Wiring Codex Computer Use into CC
OpenAI Codex Mac App’s Computer Use, plugged into CC’s dispatch. user-ask queue, 90%+ eaten.
The orchestrator has been running for a while. Most of the flow auto-completes now, but one class of tasks never went away — the kind that needs me to drive a browser or a GUI App by hand. Register an account, grab a cookie, fill a canvas form, screenshot a logged-in state. I put those into a separate user-ask queue. CC sees a GUI-class task and drops it in there for me.
The problem is only I can consume user-ask. When I sleep, the queue piles up. The whole loop stalls until I wake.
This week I wired OpenAI’s Codex Mac App Computer Use into CC’s dispatch chain. When CC sees a user-ask that’s GUI-class, it pushes the task straight to Codex App via the codex:// URL scheme and hands over the screen. 90%+ of the queue items got eaten by Codex.
The wiring is simple. Codex App registers a deeplink. One shell line pushes a task across:
open "codex://new?path=$(pwd)&prompt=<url-encoded prompt>"
CC dispatch splits by task type — GUI ops go through codex://, pure CLI tasks still go to the original worker pane. Codex App spins up Computer Use on its own, runs everything inside its sandbox — can’t touch Terminal, can’t touch Codex itself, can’t escalate to admin. The safety boundary holds at the App level by default.
Red lines hard-coded in the prompt. Computer Use is a GUI agent — it sees an input box, it fills it. It can’t tell a real payment page from a fake one. So payments / credit cards / Apple Pay / 2FA / banking Apps are all banned. Hit a payment wall, kill the thread, bounce back to user-ask for me. Captcha same — Codex refuses reCAPTCHA Enterprise and Turnstile on its own and stops. These bounce back to me. Everything else it runs.
Not every GUI task should go through Computer Use. Rule of thumb: if the site has stable DOM, use Playwright. 100x cheaper, no drift. Canvas / Flutter / desktop Apps / shaky UI — that’s where Computer Use earns its keep. Codex starts losing precision past 20 steps. Long tasks get chopped short. Cross-App drags (Figma → Keynote) work but aren’t reliable.
Progress capture takes some thought. Computer Use’s execution shows up only inside Codex App’s GUI — screenshots, action log, live cursor — it doesn’t tee to file. Two ways for CC to get results back: have the Codex task write to a specified path on completion, or run codex app-server and listen for events over JSON-RPC. Day-to-day, the first one’s enough. Save the second for complex orchestration.
After wiring it in: CC no longer blocks on GUI tasks. The dispatch chain runs end-to-end. My consumption rate on user-ask was never going to catch CC’s production rate — now Codex handles the bulk and I only review what Codex can’t — payments, logins, Captcha — which were always going to be on me anyway.
Every new GUI proxy plugged in shrinks my manual surface by another notch.
Everything except payments is automated now.
把 Codex Computer Use 接进 CC
OpenAI Codex Mac App 的 Computer Use 接进 CC 的 dispatch。user-ask 队列被吃掉 90%+。
之前我那套 orchestrator 跑了一段时间,大部分流程已经能自动走完,但有一类任务一直甩不掉 — 必须手动操作浏览器或 GUI App 才能做的那种。注册账号、抓 cookie、填画布表单、截登录态。这部分我单独做了一个 user-ask 队列,CC 检测到 GUI 类任务就排进去,等我处理。
问题是 user-ask 队列只有我能消费。我睡觉的时候,它就堆在那里,整个 loop 卡住,等我醒来。
这周把 OpenAI Codex Mac App 的 Computer Use 接进了 CC 的 dispatch 链。CC 检测到 user-ask 是 GUI 类的,会直接通过 codex:// URL scheme 把任务推给 Codex App,让它接管屏幕去做。接进去之后,user-ask 队列里 90%+ 的条目直接被 Codex 吃掉了。
接法很简单。 Codex App 在系统里注册了 deeplink,一行 shell 就能把任务推过去:
open "codex://new?path=$(pwd)&prompt=<url-encoded prompt>"
CC dispatch 时按任务类型分流:涉及 GUI 操作的走 codex://,纯命令行还走原来的 worker pane。Codex App 接到任务会自己启动 Computer Use,全程在它的沙箱里跑 — 不能动 Terminal,不能动 Codex 自己,不能做 admin 提权,安全边界 App 自己就守住了。
红线写死在 prompt 里。 Computer Use 是 GUI 代理,看见输入框就填 — 没法判断真假付款页。所以付款 / 信用卡 / Apple Pay / 2FA / 银行 App 一律禁掉,撞到付款墙立刻 kill 掉 thread,退回 user-ask 让我手动处理。Captcha 同理 — Codex 主动拒绝过 reCAPTCHA Enterprise 和 Turnstile,撞到就停。这些边界丢回我,其余的它跑。
不是所有 GUI 任务都该上 Computer Use。 一个判断原则:DOM 稳的站用 Playwright,便宜 100 倍,没漂移。画布 / Flutter / 桌面 App / UI 乱跳的,才上 Computer Use。Codex 跑超过 20 步会开始掉精度,长任务要拆短;跨 App 拖拽(比如 Figma → Keynote)能跑,但不稳。
进度回流也要想清楚。Computer Use 的执行过程只在 Codex App 的 GUI 里看 — 截图、action log、实时光标 — 不 tee 到文件。CC 这边要拿结果,只有两条路:让 Codex 任务结束时把结果写到指定的 path,或者用 codex app-server 走 JSON-RPC 监听事件。日常用前者就够,复杂编排再上后者。
接进去之后,整体感受:CC 这边不再因为 GUI 类任务 block,dispatch 链是真的能一路跑到底了。我能消费 user-ask 的速度从来没追上过 CC 生产的速度,现在 Codex 帮我吃掉了大头,我只需要审 Codex 跑不动的那些 — 付款、登录、Captcha — 这些本来也就该我做。
每多接一块 GUI proxy 进去,我手动的范围就再缩一圈。
除了需要付款的事情,其他都自动化掉了。