app/src/token-rate.mjs

// Decode throughput over a rolling second. The first token establishes the
// baseline, so prompt processing and first-token latency are not counted.
export class TokenRateWindow {
  constructor(windowMs = 1000, minimumMs = 250) {
    this.windowMs = windowMs;
    this.minimumMs = minimumMs;
    this.total = 0;
    this.firstAt = null;
    this.points = [];
  }
  add(count, at) {
    if (!(count > 0)) return;
    this.firstAt ??= at;
    this.total += count;
    this.points.push({ at, total: this.total });
    this.prune(at);
  }
  prune(at) {
    const cutoff = at - this.windowMs;
    while (this.points.length > 1 && this.points[1].at <= cutoff) this.points.shift();
  }
  rate(at) {
    if (this.firstAt === null || at - this.firstAt < this.minimumMs) return null;
    this.prune(at);
    const elapsed = Math.min(this.windowMs, at - this.firstAt);
    return (this.total - this.points[0].total) * 1000 / elapsed;
  }
}