Surface for the model's internal state — extended thinking, plans, and 'is working' indicators.
Collapsible disclosure for the model’s extended-thinking output, with live streaming.
Vertical stepper for explicit, structured reasoning steps.
Lightweight "model is working" affordance — distinct from the typing indicator.