@opengeni/config 0.10.1 → 0.10.3

package/dist/index.d.ts

@@ -190,6 +190,7 @@ declare const SettingsSchema: z.ZodObject<{
     openaiReasoningEffort: z.ZodDefault<z.ZodEnum<{
         high: "high";
         low: "low";
+        max: "max";
         medium: "medium";
         minimal: "minimal";
         none: "none";
@@ -528,6 +529,7 @@ export declare const ModelCapabilitiesV1Schema: z.ZodObject<{
         efforts: z.ZodArray<z.ZodEnum<{
             high: "high";
             low: "low";
+            max: "max";
             medium: "medium";
             minimal: "minimal";
             none: "none";
@@ -536,6 +538,7 @@ export declare const ModelCapabilitiesV1Schema: z.ZodObject<{
         defaultEffort: z.ZodNullable<z.ZodEnum<{
             high: "high";
             low: "low";
+            max: "max";
             medium: "medium";
             minimal: "minimal";
             none: "none";
@@ -743,6 +746,7 @@ declare const RegistryProviderSchema: z.ZodObject<{
                 efforts: z.ZodArray<z.ZodEnum<{
                     high: "high";
                     low: "low";
+                    max: "max";
                     medium: "medium";
                     minimal: "minimal";
                     none: "none";
@@ -751,6 +755,7 @@ declare const RegistryProviderSchema: z.ZodObject<{
                 defaultEffort: z.ZodNullable<z.ZodEnum<{
                     high: "high";
                     low: "low";
+                    max: "max";
                     medium: "medium";
                     minimal: "minimal";
                     none: "none";
@@ -940,7 +945,7 @@ export interface ConfiguredModel {
     capabilities: ModelCapabilitiesV1;
     requestPolicy?: {
         gateway: {
-            only: [string];
+            only: [string, ...string[]];
             caching: "auto" | "none";
         };
     };
@@ -954,26 +959,59 @@ export interface ConfiguredModel {
     hostedWebSearch: boolean;
 }
 export declare const VERCEL_AI_GATEWAY_BASE_URL: "https://ai-gateway.vercel.sh/v1";
+export declare const VERCEL_AI_GATEWAY_AI_SDK_BASE_URL: "https://ai-gateway.vercel.sh/v4/ai";
 export declare const OPENGENI_GATEWAY_PROVIDER_ID: "opengeni-gateway";
 export declare const WORKSPACE_GATEWAY_PROVIDER_ID: "workspace-gateway";
 export declare const WORKSPACE_GATEWAY_MODEL_ID_PREFIX: "workspace-gateway/";
 export declare const VERCEL_AI_GATEWAY_CONNECTION_DOMAIN: "ai-gateway.vercel.sh";
 export declare const VERCEL_AI_GATEWAY_CONNECTION_ROLE: "vercel_ai_gateway";
+export declare const CODEX_REALTIME_MODEL_ID: "gpt-live-1-boulder-alpha";
+export declare const OPENGENI_REALTIME_MODEL_ID_PREFIX: "opengeni-gateway/";
+export declare const WORKSPACE_REALTIME_MODEL_ID_PREFIX: "workspace-gateway/";
+/** Curated voice models exposed through AI Gateway's normalized realtime API. */
+export declare const AI_GATEWAY_REALTIME_MODELS: {
+    readonly openaiRealtime21: {
+        readonly upstreamModelId: "openai/gpt-realtime-2.1";
+        readonly managedModelId: "opengeni-gateway/openai/gpt-realtime-2.1";
+        readonly workspaceModelId: "workspace-gateway/openai/gpt-realtime-2.1";
+        readonly label: "GPT Realtime 2.1";
+        readonly description: "Best overall voice intelligence";
+    };
+    readonly openaiRealtimeMini: {
+        readonly upstreamModelId: "openai/gpt-realtime-mini";
+        readonly managedModelId: "opengeni-gateway/openai/gpt-realtime-mini";
+        readonly workspaceModelId: "workspace-gateway/openai/gpt-realtime-mini";
+        readonly label: "GPT Realtime Mini";
+        readonly description: "Faster, lighter live voice";
+    };
+    readonly grokVoiceThinkFast20: {
+        readonly upstreamModelId: "xai/grok-voice-think-fast-2.0";
+        readonly managedModelId: "opengeni-gateway/xai/grok-voice-think-fast-2.0";
+        readonly workspaceModelId: "workspace-gateway/xai/grok-voice-think-fast-2.0";
+        readonly label: "Grok Voice Think Fast 2.0";
+        readonly description: "Fast, natural xAI voice";
+    };
+};
+export type AiGatewayRealtimeModel = (typeof AI_GATEWAY_REALTIME_MODELS)[keyof typeof AI_GATEWAY_REALTIME_MODELS];
+export declare function resolveAiGatewayRealtimeModel(modelId: string): {
+    source: "managed" | "workspace";
+    upstreamModelId: string;
+} | null;
 export declare const OPENGENI_GATEWAY_MODELS: {
     readonly deepseek: {
         readonly productId: "deepseek-v4-flash-0731";
         readonly workspaceProductId: "workspace-gateway/deepseek-v4-flash-0731";
         readonly upstreamModelId: "deepseek/deepseek-v4-flash-0731";
         readonly label: "DeepSeek V4 Flash 0731";
-        readonly provider: "deepinfra";
+        readonly providers: readonly ["baseten", "novita", "deepinfra"];
         readonly implicitCaching: true;
     };
     readonly kimi: {
-        readonly productId: "kimi-k3-fast";
-        readonly workspaceProductId: "workspace-gateway/kimi-k3-fast";
-        readonly upstreamModelId: "moonshotai/kimi-k3-fast";
-        readonly label: "Kimi K3 Fast";
-        readonly provider: "wafer";
+        readonly productId: "kimi-k3";
+        readonly workspaceProductId: "workspace-gateway/kimi-k3";
+        readonly upstreamModelId: "moonshotai/kimi-k3";
+        readonly label: "Kimi K3";
+        readonly providers: readonly ["baseten", "fireworks"];
         readonly implicitCaching: true;
     };
 };
@@ -1184,6 +1222,14 @@ export declare function configuredEntitlements(settings: Settings): Entitlements
 export declare function calculateModelUsageCostMicros(settings: Settings, model: string, usage: ModelUsageInput, options?: {
     latencyMode?: LatencyMode;
 }): number;
+/**
+ * Convert AI Gateway's exact USD inference cost to OpenGeni credit micros and
+ * apply the configured model margin. Decimal arithmetic is integer-only so a
+ * sub-micro provider charge cannot be lost to floating-point rounding.
+ */
+export declare function calculateGatewayReportedCostMicros(settings: Settings, model: string, inferenceCostUsd: string, options?: {
+    inputTokens?: number;
+}): number;
 export declare function configuredAllowedReasoningEfforts(settings: Settings): Array<z.infer<typeof ReasoningEffort>>;
 /**
  * Decodes OPENGENI_ENVIRONMENTS_ENCRYPTION_KEY (base64, exactly 32 bytes) for

package/dist/index.js

@@ -367,7 +367,7 @@ var SettingsSchema = z.object({
   // match the UI danger flip (UsageBar danger at pct >= 90). OPENGENI_CODEX_ROTATION_NEAR_EXHAUSTION_PCT.
   codexRotationNearExhaustionPct: z.coerce.number().int().min(1).max(100).default(90),
   openaiReasoningEffort: ReasoningEffort.default("low"),
-  openaiAllowedReasoningEfforts: z.string().default("low,medium,high,xhigh"),
+  openaiAllowedReasoningEfforts: z.string().default("low,medium,high,xhigh,max"),
   openaiResponsesTransport: z.enum(["http", "websocket"]).default("http"),
   // Provider-assigned item ids (rs_/msg_/fc_…) in Responses API input are
   // resolved against the provider's server-side response store. That store is
@@ -1106,26 +1106,64 @@ var IntegrationOAuthClientConfigSchema = z.object({
   tokenEndpointAuthMethod: z.enum(["none", "client_secret_post", "client_secret_basic"]).default("none")
 });
 var VERCEL_AI_GATEWAY_BASE_URL = "https://ai-gateway.vercel.sh/v1";
+var VERCEL_AI_GATEWAY_AI_SDK_BASE_URL = "https://ai-gateway.vercel.sh/v4/ai";
 var OPENGENI_GATEWAY_PROVIDER_ID = "opengeni-gateway";
 var WORKSPACE_GATEWAY_PROVIDER_ID = "workspace-gateway";
 var WORKSPACE_GATEWAY_MODEL_ID_PREFIX = "workspace-gateway/";
 var VERCEL_AI_GATEWAY_CONNECTION_DOMAIN = "ai-gateway.vercel.sh";
 var VERCEL_AI_GATEWAY_CONNECTION_ROLE = "vercel_ai_gateway";
+var CODEX_REALTIME_MODEL_ID = "gpt-live-1-boulder-alpha";
+var OPENGENI_REALTIME_MODEL_ID_PREFIX = "opengeni-gateway/";
+var WORKSPACE_REALTIME_MODEL_ID_PREFIX = "workspace-gateway/";
+var AI_GATEWAY_REALTIME_MODELS = {
+  openaiRealtime21: {
+    upstreamModelId: "openai/gpt-realtime-2.1",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-2.1`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-2.1`,
+    label: "GPT Realtime 2.1",
+    description: "Best overall voice intelligence"
+  },
+  openaiRealtimeMini: {
+    upstreamModelId: "openai/gpt-realtime-mini",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-mini`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}openai/gpt-realtime-mini`,
+    label: "GPT Realtime Mini",
+    description: "Faster, lighter live voice"
+  },
+  grokVoiceThinkFast20: {
+    upstreamModelId: "xai/grok-voice-think-fast-2.0",
+    managedModelId: `${OPENGENI_REALTIME_MODEL_ID_PREFIX}xai/grok-voice-think-fast-2.0`,
+    workspaceModelId: `${WORKSPACE_REALTIME_MODEL_ID_PREFIX}xai/grok-voice-think-fast-2.0`,
+    label: "Grok Voice Think Fast 2.0",
+    description: "Fast, natural xAI voice"
+  }
+};
+function resolveAiGatewayRealtimeModel(modelId) {
+  for (const model of Object.values(AI_GATEWAY_REALTIME_MODELS)) {
+    if (model.managedModelId === modelId) {
+      return { source: "managed", upstreamModelId: model.upstreamModelId };
+    }
+    if (model.workspaceModelId === modelId) {
+      return { source: "workspace", upstreamModelId: model.upstreamModelId };
+    }
+  }
+  return null;
+}
 var OPENGENI_GATEWAY_MODELS = {
   deepseek: {
     productId: "deepseek-v4-flash-0731",
     workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}deepseek-v4-flash-0731`,
     upstreamModelId: "deepseek/deepseek-v4-flash-0731",
     label: "DeepSeek V4 Flash 0731",
-    provider: "deepinfra",
+    providers: ["baseten", "novita", "deepinfra"],
     implicitCaching: true
   },
   kimi: {
-    productId: "kimi-k3-fast",
-    workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}kimi-k3-fast`,
-    upstreamModelId: "moonshotai/kimi-k3-fast",
-    label: "Kimi K3 Fast",
-    provider: "wafer",
+    productId: "kimi-k3",
+    workspaceProductId: `${WORKSPACE_GATEWAY_MODEL_ID_PREFIX}kimi-k3`,
+    upstreamModelId: "moonshotai/kimi-k3",
+    label: "Kimi K3",
+    providers: ["baseten", "fireworks"],
     implicitCaching: true
   }
 };
@@ -1188,24 +1226,24 @@ var defaultModelPricing = {
       }
     ]
   },
-  // Vercel AI Gateway endpoint prices, provider-pinned in the runtime.
-  // Snapshot: 2026-08-02. Both pinned routes returned discounted implicit
-  // cache reads in live Gateway responses. Wafer/Kimi reported $0.45/M even
-  // though the provider-discovery flag currently says otherwise; bill from
-  // the response-backed rate, not that inconsistent boolean.
+  // Conservative Vercel AI Gateway fallback prices. Normal managed Gateway
+  // billing uses the exact response Gateway `cost` / `inferenceCost` and applies
+  // the same margin. These token rates are used only if that
+  // metadata is absent. DeepSeek therefore carries the highest approved route
+  // (Novita); both approved Kimi routes have the same list price.
   [OPENGENI_GATEWAY_MODELS.deepseek.productId]: {
     default: {
-      inputMicrosPerMillionTokens: 9e4,
-      cachedInputMicrosPerMillionTokens: 18e3,
-      outputMicrosPerMillionTokens: 18e4,
+      inputMicrosPerMillionTokens: 14e4,
+      cachedInputMicrosPerMillionTokens: 28e3,
+      outputMicrosPerMillionTokens: 28e4,
       marginBps: 2500
     }
   },
   [OPENGENI_GATEWAY_MODELS.kimi.productId]: {
     default: {
-      inputMicrosPerMillionTokens: 45e5,
-      cachedInputMicrosPerMillionTokens: 45e4,
-      outputMicrosPerMillionTokens: 225e5,
+      inputMicrosPerMillionTokens: 3e6,
+      cachedInputMicrosPerMillionTokens: 3e5,
+      outputMicrosPerMillionTokens: 15e6,
       marginBps: 2500
     }
   },
@@ -1809,7 +1847,7 @@ function gatewayRequestPolicyForUpstreamModel(upstreamModelId) {
   }
   return {
     gateway: {
-      only: [model.provider],
+      only: [...model.providers],
       caching: model.implicitCaching ? "auto" : "none"
     }
   };
@@ -1828,26 +1866,29 @@ function gatewayModelCapabilities(settings, input) {
       sse: { upstream: "supported", runnable: true }
     },
     promptCaching: input.implicitCaching ? { upstream: "supported", runnable: true, mode: "implicit" } : { upstream: "unsupported", runnable: false, mode: "none" },
-    // "Fast" is part of Kimi's product name, not OpenGeni's separately billed
-    // latency mode. Both Gateway products expose only standard here.
+    // Both Gateway products expose one reviewed route policy and no separately
+    // billed latency mode.
     latencyModes: [{ id: "standard", upstream: "supported", runnable: true }]
   });
 }
 function gatewayRegistryProvider(settings, input) {
   const workspace = input.kind === "vercel-gateway-workspace";
-  const models = [OPENGENI_GATEWAY_MODELS.deepseek, OPENGENI_GATEWAY_MODELS.kimi].map((model) => ({
-    id: workspace ? model.workspaceProductId : model.productId,
-    upstreamModelId: model.upstreamModelId,
-    label: model.label,
-    capabilities: gatewayModelCapabilities(settings, {
-      implicitCaching: model.implicitCaching,
-      vision: model === OPENGENI_GATEWAY_MODELS.kimi
-    }),
-    contextWindowTokens: 1e6,
-    effectiveContextWindowTokens: 9e5,
-    autoCompactTokenLimit: 85e4,
-    toolOutputTruncationTokens: settings.modelToolOutputTruncationTokens
-  }));
+  const models = [OPENGENI_GATEWAY_MODELS.deepseek, OPENGENI_GATEWAY_MODELS.kimi].map((model) => {
+    const kimi = model === OPENGENI_GATEWAY_MODELS.kimi;
+    return {
+      id: workspace ? model.workspaceProductId : model.productId,
+      upstreamModelId: model.upstreamModelId,
+      label: model.label,
+      capabilities: gatewayModelCapabilities(settings, {
+        implicitCaching: model.implicitCaching,
+        vision: kimi
+      }),
+      contextWindowTokens: 1e6,
+      effectiveContextWindowTokens: 9e5,
+      autoCompactTokenLimit: 85e4,
+      toolOutputTruncationTokens: settings.modelToolOutputTruncationTokens
+    };
+  });
   return {
     kind: input.kind,
     id: workspace ? WORKSPACE_GATEWAY_PROVIDER_ID : OPENGENI_GATEWAY_PROVIDER_ID,
@@ -2498,6 +2539,28 @@ function calculateModelUsageCostMicros(settings, model, usage, options) {
   }
   return total;
 }
+function calculateGatewayReportedCostMicros(settings, model, inferenceCostUsd, options) {
+  const schedule = configuredModelPricingSchedules(settings)[model];
+  if (!schedule) {
+    throw new Error(`Missing model pricing for ${model}`);
+  }
+  const pricing = selectModelPricing(schedule, positiveInt(options?.inputTokens));
+  const match = /^(0|[1-9]\d*)(?:\.(\d{1,18}))?$/.exec(inferenceCostUsd);
+  if (!match) {
+    throw new Error("Invalid AI Gateway inference cost");
+  }
+  const fraction = match[2] ?? "";
+  const decimalDigits = BigInt(`${match[1]}${fraction}`);
+  const decimalScale = 10n ** BigInt(fraction.length);
+  const marginBps = BigInt(1e4 + (pricing.marginBps ?? 0));
+  const numerator = decimalDigits * 1000000n * marginBps;
+  const denominator = decimalScale * 10000n;
+  const micros = (numerator + denominator - 1n) / denominator;
+  if (micros > BigInt(Number.MAX_SAFE_INTEGER)) {
+    throw new Error("AI Gateway inference cost exceeds the supported billing range");
+  }
+  return Number(micros);
+}
 function configuredAllowedReasoningEfforts(settings) {
   return uniqueValues([
     settings.openaiReasoningEffort,
@@ -3431,6 +3494,8 @@ function delay(ms) {
 }
 export {
   AGENT_INSTRUCTIONS_CORE_PLACEHOLDER,
+  AI_GATEWAY_REALTIME_MODELS,
+  CODEX_REALTIME_MODEL_ID,
   CapabilityStateV1Schema,
   CapabilitySupportV1,
   DEFAULT_AGENT_INSTRUCTIONS,
@@ -3440,17 +3505,21 @@ export {
   ModelProviderApi,
   OPENGENI_GATEWAY_MODELS,
   OPENGENI_GATEWAY_PROVIDER_ID,
+  OPENGENI_REALTIME_MODEL_ID_PREFIX,
   RegistryProviderKind,
   SANDBOX_REQUIRED_ENV,
   SocialOAuthClientConfigSchema,
+  VERCEL_AI_GATEWAY_AI_SDK_BASE_URL,
   VERCEL_AI_GATEWAY_BASE_URL,
   VERCEL_AI_GATEWAY_CONNECTION_DOMAIN,
   VERCEL_AI_GATEWAY_CONNECTION_ROLE,
   WORKSPACE_GATEWAY_MODEL_ID_PREFIX,
   WORKSPACE_GATEWAY_PROVIDER_ID,
+  WORKSPACE_REALTIME_MODEL_ID_PREFIX,
   applyGitAuthPointerEnvironment,
   assertTurnExecutionPolicyMatchesConfigV1,
   builtinProviderId,
+  calculateGatewayReportedCostMicros,
   calculateModelUsageCostMicros,
   canonicalizeConfiguredModelId,
   collectGitIdentityEnvironment,
@@ -3487,6 +3556,7 @@ export {
   policyProviderIdForModel,
   productLabelForModelId,
   requiredSandboxEnvForBackend,
+  resolveAiGatewayRealtimeModel,
   resolveEnrollmentSigningSecret,
   resolveFirstPartyDelegationSecret,
   resolveModelProvider,