Claude Code 发布

终端最诚实的地方，是失败不会被藏起来。命令返回非零状态，测试把堆栈打印满屏，Git 直接列出哪些文件还没提交。许多工程师可以在编辑器里写代码，却要回到这里确认软件究竟能不能运行。

2025 年 2 月 24 日，Anthropic 随 Claude 3.7 Sonnet 发布 Claude Code 研究预览。它没有在编辑器旁再放一个聊天框，而是在本地进程里读取代码库、修改文件、执行命令，再根据输出继续工作。模型第一次以产品化方式进入了完整的反馈回路：看状态，做改动，跑测试，读错误，然后重试。

这条路径也让风险变得具体。终端附近可能有未提交改动、环境变量、部署命令和访问凭据。Claude Code 在较高风险操作前请求权限，目的不是制造礼貌感，而是划定信任边界。模型可以建议下一步，能否执行仍取决于用户授予什么能力。离真实环境越近，授权的颗粒度越不能含糊。

项目还可以用 CLAUDE.md 保存仓库级指令，把构建命令、代码约定和注意事项留在版本控制附近；GitHub 与 MCP 连接则把可访问范围继续向外扩展。上下文因此不只是一段临时提示，而开始成为项目资产。但指令文件不会替代测试，连接更多工具也不会自动提高判断力。错误的项目说明会被忠实遵循，宽泛的权限会把一次误判放大成真实操作。

“研究预览”限定了这次发布的分量。它展示了一种协作节奏，而不是宣告所有仓库都能交给 Agent。人与模型共享同一个 shell 反馈，人的工作从逐行输入命令，部分转向审查计划、批准操作和验收结果。代码审查也随之扩大：要看的不只是最终 diff，还有中间运行过什么、失败后为什么选择了下一步。

Claude Code 最有代表性的画面不是模型连续敲出多少代码，而是执行前停住，等待一个许可。那个停顿承认了终端的双重性质：它既是软件工程最完整的反馈界面，也是把错误迅速变成后果的地方。

The terminal is honest because failure has nowhere to hide. A command returns a nonzero status. A test fills the screen with a stack trace. Git lists the files that remain uncommitted. Many engineers write inside an editor and return here to find out whether the software actually works.

On 24 February 2025, Anthropic released a research preview of Claude Code alongside Claude 3.7 Sonnet. It did not merely add another chat panel beside an editor. Running as a local process, it could read a repository, edit files, execute commands, and continue from their output. A model entered the full feedback loop as a product: inspect state, change something, run tests, read the failure, and try again.

That path also made risk concrete. The neighborhood around a terminal can contain uncommitted work, environment variables, deployment commands, and credentials. Claude Code requested permission before higher-risk actions. The prompt was not interface politeness; it marked a trust boundary. The model might propose the next step, while actual execution still depended on what the user granted. The closer a model moves to a real environment, the less vague permission can be.

A repository could also keep project instructions in CLAUDE.md, placing build commands, conventions, and warnings near version control. GitHub and MCP connections extended the reachable environment further. Context began to look less like a disposable prompt and more like a project asset. Yet an instruction file could not replace tests, and additional tools could not supply judgment automatically. Incorrect project guidance would be followed faithfully; broad permissions could turn one mistaken inference into a real operation.

“Research preview” bounded the claim. The release demonstrated a collaboration rhythm, not a guarantee that every repository could be handed over. A person and model shared the same shell feedback. Human work shifted partly from typing each command to reviewing a plan, authorizing actions, and accepting results. Code review expanded as well: the final diff mattered, but so did which commands ran and why the agent chose its next step after a failure.

The defining image of Claude Code is not an uninterrupted stream of generated code. It is the process stopping before execution and waiting for permission. That pause admits the terminal's double nature: it is software engineering's most complete feedback surface and also the place where an error can most quickly become a consequence.