@zuplo/runtime 6.70.62 → 6.70.63

package/out/types/index.d.ts

@@ -176,6 +176,22 @@ export declare class AIGatewayMeteringInboundPolicy extends InboundPolicy<AIGate
     increments: AIGatewayMeterIncrements
   ): void;
   static getIncrements(context: ZuploContext): AIGatewayMeterIncrements;
+  /**
+   * Record the global quota fallback models for the current request. Set by the
+   * metering policy when a quota is exceeded and a fallback is configured; read
+   * by the LLM translation handler, which routes to the capability-appropriate
+   * model instead of the (over-budget) primary.
+   *
+   * @param context - The ZuploContext
+   * @param quotaFallback - The validated quota fallback models.
+   */
+  static setQuotaFallback(
+    context: ZuploContext,
+    quotaFallback: QuotaFallbackModels
+  ): void;
+  static getQuotaFallback(
+    context: ZuploContext
+  ): QuotaFallbackModels | undefined;
   constructor(
     options: AIGatewayMeteringInboundPolicyOptions,
     policyName: string
@@ -185,6 +201,23 @@ export declare class AIGatewayMeteringInboundPolicy extends InboundPolicy<AIGate
     context: ZuploContext
   ): Promise<Response | ZuploRequest<RequestGeneric_2>>;
   private fetchCurrentMeters;
+  /**
+   * The capability a request targets, derived from its path, or `undefined` for
+   * paths that do not support quota fallback (e.g. `/v1/responses`, which the LLM
+   * handler serves without the fallback chain). Used to decide whether a quota
+   * fallback actually applies to *this* request.
+   */
+  private requestCapability;
+  /**
+   * Validate and return the configured quota fallback when it applies to *this*
+   * request's capability, or `undefined` otherwise (no config, malformed config,
+   * an unsupported path, or no model configured for the request's capability).
+   * Returning `undefined` blocks the request with a 429 — so a fallback is only
+   * counted and applied when the handler can actually serve the request with it.
+   * The config is customer-authored and reaches us through an `unknown` cast, so
+   * we validate it through Zod before acting on it.
+   */
+  private resolveQuotaFallback;
   private checkHierarchicalQuotaLimits;
   /**
    * Increment meters via API. Can be used by providers for streaming responses.
@@ -2969,6 +3002,7 @@ declare const EventType: {
   readonly AI_GATEWAY_LATENCY_HISTOGRAM: "ai_gateway_latency_histogram";
   readonly AI_GATEWAY_WARNING_COUNT: "ai_gateway_warning_count";
   readonly AI_GATEWAY_BLOCKED_COUNT: "ai_gateway_blocked_count";
+  readonly AI_GATEWAY_FALLBACK_COUNT: "ai_gateway_fallback_count";
   readonly MCP_REQUEST_RECEIVED: "mcp_request_received";
   readonly MCP_REQUEST_COMPLETED: "mcp_request_completed";
   readonly MCP_REQUEST_REJECTED: "mcp_request_rejected";
@@ -6844,6 +6878,12 @@ export declare interface MockApiInboundOptions {
  */
 export declare const MockApiInboundPolicy: InboundPolicyHandler<MockApiInboundOptions>;
 
+declare interface ModelConfiguration {
+  environmentVariable: string;
+  model: string;
+  provider: string;
+}
+
 declare type Modify<T, R> = Omit<T, keyof R> & R;
 
 declare interface MoesifContext {
@@ -8667,6 +8707,15 @@ export declare interface QuotaDetail {
   };
 }
 
+/**
+ * Global quota fallback models, keyed by capability. Applied whenever any
+ * configured quota limit (any meter/period) is exceeded.
+ */
+declare interface QuotaFallbackModels {
+  completions?: ModelConfiguration;
+  embeddings?: ModelConfiguration;
+}
+
 /**
  * The Quota policy enables you to set monthly, weekly, daily or hourly quotas on your API.
  *

package/out/types/mcp-gateway/index.d.ts

@@ -103,6 +103,7 @@ declare const EventType: {
   readonly AI_GATEWAY_LATENCY_HISTOGRAM: "ai_gateway_latency_histogram";
   readonly AI_GATEWAY_WARNING_COUNT: "ai_gateway_warning_count";
   readonly AI_GATEWAY_BLOCKED_COUNT: "ai_gateway_blocked_count";
+  readonly AI_GATEWAY_FALLBACK_COUNT: "ai_gateway_fallback_count";
   readonly MCP_REQUEST_RECEIVED: "mcp_request_received";
   readonly MCP_REQUEST_COMPLETED: "mcp_request_completed";
   readonly MCP_REQUEST_REJECTED: "mcp_request_rejected";

package/out/types/mocks/index.d.ts

@@ -84,6 +84,7 @@ declare const EventType: {
   readonly AI_GATEWAY_LATENCY_HISTOGRAM: "ai_gateway_latency_histogram";
   readonly AI_GATEWAY_WARNING_COUNT: "ai_gateway_warning_count";
   readonly AI_GATEWAY_BLOCKED_COUNT: "ai_gateway_blocked_count";
+  readonly AI_GATEWAY_FALLBACK_COUNT: "ai_gateway_fallback_count";
   readonly MCP_REQUEST_RECEIVED: "mcp_request_received";
   readonly MCP_REQUEST_COMPLETED: "mcp_request_completed";
   readonly MCP_REQUEST_REJECTED: "mcp_request_rejected";

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@zuplo/runtime",
   "type": "module",
-  "version": "6.70.62",
+  "version": "6.70.63",
   "repository": "https://github.com/zuplo/zuplo",
   "author": "Zuplo, Inc.",
   "exports": {