Add ability to immediately terminate agent

[malhotra5]

Today we expose an API conversation.pause to stop the agent. However, this will wait for the agent to wrap up its work (mid llm response, tool call, etc) before returning control. This can take up to minutes nowadays (e.g when the LLM is performing a long reasoning for one llm response)

We want to expose an API conversation.interrupt to stop the agent immediately. It should return control immediately and cancel any mid-fight LLM completions or tool executions

[linear]

from rohit:

I've proposed a PR to resolve this issue. It creates an async task for the llm completion. During interrupt we abandon the in-flight request (its still undergoing completion) and return control the to user.

However, ideally we would have our own async alternatives everywhere instead of this hack. We'd need to decide whether we'd like to maintain async alternatives to all sync apis. Eg. conversation.run_async in addition to conversation.run

cc @xingyao

[jpelletier1]

Agree with this. Would be great to have a follow-on ticket also created for CLI once this is implemented.

[linear]

from rohit:

Yep I've already opened a ticket and implemented the interrupt ability using the proposed changes in the SDK: OpenHands/OpenHands-CLI#548

It seems to work very well