Require approval in cli

[jsibbison-square]

This POC shows how tool requests can be approved or rejected by the user. The interaction pattern is:

• Model returns some ToolRequests.
• Our agent.rs (Agent) adds approval ids to tool requests
• Agent passes the tool request Messages as normal up to the interface
• If there were any tool requests requiring approval the Agent.reply stream pauses and waits on a channel for approval/rejection for each tool request
• User uses the Interface to approve/reject each tool request which is passed from the Interface to the Agent via a channel.
• Once all tool request approvals are actioned the rejected requests will get a failed tool result message added and approved requests will be sent to the systems.

So the paradigm here is that systems are ONLY called when the user approves a tool request (which offcourse we can layer in auto-approve fancy logic for systems/tools).

It's also relatively easy with this logic to incorporate a system triggered approval request if there is an api that can be called on the system which decides if approval is required.

The way the approval happens in the interface can be customised by the approval type and details where you could redirect the user to do something externally if needed. The interface just needs to be able to tell the agent the user has finished actioning approval requests.