/** Transitions.dev Reasoning stream, adapted to appended content, never a loop. */ export const REASON_HOLD = 840; export const REASON_STEP = 500; export const REASON_LINES = 2; /** Ceiling on how much text one catch-up step may cover. */ export const REASON_MAX_LINES = 40; /** Beat between catch-up steps, while the transcript is still well behind. */ export const REASON_CHASE_HOLD = 40; export function reasoningTarget( top: number, contentHeight: number, viewportHeight: number, lineHeight: number, backlogLines = 0, ): number { const end = Math.max(0, contentHeight - viewportHeight); const current = Math.min(end, Math.max(0, top)); // `backlogLines` is how far the newest text sits below what is painted. The // reference step size is a reading pace: two lines, then a hold. That pace is // far slower than a fast model writes, so the viewport used to crawl behind in // fixed hops no matter how quickly the text arrived. Covering the backlog in // one step keeps the newest line with the text instead; a model that writes at // or below the reading pace still leaves the backlog at two lines, unchanged. const lines = Math.max(REASON_LINES, Math.min(REASON_MAX_LINES, Math.ceil(backlogLines))); return Math.min(end, current + Math.max(1, lineHeight) * lines); }