How to stop a chat from a tool?

[joshforbes]

Is there any way to explicitly stop a chat from within a tool call? Here is the use case:

Our assistant will suggest a script to the end user, but needs to wait for approval before actually running it on the selected device. This approval may take hours, day, or never come (think something like Cursor approve/deny).

So I was thinking I would do something like this:

• RunScriptTool
• The execute methods creates some kind of approve/deny record that is attached to the chat
• The chat needs to exit indefinitely until the user responds
• Once the user responds, I'll run the script in the background (outside of the LLM)
• I'll load the chat from the database and feed the output back into the chat as a new system message

But I can't quite figure out how I'm intended to do something like this. Is there a expected method to guarantee a tool call to exit the chat once it is finished? I imagine that I can't just reliably tell the LLM that it should do that. The only thing I can figure is to fire some kind of custom error, but that seems... off and I'm not sure if it would have other negative impacts.

Anyone else doing something like this?

I saw that "human in the loop" was discussed previously here: #48. But that seemed more like a long-running execute method, instead of trying to return control to the app.

Thanks!

[tpaulshippy]

Similar to #256

But if you just return "done" from the tool and you are ok with that one last call to the LLM, I'm not sure what the issue is.

[joshforbes]

Honestly, didn't see that issue. Thanks for linking it, it is very similar for sure.

But if you just return "done" from the tool and you are ok with that one last call to the LLM, I'm not sure what the issue is.

I'm not really sure about this. To be completely honest, my company is just getting into this space, so my experience is very limited thus far. But it seems like returning "done" would still leave it open for the LLM to decide to do something else. Maybe it was going to perform a chain of events, but we want to stop it until we can perform some actions. Idk, could be totally wrong.

Thank you for responding.

[tpaulshippy]

You are right. It is a possibility.