同一台引擎,AGENTS.md、config、skills、MCP 跨入口通用;云端把活委派给隔离环境、跑完交回可开 PR 的 diff。截至 2026-07-17。OpenAI Codex Docs · Quickstart / Codex cloud — Codex's four entrances (the ChatGPT desktop app and web, the CLI, the IDE extension, the cloud) are wired to one engine, with AGENTS.md, config, skills, and MCP carrying across; the cloud delegates work to isolated environments and returns a diff you can open as a PR. As of 2026-07-17.
— II
你只做机器替不了的两件事You Do Only What the Machine Can't.
四面成形之后,你的工作只剩两件,恰好是机器替不了的两件:决定派什么,和确认它真做成了。前者是品味和取舍 —— 哪件活值得派、派到哪一面、边界画在哪;后者是这本书反复回来的那句 —— 给它一个能自己跑的检查,要证据,不要断言。2注 2Note 2OpenAI Codex 文档 · Best practices —— 别停在让它改;给它一个它能自己跑的检查、要证据不要断言,让它写测试、跑检查、审自己的 diff,再由你接受。原话:「在 OpenAI,Codex 审 100% 的 PR。」截至 2026-07-17。OpenAI Codex Docs · Best practices — don't stop at asking it to change; give it a check it can run, evidence over assertions, have it write tests, run checks, review its own diff, then you accept. Verbatim: 'At OpenAI, Codex reviews 100% of PRs.' As of 2026-07-17.派活的人已经太多,这两件事上的功夫,才是四面开工的时代里你真正的手艺。Once the four fronts take shape, your work reduces to two things — precisely the two a machine can't do: deciding what to send, and confirming it was truly done. The first is taste and trade-off — which job is worth sending, to which front, where the boundary sits; the second is the line this book keeps returning to — give it a check it can run, evidence over assertions.2注 2Note 2OpenAI Codex 文档 · Best practices —— 别停在让它改;给它一个它能自己跑的检查、要证据不要断言,让它写测试、跑检查、审自己的 diff,再由你接受。原话:「在 OpenAI,Codex 审 100% 的 PR。」截至 2026-07-17。OpenAI Codex Docs · Best practices — don't stop at asking it to change; give it a check it can run, evidence over assertions, have it write tests, run checks, review its own diff, then you accept. Verbatim: 'At OpenAI, Codex reviews 100% of PRs.' As of 2026-07-17. Delegators are already plentiful; craft in these two is what's actually yours in the era of four fronts.
— III
边界,和一把不过期的尺The Boundary, and a Ruler That Doesn't Expire.
剩下的是画边界。哪些活自己盯:不可逆的、高风险的、要品味的。哪些活派出去:可验证的、可并行的、重复的 —— 这些正是云端的甜点位。哪一步切走:该在编辑器里陪着改的、长在 GitHub 流程里的、以终端本身为中心的,入口跟着活走(第 10 章)。四面不是让你四处忙,是让你在对的面上派对的活。What remains is drawing boundaries. Watch yourself: the irreversible, the high-stakes, the matters of taste. Send out: the verifiable, the parallel, the repeated — exactly the cloud's sweet spot. Switch away: what wants an editor at your side, what lives in GitHub flow, what centers on the terminal itself — entrances follow the work (chapter 10). Four fronts isn't for keeping you busy everywhere; it's for sending the right work to the right front.最后说句实话:这本书会过期,而且很快 —— Codex 以周为单位在变,模型名、命令、档位,书里每个数字都标了核对日期,过期请以官方文档为准。但这把尺子不过期:拿到任何一个新特性、新入口,还是那三问 —— 它省什么,什么活用它,怎么验收。问完这三问,你就不需要下一本教程了。One honest closing note: this book will expire, and quickly — Codex changes week by week, model names, commands, tiers, and every number here carries its verification date; when one expires, trust the official docs. But the ruler doesn't expire: hand it any new feature or entrance and ask the same three questions — what does it save, which work is it for, how do I verify it. Once you ask those three, you won't need the next tutorial.
一个人,四面开工:
你派活,你验收,其余交给它们.
One person, four fronts:
you delegate, you verify — the rest is theirs.