@juspay/neurolink 9.63.0 → 9.64.0

package/dist/cli/factories/commandFactory.d.ts

@@ -3,6 +3,20 @@ import type { CommandModule } from "yargs";
  * CLI Command Factory for generate commands
  */
 export declare class CLICommandFactory {
+    /**
+     * Normalize loop session variables before merging them into provider options.
+     *
+     * The CLI loop schema models some fields (e.g. `stopSequences`,
+     * `enabledToolNames`) as a single comma-separated string for ergonomic
+     * input, but providers expect `string[]`. Without conversion,
+     * `set stopSequences a,b` would be sent as one stop token "a,b" instead
+     * of two ("a", "b"); `set enabledToolNames read,write` would be cast to
+     * a `string[]` containing the single literal "read,write" and silently
+     * filter out every tool. This helper splits and trims those fields so
+     * the spread into `enhancedOptions` produces the correct shape across
+     * generate / batch / stream paths.
+     */
+    private static normalizeLoopSessionVariables;
     private static readonly commonOptions;
     private static buildOptions;
     private static processCliImages;

package/dist/cli/factories/commandFactory.js

@@ -28,6 +28,35 @@ import { SageMakerCommandFactory } from "./sagemakerCommandFactory.js";
  * CLI Command Factory for generate commands
  */
 export class CLICommandFactory {
+    /**
+     * Normalize loop session variables before merging them into provider options.
+     *
+     * The CLI loop schema models some fields (e.g. `stopSequences`,
+     * `enabledToolNames`) as a single comma-separated string for ergonomic
+     * input, but providers expect `string[]`. Without conversion,
+     * `set stopSequences a,b` would be sent as one stop token "a,b" instead
+     * of two ("a", "b"); `set enabledToolNames read,write` would be cast to
+     * a `string[]` containing the single literal "read,write" and silently
+     * filter out every tool. This helper splits and trims those fields so
+     * the spread into `enhancedOptions` produces the correct shape across
+     * generate / batch / stream paths.
+     */
+    static normalizeLoopSessionVariables(vars) {
+        const normalized = { ...vars };
+        if (typeof normalized.stopSequences === "string") {
+            normalized.stopSequences = normalized.stopSequences
+                .split(",")
+                .map((s) => s.trim())
+                .filter(Boolean);
+        }
+        if (typeof normalized.enabledToolNames === "string") {
+            normalized.enabledToolNames = normalized.enabledToolNames
+                .split(",")
+                .map((s) => s.trim())
+                .filter(Boolean);
+        }
+        return normalized;
+    }
     // Common options available on all commands
     static commonOptions = {
         // Core generation options
@@ -621,6 +650,13 @@ export class CLICommandFactory {
             model: argv.model,
             temperature: argv.temperature,
             maxTokens: argv.maxTokens,
+            // Sampling controls — surfaced here so all three command paths
+            // (generate / stream / batch) get them consistently typed instead
+            // of relying on an ad-hoc cast at each sdk call site.
+            topP: argv.topP,
+            topK: argv.topK,
+            stopSequences: argv.stopSequences,
+            enabledToolNames: argv.enabledToolNames,
             systemPrompt: argv.system,
             timeout: argv.timeout,
             disableTools: argv.disableTools,
@@ -1930,18 +1966,8 @@ export class CLICommandFactory {
             }
             // Initialize SDK and session
             const sdk = globalSession.getOrCreateNeuroLink();
-            const sessionVariables = globalSession.getSessionVariables();
-            const enhancedOptions = {
-                ...options,
-                ...sessionVariables,
-                // enabledToolNames must be string[] for the SDK — normalize from CSV string
-                ...(typeof sessionVariables.enabledToolNames === "string" && {
-                    enabledToolNames: sessionVariables.enabledToolNames
-                        .split(",")
-                        .map((t) => t.trim())
-                        .filter(Boolean),
-                }),
-            };
+            const sessionVariables = CLICommandFactory.normalizeLoopSessionVariables(globalSession.getSessionVariables());
+            const enhancedOptions = { ...options, ...sessionVariables };
             const sessionId = globalSession.getCurrentSessionId();
             const context = sessionId
                 ? { ...options.context, sessionId }
@@ -1993,6 +2019,9 @@ export class CLICommandFactory {
                 model: enhancedOptions.model,
                 temperature: enhancedOptions.temperature,
                 maxTokens: enhancedOptions.maxTokens,
+                topP: enhancedOptions.topP,
+                topK: enhancedOptions.topK,
+                stopSequences: enhancedOptions.stopSequences,
                 systemPrompt: enhancedOptions.systemPrompt,
                 timeout: enhancedOptions.timeout
                     ? enhancedOptions.timeout * 1000
@@ -2166,7 +2195,7 @@ export class CLICommandFactory {
      */
     static async executeRealStream(argv, options, inputText, contextMetadata) {
         const sdk = globalSession.getOrCreateNeuroLink();
-        const sessionVariables = globalSession.getSessionVariables();
+        const sessionVariables = CLICommandFactory.normalizeLoopSessionVariables(globalSession.getSessionVariables());
         const enhancedOptions = { ...options, ...sessionVariables };
         const sessionId = globalSession.getCurrentSessionId();
         const context = sessionId
@@ -2201,6 +2230,9 @@ export class CLICommandFactory {
             model: enhancedOptions.model,
             temperature: enhancedOptions.temperature,
             maxTokens: enhancedOptions.maxTokens,
+            topP: enhancedOptions.topP,
+            topK: enhancedOptions.topK,
+            stopSequences: enhancedOptions.stopSequences,
             systemPrompt: enhancedOptions.systemPrompt,
             timeout: enhancedOptions.timeout
                 ? enhancedOptions.timeout * 1000
@@ -2639,18 +2671,8 @@ export class CLICommandFactory {
             }
             const results = [];
             const sdk = globalSession.getOrCreateNeuroLink();
-            const sessionVariables = globalSession.getSessionVariables();
-            const enhancedOptions = {
-                ...options,
-                ...sessionVariables,
-                // enabledToolNames must be string[] for the SDK — normalize from CSV string
-                ...(typeof sessionVariables.enabledToolNames === "string" && {
-                    enabledToolNames: sessionVariables.enabledToolNames
-                        .split(",")
-                        .map((t) => t.trim())
-                        .filter(Boolean),
-                }),
-            };
+            const sessionVariables = CLICommandFactory.normalizeLoopSessionVariables(globalSession.getSessionVariables());
+            const enhancedOptions = { ...options, ...sessionVariables };
             const sessionId = globalSession.getCurrentSessionId();
             for (let i = 0; i < prompts.length; i++) {
                 if (spinner) {
@@ -2692,6 +2714,9 @@ export class CLICommandFactory {
                         model: enhancedOptions.model,
                         temperature: enhancedOptions.temperature,
                         maxTokens: enhancedOptions.maxTokens,
+                        topP: enhancedOptions.topP,
+                        topK: enhancedOptions.topK,
+                        stopSequences: enhancedOptions.stopSequences,
                         systemPrompt: enhancedOptions.systemPrompt,
                         timeout: enhancedOptions.timeout
                             ? enhancedOptions.timeout * 1000

package/dist/cli/loop/optionsSchema.d.ts

@@ -4,4 +4,4 @@ import type { OptionSchema, TextGenerationOptions } from "../../lib/types/index.
  * This object provides metadata for validation and help text in the CLI loop.
  * It is derived from the main TextGenerationOptions interface to ensure consistency.
  */
-export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "tts" | "stt" | "thinkingConfig" | "requestId" | "fileRegistry" | "abortSignal" | "toolFilter" | "excludeTools" | "toolChoice" | "prepareStep" | "credentials">, OptionSchema>;
+export declare const textGenerationOptionsSchema: Record<keyof Omit<TextGenerationOptions, "prompt" | "input" | "schema" | "tools" | "context" | "conversationHistory" | "conversationMessages" | "conversationMemoryConfig" | "originalPrompt" | "middleware" | "expectedOutcome" | "evaluationCriteria" | "region" | "csvOptions" | "tts" | "stt" | "thinkingConfig" | "requestId" | "fileRegistry" | "abortSignal" | "toolFilter" | "excludeTools" | "toolChoice" | "prepareStep" | "credentials" | "onFinish" | "onError">, OptionSchema>;

package/dist/cli/loop/optionsSchema.js

@@ -24,6 +24,18 @@ export const textGenerationOptionsSchema = {
         type: "number",
         description: "The maximum number of tokens to generate.",
     },
+    topP: {
+        type: "number",
+        description: "Top-p (nucleus) sampling parameter. Controls diversity of generated tokens (0.0-1.0).",
+    },
+    topK: {
+        type: "number",
+        description: "Top-k sampling parameter. Limits the number of tokens considered (Google/Gemini models only).",
+    },
+    stopSequences: {
+        type: "string",
+        description: "Stop sequences that will halt generation when encountered (comma-separated).",
+    },
     output: {
         type: "string",
         description: "AI response format - specify just the format value (e.g., 'json', 'structured'). Note: This is automatically transformed to { format: value } for the API.",

package/dist/core/baseProvider.d.ts

@@ -352,7 +352,7 @@ export declare abstract class BaseProvider implements AIProvider {
      * // result.video contains the generated video
      * ```
      */
-    private handleVideoGeneration;
+    protected handleVideoGeneration(options: TextGenerationOptions, startTime: number): Promise<EnhancedGenerateResult>;
     /**
      * Create analytics - delegated to TelemetryHandler
      */

package/dist/core/modules/MessageBuilder.js

@@ -106,6 +106,16 @@ export class MessageBuilder {
                     fileRegistry: options.fileRegistry,
                 };
                 messages = await buildMultimodalMessagesArray(multimodalOptions, this.providerName, this.modelName);
+                // Propagate any systemPrompt augmentation (e.g. inline-file
+                // handling guidance from processUnifiedFilesArray) back to the
+                // caller's options. Providers like GoogleVertex's native
+                // @google/genai stream path read `options.systemPrompt` directly
+                // — without this propagation the augmentation lives only on the
+                // local `multimodalOptions` clone and never reaches the model.
+                if (multimodalOptions.systemPrompt &&
+                    multimodalOptions.systemPrompt !== options.systemPrompt) {
+                    options.systemPrompt = multimodalOptions.systemPrompt;
+                }
             }
             else {
                 if (process.env.NEUROLINK_DEBUG === "true") {
@@ -214,6 +224,16 @@ export class MessageBuilder {
                     fileRegistry: options.fileRegistry,
                 };
                 messages = await buildMultimodalMessagesArray(multimodalOptions, this.providerName, this.modelName);
+                // Propagate any systemPrompt augmentation (e.g. inline-file
+                // handling guidance from processUnifiedFilesArray) back to the
+                // caller's options. Providers like GoogleVertex's native
+                // @google/genai stream path read `options.systemPrompt` directly
+                // — without this propagation the augmentation lives only on the
+                // local `multimodalOptions` clone and never reaches the model.
+                if (multimodalOptions.systemPrompt &&
+                    multimodalOptions.systemPrompt !== options.systemPrompt) {
+                    options.systemPrompt = multimodalOptions.systemPrompt;
+                }
             }
             else {
                 if (process.env.NEUROLINK_DEBUG === "true") {

package/dist/core/redisConversationMemoryManager.js

@@ -958,9 +958,6 @@ User message: "${userMessage}"`;
             title = title.replace(/^(Title:|Here's a title:|The title is:)\s*/i, "");
             title = title.replace(/['"]/g, ""); // Remove quotes
             title = title.replace(/\.$/, ""); // Remove trailing period
-            if (title.length > 60) {
-                title = title.substring(0, 57) + "...";
-            }
             if (title.length < 3) {
                 title = "New Conversation";
             }

package/dist/factories/providerRegistry.js

@@ -135,7 +135,11 @@ export class ProviderRegistry {
                 const openrouterCreds = credentials;
                 const { OpenRouterProvider } = await import("../providers/openRouter.js");
                 return new OpenRouterProvider(modelName, sdk, undefined, openrouterCreds);
-            }, process.env.OPENROUTER_MODEL || "anthropic/claude-3-5-sonnet", ["openrouter", "or"]);
+            }, 
+            // Default updated from claude-3-5-sonnet (sunset by OpenRouter)
+            // to claude-sonnet-4.5. Must match getDefaultOpenRouterModel()
+            // in src/lib/providers/openRouter.ts.
+            process.env.OPENROUTER_MODEL || "anthropic/claude-sonnet-4.5", ["openrouter", "or"]);
             // Register Amazon SageMaker provider
             ProviderFactory.registerProvider(AIProviderName.SAGEMAKER, async (modelName, _providerName, _sdk, region, credentials) => {
                 const sagemakerCreds = credentials;

package/dist/lib/adapters/video/vertexVideoHandler.js

@@ -89,8 +89,15 @@ export function isVertexVideoConfigured() {
  * @throws VideoError if project cannot be determined
  */
 async function getVertexConfig() {
-    const location = process.env.GOOGLE_VERTEX_LOCATION ||
-        process.env.GOOGLE_CLOUD_LOCATION ||
+    // Veo 3.1 is only published in us-central1 (and the synthetic
+    // `global` endpoint). When the operator's environment defaults to
+    // a region where Veo 3.1 isn't available (e.g. us-east5 set for
+    // Gemini chat traffic), the regional endpoint returns a 404
+    // "Publisher Model … was not found". Allow `GOOGLE_VEO_LOCATION` as
+    // a Veo-specific override and fall through to us-central1 instead
+    // of inheriting the default Vertex region.
+    const location = process.env.GOOGLE_VEO_LOCATION ||
+        process.env.GOOGLE_VERTEX_VIDEO_LOCATION ||
         DEFAULT_LOCATION;
     // Try environment variables first
     let project = process.env.GOOGLE_VERTEX_PROJECT ||

package/dist/lib/core/baseProvider.d.ts

@@ -352,7 +352,7 @@ export declare abstract class BaseProvider implements AIProvider {
      * // result.video contains the generated video
      * ```
      */
-    private handleVideoGeneration;
+    protected handleVideoGeneration(options: TextGenerationOptions, startTime: number): Promise<EnhancedGenerateResult>;
     /**
      * Create analytics - delegated to TelemetryHandler
      */

package/dist/lib/core/modules/MessageBuilder.js

@@ -106,6 +106,16 @@ export class MessageBuilder {
                     fileRegistry: options.fileRegistry,
                 };
                 messages = await buildMultimodalMessagesArray(multimodalOptions, this.providerName, this.modelName);
+                // Propagate any systemPrompt augmentation (e.g. inline-file
+                // handling guidance from processUnifiedFilesArray) back to the
+                // caller's options. Providers like GoogleVertex's native
+                // @google/genai stream path read `options.systemPrompt` directly
+                // — without this propagation the augmentation lives only on the
+                // local `multimodalOptions` clone and never reaches the model.
+                if (multimodalOptions.systemPrompt &&
+                    multimodalOptions.systemPrompt !== options.systemPrompt) {
+                    options.systemPrompt = multimodalOptions.systemPrompt;
+                }
             }
             else {
                 if (process.env.NEUROLINK_DEBUG === "true") {
@@ -214,6 +224,16 @@ export class MessageBuilder {
                     fileRegistry: options.fileRegistry,
                 };
                 messages = await buildMultimodalMessagesArray(multimodalOptions, this.providerName, this.modelName);
+                // Propagate any systemPrompt augmentation (e.g. inline-file
+                // handling guidance from processUnifiedFilesArray) back to the
+                // caller's options. Providers like GoogleVertex's native
+                // @google/genai stream path read `options.systemPrompt` directly
+                // — without this propagation the augmentation lives only on the
+                // local `multimodalOptions` clone and never reaches the model.
+                if (multimodalOptions.systemPrompt &&
+                    multimodalOptions.systemPrompt !== options.systemPrompt) {
+                    options.systemPrompt = multimodalOptions.systemPrompt;
+                }
             }
             else {
                 if (process.env.NEUROLINK_DEBUG === "true") {

package/dist/lib/core/redisConversationMemoryManager.js

@@ -958,9 +958,6 @@ User message: "${userMessage}"`;
             title = title.replace(/^(Title:|Here's a title:|The title is:)\s*/i, "");
             title = title.replace(/['"]/g, ""); // Remove quotes
             title = title.replace(/\.$/, ""); // Remove trailing period
-            if (title.length > 60) {
-                title = title.substring(0, 57) + "...";
-            }
             if (title.length < 3) {
                 title = "New Conversation";
             }

package/dist/lib/factories/providerRegistry.js

@@ -135,7 +135,11 @@ export class ProviderRegistry {
                 const openrouterCreds = credentials;
                 const { OpenRouterProvider } = await import("../providers/openRouter.js");
                 return new OpenRouterProvider(modelName, sdk, undefined, openrouterCreds);
-            }, process.env.OPENROUTER_MODEL || "anthropic/claude-3-5-sonnet", ["openrouter", "or"]);
+            }, 
+            // Default updated from claude-3-5-sonnet (sunset by OpenRouter)
+            // to claude-sonnet-4.5. Must match getDefaultOpenRouterModel()
+            // in src/lib/providers/openRouter.ts.
+            process.env.OPENROUTER_MODEL || "anthropic/claude-sonnet-4.5", ["openrouter", "or"]);
             // Register Amazon SageMaker provider
             ProviderFactory.registerProvider(AIProviderName.SAGEMAKER, async (modelName, _providerName, _sdk, region, credentials) => {
                 const sagemakerCreds = credentials;

package/dist/lib/memory/hippocampusInitializer.d.ts

@@ -1,2 +1,2 @@
-import { Hippocampus, type HippocampusConfig } from "@juspay/hippocampus";
-export declare function initializeHippocampus(config: HippocampusConfig): Hippocampus | null;
+import type { HippocampusConfig, HippocampusLike } from "../types/index.js";
+export declare function initializeHippocampus(config: HippocampusConfig): HippocampusLike | null;

package/dist/lib/memory/hippocampusInitializer.js

@@ -1,8 +1,38 @@
-import { Hippocampus } from "@juspay/hippocampus";
+import { createRequire } from "node:module";
 import { logger } from "../utils/logger.js";
+// Lazy require so importing NeuroLink core does not fail when the optional
+// peer @juspay/hippocampus is not installed. The package was previously a
+// hard runtime dependency, but Hippocampus declares a peer on
+// @juspay/neurolink which made pnpm pull a registry NeuroLink that
+// transitively required @ai-sdk/google + @ai-sdk/google-vertex into the
+// production graph. Making memory optional breaks that cycle while keeping
+// the same runtime behavior whenever the package is installed.
+const lazyRequire = createRequire(import.meta.url);
+let cachedModule;
+function loadHippocampusModule() {
+    if (cachedModule !== undefined) {
+        return cachedModule;
+    }
+    try {
+        cachedModule = lazyRequire("@juspay/hippocampus");
+        return cachedModule;
+    }
+    catch (error) {
+        cachedModule = null;
+        logger.debug("[memoryInitializer] @juspay/hippocampus is not installed; memory features disabled.", {
+            error: error instanceof Error ? error.message : String(error),
+        });
+        return null;
+    }
+}
 export function initializeHippocampus(config) {
+    const mod = loadHippocampusModule();
+    if (!mod) {
+        logger.warn("[memoryInitializer] Memory configuration provided but @juspay/hippocampus is not installed. Run `pnpm add @juspay/hippocampus` (or your package manager equivalent) to enable memory.");
+        return null;
+    }
     try {
-        const instance = new Hippocampus(config);
+        const instance = new mod.Hippocampus(config);
         logger.info("[memoryInitializer] Memory initialized successfully", {
             storageType: config.storage?.type || "sqlite",
             maxWords: config.maxWords || 50,

package/dist/lib/middleware/builtin/lifecycle.js

@@ -9,6 +9,7 @@
  */
 import { logger } from "../../utils/logger.js";
 import { isRecoverableError } from "../../utils/errorHandling.js";
+import { fireOnErrorOnce } from "../../utils/lifecycleCallbacks.js";
 export function createLifecycleMiddleware(config = {}) {
     const metadata = {
         id: "lifecycle",
@@ -50,22 +51,12 @@ export function createLifecycleMiddleware(config = {}) {
                 return result;
             }
             catch (error) {
-                if (config.onError) {
-                    const err = error instanceof Error ? error : new Error(String(error));
-                    try {
-                        const callbackResult = config.onError({
-                            error: err,
-                            duration: Date.now() - startTime,
-                            recoverable: isRecoverableError(err),
-                        });
-                        Promise.resolve(callbackResult).catch((e) => {
-                            logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                        });
-                    }
-                    catch (e) {
-                        logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                    }
-                }
+                const err = error instanceof Error ? error : new Error(String(error));
+                fireOnErrorOnce(config.onError, error, {
+                    error: err,
+                    duration: Date.now() - startTime,
+                    recoverable: isRecoverableError(err),
+                });
                 throw error;
             }
         },
@@ -102,22 +93,12 @@ export function createLifecycleMiddleware(config = {}) {
                             controller.enqueue(chunk);
                         }
                         catch (error) {
-                            if (config.onError) {
-                                const err = error instanceof Error ? error : new Error(String(error));
-                                try {
-                                    const callbackResult = config.onError({
-                                        error: err,
-                                        duration: Date.now() - startTime,
-                                        recoverable: isRecoverableError(err),
-                                    });
-                                    Promise.resolve(callbackResult).catch((e) => {
-                                        logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                                    });
-                                }
-                                catch (e) {
-                                    logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                                }
-                            }
+                            const err = error instanceof Error ? error : new Error(String(error));
+                            fireOnErrorOnce(config.onError, error, {
+                                error: err,
+                                duration: Date.now() - startTime,
+                                recoverable: isRecoverableError(err),
+                            });
                             throw error;
                         }
                     },
@@ -144,22 +125,12 @@ export function createLifecycleMiddleware(config = {}) {
                 };
             }
             catch (error) {
-                if (config.onError) {
-                    const err = error instanceof Error ? error : new Error(String(error));
-                    try {
-                        const callbackResult = config.onError({
-                            error: err,
-                            duration: Date.now() - startTime,
-                            recoverable: isRecoverableError(err),
-                        });
-                        Promise.resolve(callbackResult).catch((e) => {
-                            logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                        });
-                    }
-                    catch (e) {
-                        logger.warn("[LifecycleMiddleware] onError callback error:", e);
-                    }
-                }
+                const err = error instanceof Error ? error : new Error(String(error));
+                fireOnErrorOnce(config.onError, error, {
+                    error: err,
+                    duration: Date.now() - startTime,
+                    recoverable: isRecoverableError(err),
+                });
                 throw error;
             }
         },

package/dist/lib/neurolink.js

@@ -64,6 +64,7 @@ import { getConversationMessages, storeConversationTurn, } from "./utils/convers
 import { CircuitBreaker, ERROR_CODES, ErrorFactory, isAbortError, isRetriableError, logStructuredError, NeuroLinkError, withRetry, withTimeout, } from "./utils/errorHandling.js";
 // Factory processing imports
 import { createCleanStreamOptions, enhanceTextGenerationOptions, processFactoryOptions, processStreamingFactoryOptions, validateFactoryConfig, } from "./utils/factoryProcessing.js";
+import { fireOnErrorOnce } from "./utils/lifecycleCallbacks.js";
 import { logger, mcpLogger } from "./utils/logger.js";
 import { extractMcpErrorText } from "./utils/mcpErrorText.js";
 import { createCustomToolServerInfo, detectCategory, } from "./utils/mcpDefaults.js";
@@ -2753,7 +2754,29 @@ Current user's request: ${currentInput}`;
      * @since 1.0.0
      */
     async generate(optionsOrPrompt) {
-        return this.runWithFallbackOrchestration(optionsOrPrompt, "generate", (opts) => tracers.sdk.startActiveSpan("neurolink.generate", { kind: SpanKind.INTERNAL }, (generateSpan) => this.executeGenerateWithMetricsContext(opts, generateSpan)));
+        const startTime = Date.now();
+        try {
+            return await this.runWithFallbackOrchestration(optionsOrPrompt, "generate", (opts) => tracers.sdk.startActiveSpan("neurolink.generate", { kind: SpanKind.INTERNAL }, (generateSpan) => this.executeGenerateWithMetricsContext(opts, generateSpan)));
+        }
+        catch (error) {
+            // Fire `onError` lifecycle callback for ANY thrown error — including
+            // ones raised before the provider is even instantiated (invalid
+            // provider name, missing credentials, etc.). The downstream
+            // LifecycleMiddleware only fires `onError` once it has wrapped the
+            // AI SDK doGenerate, which is too late for early-resolution failures.
+            // `fireOnErrorOnce` dedupes against the middleware path so the
+            // consumer callback fires at most once per logical failure.
+            const onError = typeof optionsOrPrompt === "object" && optionsOrPrompt !== null
+                ? optionsOrPrompt.onError
+                : undefined;
+            const err = error instanceof Error ? error : new Error(String(error));
+            fireOnErrorOnce(onError, error, {
+                error: err,
+                duration: Date.now() - startTime,
+                recoverable: false,
+            });
+            throw error;
+        }
     }
     /**
      * Curator P2-3: wraps a generate/stream call with the fallback
@@ -3167,6 +3190,12 @@ Current user's request: ${currentInput}`;
             middleware: options.middleware,
             conversationMessages: options.conversationMessages,
             credentials: options.credentials,
+            // Lifecycle callbacks must reach the provider so non-AI-SDK paths
+            // (Vertex's native @google/genai, native Bedrock, Ollama, etc.) can
+            // invoke them directly. Pipeline A also still receives them via the
+            // wrapped middleware config set by applyGenerateLifecycleMiddleware.
+            onFinish: options.onFinish,
+            onError: options.onError,
         };
         const extraContext = options;
         if (extraContext.sessionId || extraContext.userId) {
@@ -5032,7 +5061,25 @@ Current user's request: ${currentInput}`;
                 : [],
             optionKeys: Object.keys(options),
         });
-        return this.streamWithIterationFallback(options);
+        const startTime = Date.now();
+        try {
+            return await this.streamWithIterationFallback(options);
+        }
+        catch (error) {
+            // Fire `onError` for early-resolution failures (invalid provider,
+            // missing credentials, etc.) that surface before the per-chunk
+            // wrapper installed by GoogleVertex / LifecycleMiddleware can run.
+            // `fireOnErrorOnce` dedupes against the middleware path so the
+            // consumer callback fires at most once per logical failure.
+            const onError = options.onError;
+            const err = error instanceof Error ? error : new Error(String(error));
+            fireOnErrorOnce(onError, error, {
+                error: err,
+                duration: Date.now() - startTime,
+                recoverable: false,
+            });
+            throw error;
+        }
     }
     /**
      * Curator P2-3 / Reviewer Finding #2: stream-fallback that also covers

package/dist/lib/providers/googleAiStudio.d.ts

@@ -41,7 +41,8 @@ export declare class GoogleAIStudioProvider extends BaseProvider {
     getProviderName(): AIProviderName;
     getDefaultModel(): string;
     /**
-     * 🔧 PHASE 2: Return AI SDK model instance for tool calling
+     * AI SDK model instance — no longer used.
+     * All models are routed through native @google/genai SDK directly.
      */
     getAISDKModel(): LanguageModel;
     protected formatProviderError(error: unknown): Error;
@@ -62,8 +63,8 @@ export declare class GoogleAIStudioProvider extends BaseProvider {
     private estimateTokenCount;
     protected executeStream(options: StreamOptions, analysisSchema?: ZodUnknownSchema | Schema<unknown>): Promise<StreamResult>;
     /**
-     * Execute stream using native @google/genai SDK for Gemini 3 models
-     * This bypasses @ai-sdk/google to properly handle thought_signature
+     * Execute stream using native @google/genai SDK
+     * Uses @google/genai directly for all Gemini models (2.0, 2.5, 3.x)
      */
     private executeNativeGemini3Stream;
     /**
@@ -75,6 +76,13 @@ export declare class GoogleAIStudioProvider extends BaseProvider {
      * Override generate to route Gemini 3 models with tools to native SDK
      */
     generate(optionsOrPrompt: TextGenerationOptions | string): Promise<EnhancedGenerateResult | null>;
+    /**
+     * Emit `generation:end` so the Pipeline B observability listener creates
+     * a `model.generation` span for native Google AI Studio generate calls.
+     * Without this hand-off the native path silently disappears from
+     * Pipeline B exporters (Langfuse, custom OTEL collectors).
+     */
+    private emitPipelineBGenerationEvent;
     private executeAudioStreamViaGeminiLive;
     protected getDefaultEmbeddingModel(): string;
     /**