Surfacing reasoning

If your model exposes a reasoning/thinking stream, deciding whether to show it is a UX choice:

• Show it on interactive surfaces where users value seeing the agent's process; it builds trust and fills the latency productively.
• Hide it (or don't even request it) on bulk/background tasks where it's noise and cost.

Make it a per-call decision (Chapter 4), and stream reasoning as its own event type so the UI can render it distinctly (often collapsed/greyed) from the final answer.