One turn as a sequence of events, so the colour arrives first.

The blocking /chat runs tool, then a second model round for commentary, then returns everything together -- so the customer waits the full turn to see a palette that existed a third of the way in.

This yields the visual the moment the tool returns and streams the commentary after it. Same tools, same metering, same entitlement check, same result object; only the delivery changed. /chat is untouched for anything that wants one JSON body.