In our use case, the thinking phase of the agent is relatively long. We want to allow one interruption from the user that immediately stops the current TTS and triggers a new LLM + TTS response — this already works as expected.

However, the problem arises when the user interrupts again during the thinking phase. Currently, this triggers a second LLM call which breaks the UX and flow. We want to avoid that.

Ideally, we would like to dynamically disable interruptions only during the thinking phase, after the first interruption has occurred. Right now, updating allow_interrupts via AgentSessionOptions mid-session doesn’t seem to have any effect.

Request:
• Add support for dynamically modifying the allow_interrupts behavior during a session
• Specifically, we’d like to:
Allow interruptions during only during "speaking"

Is there a recommended way to do this? Or could this be supported in the SDK/runtime?