Aider 把 Git 结对写进终端

2023 年春天，AI 编程最醒目的画面仍是光标后面自动长出的代码。GitHub Copilot 已经让补全进入日常，但补全只负责眼前几行；一旦任务变成“在这个仓库里改三个文件，再确认没有破坏测试”，开发者通常还得把代码复制到网页聊天，解释目录关系，再亲手把答案搬回来。

Aider 选择了一个不那么华丽的入口：终端。开发者在本地 Git 仓库里发出指令，工具读取相关文件，把模型的回答变成实际补丁，再把变化交给 Git 展示。模型可以提议改动，仓库仍用 diff 回答一个更重要的问题：它到底改了什么？

这套设计没有试图把版本控制藏起来。恰恰相反，Git 是协作契约。一次修改可以被逐行检查，可以撤销，也可以形成提交记录。

对生成代码来说，这些旧工具不是负担，而是约束失误的缰绳。模型换了，API 供应商换了，审查和回退的路径仍留在开发者手里。

Aider 也逐渐把代码库映射、测试与提交纳入同一条终端工作流。它并没有因此成为一个可以独自负责项目的“虚拟工程师”。仓库越大，相关上下文越难选；测试缺失时，漂亮的 diff 也可能藏着错误；模型给出的解释仍需要工程判断。Aider 的可信感并不来自模型永远正确，而来自错误能够以熟悉的工程形式暴露出来。

Apache 许可和自带密钥的路径，又让这套方法脱离单一商业编辑器。团队可以选择模型、查看实现，并把工具放进已有的开发习惯里。这与后来强调一体化体验的 AI IDE 并不冲突：一边优化“所有能力都在同一个界面”，另一边坚持“仓库和模型不必交给同一个产品”。

后来终端里出现了更多编程 Agent，功能更强、名字更响。Aider 留下的判断却很朴素：让模型进入真实仓库之前，先让每一次改动都能被看见、被拒绝、被退回。终端里滚动的彩色 diff，才是这段开源结对故事真正的主角。

In the spring of 2023, the defining image of AI coding was still text appearing after a cursor. GitHub Copilot had made completion ordinary, but completion usually stopped at the code in front of the developer. When the task became “change three files in this repository and make sure the tests still pass,” people often returned to browser chat, explained the directory structure, and carried the answer back by hand.

Aider chose a less theatrical entrance: the terminal. A developer opened a local Git repository, described a change, and let the tool turn a model response into edits against real files. Git then asked the question that mattered more than the fluency of the answer: what, exactly, changed?

Version control was not hidden behind a friendly interface. It was the contract. A proposed edit could be inspected line by line, reverted, or preserved in a commit.

For generated code, those old mechanisms were not ceremony; they were reins. Models could change and API providers could change while review and rollback remained in the developer's hands.

Aider's workflow grew to include repository mapping, test execution, and commits. None of that made it an engineer that could safely own a project alone. Large repositories still made context selection difficult. A convincing diff could still be wrong when tests were weak. Explanations still required engineering judgment. The tool's credibility came less from a promise that the model would always be right than from the fact that mistakes appeared in familiar engineering artifacts.

Its Apache license and bring-your-own-key approach also separated the workflow from a single commercial editor. Teams could choose models, inspect the implementation, and fit the tool around habits they already had. That did not invalidate the integrated AI IDE. It established a different bargain: one product could optimize for everything living in one interface, while an open terminal tool could insist that the repository and the model did not need the same owner.

More powerful and more famous coding agents later arrived in the terminal. Aider's durable idea remained modest: before allowing a model into a real repository, make every change visible, rejectable, and reversible. The colored diff scrolling past the prompt—not a claim of autonomy—is the central image of its story.