Claude Sonnet 5.5 via Claude Desktop, but using the 'Code' model therein. It seems trained to be engaging, with followup questions, suggesting next steps, and that crap. This is REALLY BAD for research coding, I think. In those situations, I often have no clear goal, just a general direction. I develop a bit. I see how it behaves. And I change goal. Having an engagement maximizing agent just opens new rabbit holes everywhere. It never closes things. You never get done. Horrible.