Interrupts and Placeholders

How to fit asynchronous interrupts into a synchronous format: normally the LLM sees an ideal synchronous trajectory; only on interrupt is a placeholder inserted.

Rules:

assistant message is written immediately;
tool result is written upon completion;
interrupt during tool execution gets a placeholder tool result (“tool is running in the background”);
interrupt during thinking discards unwritten reasoning;
non-urgent events are queued.

Risk: the model may hallucinate the result of an unfinished tool – it hasn’t been trained on the “result not yet arrived” situation.