@juspay/neurolink 9.30.0 → 9.31.1

package/dist/constants/tokens.js

@@ -89,14 +89,16 @@ export const PROVIDER_TOKEN_LIMITS = {
     },
     /** Google AI model limits */
     GOOGLE_AI: {
-        // Gemini 3 Series (Preview - November 2025)
-        "gemini-3-pro-preview": 8192,
-        "gemini-3-pro-preview-11-2025": 8192,
-        "gemini-3-pro-latest": 8192,
-        // Gemini 3 Flash Series
-        "gemini-3-flash": 65536,
+        // Gemini 3.1 Series (all require -preview suffix)
+        "gemini-3.1-pro-preview": 8192,
+        "gemini-3.1-flash-lite-preview": 65536,
+        "gemini-3.1-flash-image-preview": 8192,
+        "gemini-3.1-pro-preview-customtools": 8192,
+        // Gemini 3 Series
         "gemini-3-flash-preview": 65536,
-        "gemini-3-flash-latest": 65536,
+        "gemini-3-pro-image-preview": 8192,
+        /** @deprecated SHUT DOWN March 9, 2026. Migrate to gemini-3.1-pro-preview. */
+        "gemini-3-pro-preview": 8192,
         // Gemini 2.5 Series
         "gemini-2.5-pro": 8192,
         "gemini-2.5-flash": 8192,
@@ -112,14 +114,16 @@ export const PROVIDER_TOKEN_LIMITS = {
     },
     /** Google Vertex AI model limits */
     VERTEX: {
-        // Gemini 3 Series (Preview - November 2025)
-        "gemini-3-pro-preview": 8192,
-        "gemini-3-pro-preview-11-2025": 8192,
-        "gemini-3-pro-latest": 8192,
-        // Gemini 3 Flash Series
-        "gemini-3-flash": 65536,
+        // Gemini 3.1 Series (all require -preview suffix)
+        "gemini-3.1-pro-preview": 8192,
+        "gemini-3.1-flash-lite-preview": 65536,
+        "gemini-3.1-flash-image-preview": 8192,
+        "gemini-3.1-pro-preview-customtools": 8192,
+        // Gemini 3 Series
         "gemini-3-flash-preview": 65536,
-        "gemini-3-flash-latest": 65536,
+        "gemini-3-pro-image-preview": 8192,
+        /** @deprecated SHUT DOWN March 9, 2026. Migrate to gemini-3.1-pro-preview. */
+        "gemini-3-pro-preview": 8192,
         // Gemini 2.5 Series
         "gemini-2.5-pro": 8192,
         "gemini-2.5-flash": 8192,
@@ -218,10 +222,13 @@ export const CONTEXT_WINDOWS = {
  * For models that require explicit input and output token limits
  */
 export const MODEL_TOKEN_LIMITS = {
+    /** Gemini 3.1 Series */
+    "gemini-3.1-pro-preview": { input: 1048576, output: 65536 },
+    "gemini-3.1-flash-lite-preview": { input: 1048576, output: 65536 },
+    "gemini-3.1-flash-image-preview": { input: 1048576, output: 8192 },
+    "gemini-3.1-pro-preview-customtools": { input: 1048576, output: 65536 },
     /** Gemini 3 Flash Series */
-    "gemini-3-flash": { input: 1000000, output: 65536 },
     "gemini-3-flash-preview": { input: 1000000, output: 65536 },
-    "gemini-3-flash-latest": { input: 1000000, output: 65536 },
 };
 /**
  * Token estimation utilities

package/dist/core/baseProvider.d.ts

@@ -1,4 +1,4 @@
-import type { CoreMessage, LanguageModelV1, Tool } from "ai";
+import type { ModelMessage, LanguageModel, Tool } from "ai";
 import type { AIProviderName } from "../constants/enums.js";
 import type { EvaluationData } from "../index.js";
 import type { NeuroLink } from "../neurolink.js";
@@ -83,9 +83,9 @@ export declare abstract class BaseProvider implements AIProvider {
      * with automatic multimodal detection, eliminating code duplication
      *
      * @param options - Stream options or text generation options
-     * @returns Promise resolving to CoreMessage array ready for AI SDK
+     * @returns Promise resolving to ModelMessage array ready for AI SDK
      */
-    protected buildMessagesForStream(options: StreamOptions | TextGenerationOptions): Promise<CoreMessage[]>;
+    protected buildMessagesForStream(options: StreamOptions | TextGenerationOptions): Promise<ModelMessage[]>;
     /**
      * Execute the generation with AI SDK - delegated to GenerationHandler
      */
@@ -200,13 +200,13 @@ export declare abstract class BaseProvider implements AIProvider {
      * REQUIRED: Every provider MUST implement this method
      * Returns the Vercel AI SDK model instance for this provider
      */
-    protected abstract getAISDKModel(): LanguageModelV1 | Promise<LanguageModelV1>;
+    protected abstract getAISDKModel(): LanguageModel | Promise<LanguageModel>;
     /**
      * Get AI SDK model with middleware applied
      * This method wraps the base model with any configured middleware
      * TODO: Implement global level middlewares that can be used
      */
-    protected getAISDKModelWithMiddleware(options?: TextGenerationOptions | StreamOptions): Promise<LanguageModelV1>;
+    protected getAISDKModelWithMiddleware(options?: TextGenerationOptions | StreamOptions): Promise<LanguageModel>;
     /**
      * Extract middleware options - delegated to Utilities
      */

package/dist/core/baseProvider.js

@@ -436,7 +436,7 @@ export class BaseProvider {
      * with automatic multimodal detection, eliminating code duplication
      *
      * @param options - Stream options or text generation options
-     * @returns Promise resolving to CoreMessage array ready for AI SDK
+     * @returns Promise resolving to ModelMessage array ready for AI SDK
      */
     async buildMessagesForStream(options) {
         return this.messageBuilder.buildMessagesForStream(options);
@@ -609,17 +609,17 @@ export class BaseProvider {
                                 messages: [
                                     { role: "user", content: formattingPrompt },
                                 ],
-                                maxTokens: options.maxTokens || 8192,
+                                maxOutputTokens: options.maxTokens || 8192,
                                 temperature: 0.3,
                                 abortSignal: options.abortSignal,
                                 experimental_telemetry: this.telemetryHandler?.getTelemetryConfig(options, "generate"),
                             });
                             formattedContent = formattedResult.text;
                             usage = {
-                                input: formattedResult.usage?.promptTokens || 0,
-                                output: formattedResult.usage?.completionTokens || 0,
-                                total: (formattedResult.usage?.promptTokens || 0) +
-                                    (formattedResult.usage?.completionTokens || 0),
+                                input: formattedResult.usage?.inputTokens || 0,
+                                output: formattedResult.usage?.outputTokens || 0,
+                                total: (formattedResult.usage?.inputTokens || 0) +
+                                    (formattedResult.usage?.outputTokens || 0),
                             };
                             logger.debug("[VideoAnalysis] Claude formatting complete", {
                                 formattedLength: formattedContent.length,

package/dist/core/constants.js

@@ -7,15 +7,20 @@
 export const IMAGE_GENERATION_MODELS = [
     "gemini-3-pro-image-preview",
     "gemini-2.5-flash-image",
+    "gemini-3.1-flash-image-preview",
 ];
 // PDF Image Generation Models
 // Models that support generating images from PDFs
-export const PDF_IMAGE_GENERATION_MODELS = ["gemini-3-pro-image-preview"];
+export const PDF_IMAGE_GENERATION_MODELS = [
+    "gemini-3-pro-image-preview",
+    "gemini-3.1-flash-image-preview",
+];
 // Global Location Models
 // Models that require global location configuration (uses aiplatform.googleapis.com instead of region-specific endpoints)
 export const GLOBAL_LOCATION_MODELS = [
     "gemini-3-pro-image-preview",
     "gemini-2.5-flash-image",
+    "gemini-3.1-flash-image-preview",
 ];
 // Core AI Generation Defaults
 export const DEFAULT_MAX_TOKENS = undefined; // Unlimited by default - let providers decide their own limits

package/dist/core/dynamicModels.js

@@ -19,9 +19,9 @@ const ModelConfigSchema = z.object({
 const ModelRegistrySchema = z.object({
     version: z.string(),
     lastUpdated: z.string(),
-    models: z.record(z.record(ModelConfigSchema)),
-    aliases: z.record(z.string()).optional(),
-    defaults: z.record(z.string()).optional(),
+    models: z.record(z.string(), z.record(z.string(), ModelConfigSchema)),
+    aliases: z.record(z.string(), z.string()).optional(),
+    defaults: z.record(z.string(), z.string()).optional(),
 });
 /**
  * Dynamic Model Provider
@@ -124,7 +124,9 @@ export class DynamicModelProvider {
             catch (error) {
                 clearTimeout(timeoutId);
                 if (isAbortError(error)) {
-                    throw new Error(`Request timeout after ${timeoutMs}ms`);
+                    throw new Error(`Request timeout after ${timeoutMs}ms`, {
+                        cause: error,
+                    });
                 }
                 throw error;
             }
@@ -179,11 +181,13 @@ export class DynamicModelProvider {
         catch (error) {
             clearTimeout(timeoutId);
             if (isAbortError(error)) {
-                throw new Error(`Localhost health check timeout - server may not be running`);
+                throw new Error(`Localhost health check timeout - server may not be running`, { cause: error });
             }
             // For connection refused, throw a more specific error
             if (error instanceof Error && error.message.includes("ECONNREFUSED")) {
-                throw new Error(`Localhost server not running at ${url}`);
+                throw new Error(`Localhost server not running at ${url}`, {
+                    cause: error,
+                });
             }
             // For other errors, let the main request handle them
             logger.debug(`Health check failed for ${url}, proceeding with main request`, {

package/dist/core/evaluationProviders.js

@@ -209,7 +209,7 @@ export function getProviderPerformanceAnalytics() {
         }
         const avgResponseTime = metrics.responseTime.reduce((a, b) => a + b, 0) /
             metrics.responseTime.length;
-        let recommendation = "";
+        let recommendation;
         if (metrics.successRate > PERFORMANCE_THRESHOLDS.EXCELLENT_SUCCESS_RATE &&
             avgResponseTime < PERFORMANCE_THRESHOLDS.EXCELLENT_RESPONSE_TIME_MS) {
             recommendation = "Excellent - Recommended for production";

package/dist/core/infrastructure/baseError.d.ts

@@ -1,4 +1,4 @@
-export type ErrorCode = string;
+import type { ErrorCode } from "../../types/index.js";
 export declare class NeuroLinkFeatureError extends Error {
     readonly code: ErrorCode;
     readonly feature: string;

package/dist/core/infrastructure/baseFactory.d.ts

@@ -1,9 +1,4 @@
-export type FactoryFunction<TInstance, TConfig> = (config?: TConfig) => Promise<TInstance>;
-export type FactoryRegistration<TInstance, TConfig> = {
-    factory: FactoryFunction<TInstance, TConfig>;
-    aliases: string[];
-    metadata?: Record<string, unknown>;
-};
+import type { FactoryFunction, FactoryRegistration } from "../../types/index.js";
 export declare abstract class BaseFactory<TInstance, TConfig = unknown> {
     protected items: Map<string, FactoryRegistration<TInstance, TConfig>>;
     protected aliasMap: Map<string, string>;

package/dist/core/infrastructure/baseRegistry.d.ts

@@ -1,8 +1,9 @@
-export type RegistryEntry<TItem, TMetadata = unknown> = {
-    factory: () => Promise<TItem>;
-    metadata: TMetadata;
-    instance?: TItem;
-};
+import type { InfraRegistryEntry } from "../../types/index.js";
+/**
+ * Local alias: the canonical type was renamed to InfraRegistryEntry to avoid
+ * collision with other RegistryEntry types in the codebase.
+ */
+export type RegistryEntry<TItem, TMetadata = unknown> = InfraRegistryEntry<TItem, TMetadata>;
 export declare abstract class BaseRegistry<TItem, TMetadata = unknown> {
     protected items: Map<string, RegistryEntry<TItem, TMetadata>>;
     protected initialized: boolean;

package/dist/core/infrastructure/index.d.ts

@@ -1,5 +1,7 @@
-export { createErrorFactory, type ErrorCode, NeuroLinkFeatureError, } from "./baseError.js";
-export { BaseFactory, type FactoryFunction, type FactoryRegistration, } from "./baseFactory.js";
-export { BaseRegistry, type RegistryEntry } from "./baseRegistry.js";
-export { type RetryOptions, withRetry } from "./retry.js";
+export { createErrorFactory, NeuroLinkFeatureError } from "./baseError.js";
+export { BaseFactory } from "./baseFactory.js";
+export { BaseRegistry } from "./baseRegistry.js";
+export { withRetry } from "./retry.js";
+export type { ErrorCode, FactoryFunction, FactoryRegistration, InfraRegistryEntry, InfraRegistryEntry as RegistryEntry, AsyncRetryOptions, } from "../../types/index.js";
+export type { RetryOptions } from "./retry.js";
 export { TypedEventEmitter } from "./typedEventEmitter.js";

package/dist/core/infrastructure/index.js

@@ -1,5 +1,5 @@
-export { createErrorFactory, NeuroLinkFeatureError, } from "./baseError.js";
-export { BaseFactory, } from "./baseFactory.js";
+export { createErrorFactory, NeuroLinkFeatureError } from "./baseError.js";
+export { BaseFactory } from "./baseFactory.js";
 export { BaseRegistry } from "./baseRegistry.js";
 export { withRetry } from "./retry.js";
 export { TypedEventEmitter } from "./typedEventEmitter.js";

package/dist/core/modelConfiguration.js

@@ -633,7 +633,9 @@ export class ModelConfigurationManager {
         catch (error) {
             const errorMessage = error instanceof Error ? error.message : String(error);
             logger.error(`Failed to load configurations from file: ${errorMessage}`);
-            throw new Error(`Configuration loading failed: ${errorMessage}`);
+            throw new Error(`Configuration loading failed: ${errorMessage}`, {
+                cause: error,
+            });
         }
     }
     /**

package/dist/core/modules/GenerationHandler.d.ts

@@ -12,7 +12,7 @@
  *
  * @module core/modules/GenerationHandler
  */
-import type { CoreMessage, LanguageModelV1, Tool } from "ai";
+import type { LanguageModel, ModelMessage, Tool } from "ai";
 import { generateText } from "ai";
 import type { AIProviderName, EnhancedGenerateResult, StandardRecord, TextGenerationOptions } from "../../types/index.js";
 /**
@@ -37,10 +37,10 @@ export declare class GenerationHandler {
     /**
      * Execute the generation with AI SDK
      */
-    executeGeneration(model: LanguageModelV1, messages: CoreMessage[], tools: Record<string, Tool>, options: TextGenerationOptions): Promise<Awaited<ReturnType<typeof generateText>>>;
+    executeGeneration(model: LanguageModel, messages: ModelMessage[], tools: Record<string, Tool>, options: TextGenerationOptions): Promise<Awaited<ReturnType<typeof generateText>>>;
     /**
      * Extract cache metrics from provider metadata (e.g. Anthropic's providerMetadata.anthropic)
-     * The Vercel AI SDK's LanguageModelUsage only has promptTokens/completionTokens/totalTokens.
+     * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
     private extractCacheMetricsFromProviderMetadata;

package/dist/core/modules/GenerationHandler.js

@@ -13,7 +13,8 @@
  * @module core/modules/GenerationHandler
  */
 import { SpanKind, SpanStatusCode } from "@opentelemetry/api";
-import { generateText, NoObjectGeneratedError, Output } from "ai";
+import { generateText, NoObjectGeneratedError, Output, stepCountIs } from "ai";
+import { getModelId } from "../../providers/providerTypeUtils.js";
 import { tracers } from "../../telemetry/tracers.js";
 import { logger } from "../../utils/logger.js";
 import { calculateCost } from "../../utils/pricing.js";
@@ -99,14 +100,17 @@ export class GenerationHandler {
             messages,
             ...(shouldUseTools &&
                 Object.keys(toolsWithCache).length > 0 && { tools: toolsWithCache }),
-            maxSteps: options.maxSteps ?? DEFAULT_MAX_STEPS,
+            stopWhen: stepCountIs(options.maxSteps ?? DEFAULT_MAX_STEPS),
             ...(shouldUseTools &&
                 options.toolChoice && { toolChoice: options.toolChoice }),
             ...(options.prepareStep && {
-                experimental_prepareStep: options.prepareStep,
+                experimental_prepareStep: ((stepOptions) => options.prepareStep({
+                    ...stepOptions,
+                    maxSteps: options.maxSteps ?? DEFAULT_MAX_STEPS,
+                })),
             }),
             temperature: options.temperature,
-            maxTokens: options.maxTokens,
+            maxOutputTokens: options.maxTokens,
             maxRetries: 0, // NL11: Disable AI SDK's invisible internal retries; we handle retries with OTel instrumentation
             abortSignal: options.abortSignal,
             ...(useStructuredOutput &&
@@ -173,13 +177,13 @@ export class GenerationHandler {
             span.setAttribute("neurolink.structured_output", useStructuredOutput);
             span.setAttribute("neurolink.tool_count", toolCount);
             span.setAttribute("neurolink.message_count", messages.length);
-            span.setAttribute("gen_ai.request.model", model.modelId || this.modelName || "unknown");
+            span.setAttribute("gen_ai.request.model", getModelId(model, this.modelName || "unknown"));
             const requestId = options.requestId ||
                 options.context?.requestId ||
                 "unknown";
             logger.info("[GenerationHandler] Calling generateText", {
                 requestId,
-                model: model.modelId || "unknown",
+                model: getModelId(model),
                 messageCount: messages.length,
                 toolCount,
                 maxSteps: options.maxSteps,
@@ -189,7 +193,7 @@ export class GenerationHandler {
                 try {
                     logger.debug("[Observability] Full generateText parameters", {
                         requestId,
-                        model: model.modelId || "unknown",
+                        model: getModelId(model),
                         messageCount: messages.length,
                         messages: messages.map((msg, i) => ({
                             index: i,
@@ -246,14 +250,14 @@ export class GenerationHandler {
                 }
                 // Set token usage and completion attributes on span
                 if (result.usage) {
-                    span.setAttribute("gen_ai.usage.input_tokens", result.usage.promptTokens || 0);
-                    span.setAttribute("gen_ai.usage.output_tokens", result.usage.completionTokens || 0);
+                    span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
+                    span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
                     // Cost on span so users can query "what did this trace cost?"
                     const cost = calculateCost(this.providerName, this.modelName, {
-                        input: result.usage.promptTokens || 0,
-                        output: result.usage.completionTokens || 0,
-                        total: (result.usage.promptTokens || 0) +
-                            (result.usage.completionTokens || 0),
+                        input: result.usage.inputTokens || 0,
+                        output: result.usage.outputTokens || 0,
+                        total: (result.usage.inputTokens || 0) +
+                            (result.usage.outputTokens || 0),
                     });
                     span.setAttribute("neurolink.cost", cost ?? 0);
                 }
@@ -297,13 +301,13 @@ export class GenerationHandler {
                         responseChars: result.text?.length || 0,
                     });
                     if (result.usage) {
-                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.promptTokens || 0);
-                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.completionTokens || 0);
+                        span.setAttribute("gen_ai.usage.input_tokens", result.usage.inputTokens || 0);
+                        span.setAttribute("gen_ai.usage.output_tokens", result.usage.outputTokens || 0);
                         const fallbackCost = calculateCost(this.providerName, this.modelName, {
-                            input: result.usage.promptTokens || 0,
-                            output: result.usage.completionTokens || 0,
-                            total: (result.usage.promptTokens || 0) +
-                                (result.usage.completionTokens || 0),
+                            input: result.usage.inputTokens || 0,
+                            output: result.usage.outputTokens || 0,
+                            total: (result.usage.inputTokens || 0) +
+                                (result.usage.outputTokens || 0),
                         });
                         span.setAttribute("neurolink.cost", fallbackCost ?? 0);
                     }
@@ -327,13 +331,11 @@ export class GenerationHandler {
     }
     /**
      * Extract cache metrics from provider metadata (e.g. Anthropic's providerMetadata.anthropic)
-     * The Vercel AI SDK's LanguageModelUsage only has promptTokens/completionTokens/totalTokens.
+     * The AI SDK's LanguageModelUsage only has inputTokens/outputTokens.
      * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
      */
     extractCacheMetricsFromProviderMetadata(generateResult) {
-        const providerMeta = generateResult
-            .providerMetadata ||
-            generateResult.experimental_providerMetadata;
+        const providerMeta = generateResult.providerMetadata;
         if (!providerMeta) {
             return {};
         }
@@ -421,22 +423,12 @@ export class GenerationHandler {
                         const trRecord = toolResult;
                         const toolName = trRecord.toolName || "unknown";
                         const toolId = trRecord.toolCallId || trRecord.id;
-                        let toolArgs = {};
-                        if (trRecord.args) {
-                            toolArgs = trRecord.args;
-                        }
-                        else if (trRecord.arguments) {
-                            toolArgs = trRecord.arguments;
-                        }
-                        else if (trRecord.parameters) {
-                            toolArgs = trRecord.parameters;
-                        }
-                        else if (trRecord.input) {
-                            toolArgs = trRecord.input;
-                        }
-                        else {
-                            toolArgs = toolCallArgsMap.get(toolId || toolName) || {};
-                        }
+                        const toolArgs = trRecord.args ??
+                            trRecord.arguments ??
+                            trRecord.parameters ??
+                            trRecord.input ??
+                            toolCallArgsMap.get(toolId || toolName) ??
+                            {};
                         toolExecutions.push({
                             name: toolName,
                             input: toolArgs,
@@ -518,12 +510,28 @@ export class GenerationHandler {
                 }
             }
         }
+        // Extract reasoning from AI SDK response (Anthropic thinking, Gemini thought, OpenAI o1)
+        // Handle both string and array (AI SDK v5 returns string, v6 returns ReasoningOutput[])
+        const rawReasoning = generateResult.reasoning;
+        const reasoning = rawReasoning
+            ? typeof rawReasoning === "string"
+                ? rawReasoning
+                : Array.isArray(rawReasoning)
+                    ? rawReasoning
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        .map((r) => typeof r === "string" ? r : (r.text ?? JSON.stringify(r)))
+                        .join("\n")
+                    : String(rawReasoning)
+            : undefined;
+        const reasoningTokens = usage.reasoning ?? undefined;
         return {
             content,
             usage,
             finishReason: generateResult.finishReason,
             provider: this.providerName,
             model: this.modelName,
+            reasoning,
+            reasoningTokens,
             toolCalls: generateResult.toolCalls
                 ? generateResult.toolCalls.map((tc) => ({
                     toolCallId: tc.toolCallId || "unknown",
@@ -531,7 +539,7 @@ export class GenerationHandler {
                     args: tc.args || {},
                 }))
                 : [],
-            toolResults: generateResult.toolResults || [],
+            toolResults: generateResult.toolResults ?? [],
             toolsUsed,
             toolExecutions,
             availableTools: Object.keys(tools).map((name) => {
@@ -539,7 +547,7 @@ export class GenerationHandler {
                 return {
                     name,
                     description: tool.description || "No description available",
-                    parameters: tool.parameters || {},
+                    parameters: tool.inputSchema || {},
                     server: tool.serverId || "direct",
                 };
             }),

package/dist/core/modules/MessageBuilder.d.ts

@@ -8,11 +8,11 @@
  * - Building messages from text generation options
  * - Building messages from stream options
  * - Multimodal input detection
- * - Message format conversion (to CoreMessage[])
+ * - Message format conversion (to ModelMessage[])
  *
  * @module core/modules/MessageBuilder
  */
-import type { CoreMessage } from "ai";
+import type { ModelMessage } from "ai";
 import type { AIProviderName, TextGenerationOptions } from "../../types/index.js";
 import type { StreamOptions } from "../../types/streamTypes.js";
 /**
@@ -26,14 +26,14 @@ export declare class MessageBuilder {
      * Build messages array for generation
      * Detects multimodal input and routes to appropriate message builder
      */
-    buildMessages(options: TextGenerationOptions): Promise<CoreMessage[]>;
+    buildMessages(options: TextGenerationOptions): Promise<ModelMessage[]>;
     /**
      * Build messages array for streaming operations
      * This is a protected helper method that providers can use to build messages
      * with automatic multimodal detection, eliminating code duplication
      *
      * @param options - Stream options or text generation options
-     * @returns Promise resolving to CoreMessage array ready for AI SDK
+     * @returns Promise resolving to ModelMessage array ready for AI SDK
      */
-    buildMessagesForStream(options: StreamOptions | TextGenerationOptions): Promise<CoreMessage[]>;
+    buildMessagesForStream(options: StreamOptions | TextGenerationOptions): Promise<ModelMessage[]>;
 }

package/dist/core/modules/MessageBuilder.js

@@ -8,7 +8,7 @@
  * - Building messages from text generation options
  * - Building messages from stream options
  * - Multimodal input detection
- * - Message format conversion (to CoreMessage[])
+ * - Message format conversion (to ModelMessage[])
  *
  * @module core/modules/MessageBuilder
  */
@@ -129,8 +129,7 @@ export class MessageBuilder {
                     return {
                         role: msg.role,
                         content: msg.content.map((item) => {
-                            const itemProviderOptions = item
-                                .providerOptions;
+                            const itemProviderOptions = item.providerOptions;
                             if (item.type === "text") {
                                 return {
                                     type: "text",
@@ -169,7 +168,7 @@ export class MessageBuilder {
      * with automatic multimodal detection, eliminating code duplication
      *
      * @param options - Stream options or text generation options
-     * @returns Promise resolving to CoreMessage array ready for AI SDK
+     * @returns Promise resolving to ModelMessage array ready for AI SDK
      */
     async buildMessagesForStream(options) {
         return withSpan({
@@ -238,8 +237,7 @@ export class MessageBuilder {
                     return {
                         role: msg.role,
                         content: msg.content.map((item) => {
-                            const itemProviderOptions = item
-                                .providerOptions;
+                            const itemProviderOptions = item.providerOptions;
                             if (item.type === "text") {
                                 return {
                                     type: "text",

package/dist/core/modules/StreamHandler.js

@@ -19,6 +19,7 @@ import { validateStreamOptions as validateStreamOpts, ValidationError, createVal
 import { STEP_LIMITS } from "../constants.js";
 import { createAnalytics } from "../analytics.js";
 import { nanoid } from "nanoid";
+import { NoOutputGeneratedError } from "ai";
 /**
  * StreamHandler class - Handles streaming operations for AI providers
  */
@@ -89,20 +90,33 @@ export class StreamHandler {
             let totalBytes = 0;
             const streamStart = Date.now();
             let firstChunkTime;
-            for await (const chunk of result.textStream) {
-                chunkCount++;
-                totalBytes += chunk.length;
-                if (!firstChunkTime) {
-                    firstChunkTime = Date.now();
-                    const activeSpan = trace.getSpan(otelContext.active());
-                    if (activeSpan) {
-                        activeSpan.addEvent("stream.first_chunk", {
-                            "stream.ttfc_ms": firstChunkTime - streamStart,
-                            "stream.provider": providerName,
-                        });
+            try {
+                for await (const chunk of result.textStream) {
+                    chunkCount++;
+                    totalBytes += chunk.length;
+                    if (!firstChunkTime) {
+                        firstChunkTime = Date.now();
+                        const activeSpan = trace.getSpan(otelContext.active());
+                        if (activeSpan) {
+                            activeSpan.addEvent("stream.first_chunk", {
+                                "stream.ttfc_ms": firstChunkTime - streamStart,
+                                "stream.provider": providerName,
+                            });
+                        }
                     }
+                    yield { content: chunk };
+                }
+            }
+            catch (error) {
+                // AI SDK v6 throws NoOutputGeneratedError when the stream produces no output
+                // (e.g. empty response, model refusal with no text). Treat as an empty stream
+                // rather than crashing the process with an unhandled rejection.
+                if (NoOutputGeneratedError.isInstance(error)) {
+                    logger.warn(`${providerName}: Stream produced no output (NoOutputGeneratedError), returning empty stream`);
+                }
+                else {
+                    throw error;
                 }
-                yield { content: chunk };
             }
             // Record completion metrics on the active span
             const activeSpan = trace.getSpan(otelContext.active());

package/dist/core/modules/TelemetryHandler.d.ts

@@ -36,9 +36,8 @@ export declare class TelemetryHandler {
      * Record performance metrics for a generation
      */
     recordPerformanceMetrics(usage: {
-        promptTokens: number;
-        completionTokens: number;
-        totalTokens: number;
+        inputTokens: number | undefined;
+        outputTokens: number | undefined;
     } | undefined, responseTime: number): Promise<void>;
     /**
      * Calculate actual cost based on token usage and provider configuration.
@@ -53,9 +52,8 @@ export declare class TelemetryHandler {
      * on Vertex AI ($0.000060 vs $0.106895 for the same request).
      */
     calculateActualCost(usage: {
-        promptTokens?: number;
-        completionTokens?: number;
-        totalTokens?: number;
+        inputTokens?: number | undefined;
+        outputTokens?: number | undefined;
     }): Promise<number>;
     /**
      * Create telemetry configuration for Vercel AI SDK experimental_telemetry