Workaround: You can ask the model in a prompt that when it finishes thinking, it will write you some sign that you decide on, say <thinking_end> and then you can easily do a split