默认每一步都要确认，再按需放行低风险操作。
Bi 的默认做法是：智能体的每一次操作都需要你的确认，包括读取文件。控制权先交给你，再由你决定哪些操作可以免去确认。
在保持整体安全级别的前提下，通过按需设置自动放行提升易用性。
在默认设置下，智能体的每一步都会弹出确认框：
你可以逐一批准，也可以拒绝。
如果某些操作你反复批准、并且确信没有风险，可以在「设置」中把它们设为自动放行。常见的做法是：
写入文件和运行命令的风险更高，建议保持确认。
除了逐次确认，Bi 还会拦截一部分明确危险的命令（例如删除、推送代码、修改权限）。这类操作会被直接拒绝，不会执行，详见危险操作拦截。
无论权限如何设置，文件操作都被限制在你选定的工作区内，范围之外的路径一律拒绝。这是与权限独立的一层保护，无法通过设置放宽。见工作区。
Every step asks first, and you can auto-approve low-risk actions.
By default, every action the agent takes needs your approval, including reading a file. Control starts with you; you decide which operations can skip confirmation.
Making Bi smoother is a matter of approving specific low-risk actions, not lowering the overall security level.
Out of the box, every step raises a confirmation prompt:
You can approve or reject each one.
If you approve the same kind of action repeatedly and are confident it is safe, set it to auto-approve in Settings. Common choices:
Writes and command execution carry more risk; keep them confirmed.
Beyond per-action approval, Bi intercepts a set of clearly dangerous commands (deleting, pushing code, changing permissions). Those are refused outright rather than prompted. See Dangerous command interception.
Whatever the approval settings, file operations stay inside the workspace you chose; paths outside it are refused. This is a separate layer that settings cannot relax. See Workspace.
