How it works
Widgets stream before the model speaks.
The program is classified locally and the widgets paint first. The response-time badge stops when the reply text lands.
01 · Local
Widgets first
Catalog and regex still build the program before narration. A compound question splits into clauses and paints every widget in one reply. Groq classifies a clause only when regex cannot, and those calls run together.
02 · Narration
Replies in parallel
Each clause gets its own Groq reply. Those calls run together, then one ReplyCard stitches them in order. If narration fails, the canned line stays.
03 · Real API
ESPN fills the facts
Query() loads fixtures, tables, squads, and reports after the widgets render. Matchup and final narration fetch ESPN facts for that clause only, in parallel with the others.