重新搬回的“椅子”
过去，两个程序员想一起解决同一个问题，可以把椅子挪到同一块屏幕前。
一个人负责操作键盘，另一个人负责审视思路；遇到分歧，就停下来讨论。屏幕上的代码是两个人共同产出的成果，问题的来龙去脉，也在这一问一答中逐渐对齐。
这种协作的方法，就叫做“结对编程”。
如今，AI 俨然已经成了坐在我们屏幕另一侧的新搭档。
我们向它描述需求，让它阅读代码、分析问题，并尝试修改。在一轮轮对话之后，那条会话里逐渐积累起了整个任务的背景，包括：为什么要做这个功能，哪些方案已经被否定，还有什么约束必须遵守等等。
可当后续有另外一位同事想参与进来，或接手你的工作时，事情就变得开始低效了。
我们需要把对话截图发过去，把代码复制出来，再把同事的意见转述回给 AI。对方如果想追问一句，又要等我们把问题送进去，再把回答搬回来。
AI 明明已经能够参与越来越多的执行工作了，但人与人之间的协作，仍可能被这些来回的转述给拖住。
个人的生产力被放大了，团队掌握的信息，却未必能同样顺畅地汇集到一起。
是时候，我们该重新搬回那张“椅子”了。
Codex Co² 想解决什么？
基于同一条会话的协作，正是我想用 Codex Co² 来解决的问题。
Codex Co² 的全称是 Codex Conversation Collaboration（Codex 会话协作），名字有点长，但是想做的事情很简单：让多个人参与同一条 Codex 会话，一起推进工作。
分享者可以继续在官方 Codex Desktop 中工作，只需在 Co² 里选中一条已有会话，生成邀请，发给协作者。协作者则通过浏览器打开邀请，申请加入；获准后，就能阅读共享上下文，向同一个 Agent 补充要求，并跟进它的回复。
协作者无需安装 Co²，也无需单独登录 Codex。发出的每一条消息，都会沿用分享者当前会话的工具、权限和额度。
邀请协作者
不妨看一个具体的例子。
假设有用户反馈：“任务已经处理完成，页面上的状态却一直没有变化，刷新之后才恢复。”
运营拿到 FeedbackID 后，可以先在 Codex 中分析相关的用户日志，得到初步结论：操作请求已经成功，问题可能出在页面状态的更新环节。
随后，运营就可以通过 Co² ，将这条会话分享给研发。
研发加入后，可以沿着已有的分析继续追问：服务端有没有发出状态变更通知？客户端是否收到？异常发生前，是否经历过断线和重连？
结合日志与相关代码，问题逐渐收敛：客户端在网络重连后没有恢复实时订阅，后续状态变更因此未能到达页面；而手动刷新会重新拉取数据，所以显示又恢复了正常。
到这里，一条模糊的用户反馈，才有了可以解释、可以验证的技术根因。
邀请另一协作者
但找到根因，并不意味着就走到了解决问题的最后一步。
研发可以继续让 Agent 梳理修复方案：补齐重连后的订阅恢复，并在恢复连接时校准一次状态；同时整理受影响的场景、改动范围、验证成本，以及临时处理方式，再邀请产品加入这条会话。
产品看到的，便包括了用户最初遇到的困难、运营从日志中获得的线索、研发定位根因的过程，以及不同处理方式的成本与风险。
基于这些信息，产品可以决定是否需要立即修复、如何安排版本，以及在修复上线前怎样回应用户。
运营带来现场，研发追到根因，产品决定如何推进。
不同角色的判断，在同一条会话里逐步接续。前一个人留下的分析，成为后一个人继续追问的起点，最终把一条用户反馈推进为一个有依据的处理决定。
收集分散的信息
Co² 带来的这个看起来很小的变化，恰恰对应着软件开发中一个长期存在的事实：理解一个问题所需要的信息，往往分散在不同的人手里。
产品同事知道某条看似奇怪的规则，背后对应着怎样的业务现实；测试同事知道一个偶发问题，究竟要经过哪些步骤才能复现；熟悉系统的工程师则知道，一处看似无害的改动，可能牵动哪些历史约束。
这些判断，很难仅凭一句“帮我把功能做出来”就交代清楚。
AI 能够处理的信息再多，也需要有人把这些最相关的事实带到它面前。因此，当我们讨论如何更好地使用 AI 时，更值得考虑的是：怎样让那些真正了解问题的人，更容易参与当前的工作。
Co² 尝试提供的，就是这样一个入口。
我希望它能够用在结对开发中，让另一位工程师直接参与方案讨论；用在远程排障中，让了解现场的人补充复现条件；也用在需求澄清中，让业务上的判断及时进入实现过程。
还有更多可能性
当然，Co² 目前仍是一个早期的独立实验项目。
项目代码已经在 Github 上开源，提供了 macOS 和 Windows 安装包，可以通过 Homebrew 安装后打开，也可以直接在浏览器上运行。
使用时还需要理解以下几点限制：
- 共享会话沿用原来的执行环境和额度，协作者的每一条消息，实际都是以分享者的身份在发送的，使用时需要注意权限控制和额度把控。
- 公网分享经过中继，使用普通 TLS，没有端到端加密，记得只邀请可信任的协作者，或者开启“仅内网可访问”开关；
- 退出 Co² 时，分享出去的邀请链接会变成不可访问，因为 Co² 设计上就是服务于当下的协作，而不是作为一个长期的入口。
如果你也正在和同事一起使用 Codex，欢迎试用，并把遇到的体验问题通过 Issue 反馈给我，一起参与 Co² 的共创。
结语
希望，以后我们在用 AI 来提升工作效率的同时，也能继续给身边的同事留出一个位置。
让了解需求的人、熟悉代码的人、发现问题的人，都能把各自掌握的那一部分带进来，完成一个共同的协作，这才是团队存在的意义。
曾经屏幕前的那把椅子，值得再搬近一些。
Github 地址： github.com/SpaceCoder-…
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Removed chairs
In the past, two programmers wanted to solve the same problem together and could move the chair to the front of the same screen.
One person handles the keyboard, the other person examines the thinking; when differences arise, the discussion stops. The code on the screen is the product of a joint output of two people, and the context of the problem is gradually aligned in this answer.
The method of such collaboration is called twinning programming.
Now, AI has become a new partner on the other side of our screen.
We describe needs to it, let it read codes, analyse problems and try to modify them. After a round of dialogue, that conversation gradually built up the context of the entire mandate, including: why this function should be done, which options have been rejected, what constraints must be met.
But then when another colleague wants to be involved or takes over your work, things become ineffective.
We need to send a message, copy the code and retransmit the comments of our colleagues to AI. If the other party wants to ask a question, we'll have to wait until we get the questions in and get the answers back.
Although AI has been able to participate in an increasing number of implementation efforts, people-to-people collaboration may still be delayed by these circulars.
Individual productivity has been amplified and the information available to the team has not necessarily been brought together equally smoothly.
It's time we moved back to that chair.
Codex Co2. What are you trying to solve?
Based on the collaboration of the same session, it's the problem that I want to solve with Codex Co2.
The full name of Codex Co2 is the Codex Conference Cooperation, which is a little long, but the simple thing to do is to involve a number of people in the same Cordex session and move forward.
Shareholders can continue to work in official Codex Desktops by selecting an existing session in Co2 to generate invitations and send them to collaborators. The collaborators open the invitation through the browser and apply to join; once approved, they can read shared context, add to the same Agent request and follow up on its response.
The collaborators do not need to install Co2 or log in alone. Every message sent will follow the Shareholders ' current session tools, permission and quota.
Invite collaborators
A specific example may be found.
Assume that there is feedback from users: “The task has been processed and the status on the page has remained unchanged and is not restored until it is refreshed.”
Once you get a feedbackID, you can analyze the user logs in Codex and get a preliminary conclusion that the operation request has been successful and that the problem may arise from an update of the page status.
The operation can then share this session with R & D via Co2.
With the addition of R & D, it is possible to follow up on the existing analysis: has the service provider issued a status change notice? Do you copy the client? Before the anomaly, did you experience a disconnection and reconnection?
In combination with the logs and related codes, problems are becoming less problematic: the client did not resume real-time subscriptions after reconnecting the network, and the subsequent status change did not reach the page; and the manual refreshing would retake the data, thus showing a return to normality.
Here, a vague user feedback is the source of an explanatory and verifiable technology.
Invite another collaborator
Finding the root causes, however, does not mean that the last step towards a solution has been taken.
R & D can continue to enable the Agent combo restoration programme: to restore the reconnected subscriptions and calibrate the state once the connection is restored; and to organize the affected scenes, the scope of the change, the cost of validation, and the ad hoc treatment, inviting the product to join the session.
The product sees the initial difficulties encountered by users, the clues to operations from logs, the process of developing and locating the root causes, and the costs and risks of different approaches.
On the basis of this information, the product can determine whether immediate repairs are required, how the versions are scheduled and how to respond to users before the repairs are online.
Operations bring on-site, R & D takes root and products decide how to move forward.
The judgement of the different players continues gradually in the same session. The analysis left behind by the former person serves as a starting point for the latter to continue his inquiries, ultimately pushing a user feedback into a valid handling decision.
Collect decentralized information
This seemingly small change brought about by Co2 corresponds precisely to a long-standing reality in software development: the information needed to understand a problem is often dispersed among different people.
Product colleagues know what kind of business reality is behind a seemingly strange rule; test colleagues know what kind of steps are required to repeat an occasional problem; and system-acquainted engineers know what historical constraints may be involved in a seemingly harmless change.
It is difficult for these judgements to be clear with just one sentence: “Help me make it work”.
As much information as AI can handle, those most relevant facts need to be brought before it. So, when we discuss how to make better use of AI, it is even more worth considering how to make it easier for those who really understand the problem to participate in the current work.
Co2 tried to provide one of these entrances.
I hope that it can be used in twinning development to involve another engineer directly in programme discussions; in remote barriers to allow those who know the site to supplement the recovery conditions; and in clarification of needs to allow operational judgement to enter the process in a timely manner.
There are more possibilities.
Of course, Co2 is still an early independent experiment.
The project code is already open on Github and provides the MacOS and Windows installation packages, which can be opened through Homebrew installation or run directly on the browser.
The following limitations also need to be understood:
- The sharing of sessions follows the original implementation environment and quota, and each message from a collaborator is actually sent as a sharer, using permission control and quota control.
- Public Internet sharing is relayed, using regular TLS, without end-to-end encryption, bearing in mind that only trusted collaborators are invited, or the “innernet accessible” switch is activated;
- When exiting Co2, the shared invitation link becomes unaccessible because Co2 is designed to serve immediate collaboration rather than as a permanent entry point.
If you're also using Codex with your colleagues, welcome the experiment, and give me feedback on the experience that you've experienced, through Issue, to join in Co2 co-creation.
Concluding remarks
It is hoped that, in the future, while using AI to improve efficiency, we will continue to give our colleagues a place.
To allow those who know what they need, those who know the code, those who discover the problem, to bring in the part that they have and to work together, that is what a team exists.
The chair that was in front of the screen was worth moving closer.
Github Address: gethub.com/SpaceCoder-...
Auto-fetched and translated by AI Atlas. All rights belong to the original author.