Function decideHandler

services/auto-routing/src/decide.ts:270–402

}

// Single sink for decision telemetry: one Analytics Engine data point and
// one `auto_routing_decision` log line per decision. Successes are sampled
// per the KV-configured rate; fallbacks, errors, and real model switches
// always log (failures at warn).
function recordDecision(
  env: Env,
  ctx: DecisionContext,
  durationMs: number,
  outcome: DecisionOutcome,
  autoRoutingMode: string,
  decision: AutoRoutingDecision | null = null,
  incumbent: StickyDecision | null = null
): void {
  const summary = summarizeOutcome(outcome);

  writeClassifierMetricsDataPoint(env, {
    status: summary.status,
    classifierModel: summary.classifierModel,
    requestedModel: ctx.payload.input.requestedModel,
    classification: summary.classification,
    classifierCostCredits: summary.cost,
    classifierDurationMs: durationMs,
    cacheHit: summary.cacheHit,
  });

  const incumbentModel = incumbent?.model ?? null;
  const decidedVariant =
    decision !== null && decision.source === 'benchmark' ? (decision.variant ?? null) : null;
  const switched =
    decision !== null &&
    incumbentModel !== null &&
    (decision.model !== incumbentModel ||
      (incumbent?.variant != null && decidedVariant !== incumbent.variant));
  const routeKey = summary.classification ? taxonomyRouteKey(summary.classification) : null;
  // Null when there is no incumbent route to compare against (no incumbent,
  // a pre-routeKey cache entry, or no classification this request).
  const routeChanged =
    routeKey !== null && incumbent?.routeKey != null ? routeKey !== incumbent.routeKey : null;

  // Retried decisions are rare and diagnostically valuable, so they bypass
  // sampling along with failures. Real model switches also always log:
  // within-session switch sequences are the signal the sampled stream
  // decimates.
  const isFailure = summary.status !== 'classified';
  const alwaysLog = isFailure || summary.retried || switched;
  if (!alwaysLog && Math.random() >= ctx.successSampleRate) {
    return;
  }
  const log = isFailure ? console.warn : console.log;
  log(
    JSON.stringify({
      event: 'auto_routing_decision',
      status: summary.status,
      cacheHit: summary.cacheHit,
      retried: summary.retried,
      classifierModel: summary.classifierModel,
      requestedModel: ctx.payload.input.requestedModel,
      apiKind: ctx.payload.input.apiKind,