[PRIORITY] Design a tool call approval mechanism

[dlqqq]

One thing we will need before the v3.0 official launch is to have some way for users to approve tool calls before they are run.

We need to design:

1. What is the rubric we will apply to determine when tool approval is necessary?
2. How do we send a signal from the UI to the backend to indicate tool approval?
3. How do we design "auto-approval" for trusted personas? Should we make this specific to the user invoking it, the chat it is in, or both?