Bi

内置终端

在界面中直接操作命令行,并区分它与智能体运行命令。

Markdown 原文Raw markdown

Bi 中有两种「运行命令」的入口,容易混淆,先区分清楚:

入口 谁来用 结果去哪里
终端面板 你本人 只显示在屏幕上,不会进入模型的上下文
智能体运行命令 智能体 作为运行结果反馈给智能体,供它继续处理任务

终端面板

界面中的「终端」是一个可以持续使用的命令行窗口,供你手动操作:

  • 工作目录取自当前会话的工作区;未设置会话时使用全局工作区
  • 输出逐行显示,中文不会乱码
  • 它的输出只显示给你,不会自动提供给模型

智能体运行命令

智能体执行的命令在工作区根目录下运行,结果会反馈给它,用于继续推进任务:

  • 运行前会请你确认(除非你把相应操作设为自动放行)
  • 命令会经过危险操作拦截检查

因此,「命令的输出会进入上下文」只对智能体运行命令成立。如果你希望智能体看到某条命令的结果,应当让它去运行,无需在终端面板中运行后再粘贴结果。

关于联网

智能体运行的命令可以访问网络(例如下载文件、安装依赖、拉取代码)。这是功能的一部分,但意味着:

  • 即使使用本地模型,也不等于「绝不联网」——命令本身仍可能发起网络请求
  • 运行命令前的确认是主要的人工把关环节,请留意你批准的每一次运行

相关

Built-in terminal

Use a shell inside the UI, and understand how it differs from agent commands.

Markdown 原文Raw markdown

Bi has two places where commands run. They are easy to confuse, so start with the difference:

Entry point Who uses it Where the output goes
Terminal panel You Shown on screen only; it does not enter the model’s context
Agent command The agent Fed back to the agent so it can continue the task

Terminal panel

The “Terminal” in the UI is a persistent shell for you:

  • Its working directory is the current session’s workspace; without a session it uses the global workspace
  • Output appears line by line, with no mojibake for Chinese
  • Its output is shown to you only, never sent to the model automatically

Agent commands

Commands run by the agent execute at the workspace root and feed their output back to it:

So “command output enters the context” applies only to agent commands. If you want the agent to see a command’s result, have the agent run it rather than running it yourself and pasting the output.

About network access

Agent commands can reach the network (downloading files, installing dependencies, pulling code). That is part of the feature, and it means:

  • A local model does not imply “never online” — a command can still make network requests
  • The confirmation before a command is the main human checkpoint; pay attention to every run you approve