[功能] 增加 Agent Sandbox 网络域名策略和工具返回内容安全审查

The desired result is as follows: 1. Sandbox can still connect to the network normally. 2. External return content is first screened. 3. Hit content is censored or blocked. 4. Other security results can still be used. 5. Users will not see the original high-risk text. 6. Original content will not enter the model context, run details, or ordinary persistence records. 7. Even if the model refuses or interrupts, the original tool return will not be exposed directly.