Agents should support thinking models, so we need to support thinking blocks and we should log the thinking