用户接管阶段可能被页面状态变化过早结束 | User takeover may end prematurely due to page state changes

[Ljy-0827]

问题描述

当 BrowserSkill 将浏览器控制权交给用户处理验证码、密码等步骤时，页面状态变化可能会触发 Agent 重新接管，但此时用户的操作实际上还没有完成。此类情况包括但不限于：

• Agent 请求用户填写验证码后，用户点击“换一个验证码”，页面发生刷新或局部更新，Agent 就重新开始操作。
• 用户输入密码后提示错误，可能还需要重新输入，甚至进入重置密码流程，但页面跳转后 Agent 可能提前接管。

预期行为

在用户接管期间，页面刷新、导航或状态变化不应自动表示用户操作已经完成。
Agent 应持续等待，直到满足以下条件之一：

• 用户明确点击“完成”或“交还控制权”；
• 系统检测到当前人工处理目标已经真正完成，例如已经成功登录。
  用户接管阶段的结束条件应该基于明确的完成信号，而不是任意页面状态变化。

Issue Description

When BrowserSkill hands over browser control to the user for steps such as entering a CAPTCHA or password, a page state change may cause the Agent to take control again even though the user has not actually finished the task.

Such cases include, but are not limited to:

• After the Agent asks the user to enter a CAPTCHA, the user clicks “Try another” or a similar option. The page refreshes or updates partially, and the Agent resumes operation immediately.
• After the user enters an incorrect password, they may need to try again or even go through a password reset flow. However, the Agent may take control prematurely after a page navigation or state change.

Expected Behavior

While the user has control, a page refresh, navigation, or state change should not automatically be treated as completion of the user-handled step.

The Agent should continue waiting until one of the following conditions is met:

• The user explicitly clicks “Done” or “Return control”;
• The system detects that the user-handled task has actually been completed, such as a successful login.

The end of the user takeover phase should be determined by an explicit completion signal, rather than by an arbitrary page state change.