@juspay/neurolink 9.41.0 → 9.42.1

package/dist/providers/googleVertex.js

@@ -22,6 +22,7 @@ import { createGoogleAuthConfig, createVertexProjectConfig, validateApiKey, } fr
 import { convertZodToJsonSchema, inlineJsonSchema, } from "../utils/schemaConversion.js";
 import { composeAbortSignals, createTimeoutController, TimeoutError, } from "../utils/timeout.js";
 import { estimateTokens } from "../utils/tokenEstimation.js";
+import { resolveToolChoice } from "../utils/toolChoice.js";
 import { buildNativeConfig, buildNativeToolDeclarations, collectStreamChunks, collectStreamChunksIncremental, computeMaxSteps as computeMaxStepsShared, createTextChannel, executeNativeToolCalls, extractTextFromParts, handleMaxStepsTermination, pushModelResponseToHistory, sanitizeToolsForGemini, } from "./googleNativeGemini3.js";
 import { getModelId } from "./providerTypeUtils.js";
 // Import proper types for multimodal message handling
@@ -776,251 +777,76 @@ export class GoogleVertexProvider extends BaseProvider {
         this.validateStreamOptions(options);
     }
     async executeStream(options, analysisSchema) {
-        // Check if this is a Gemini 3 model with tools - use native SDK for thought_signature
-        const gemini3CheckModelName = this.resolveAlias(options.model || this.modelName || getDefaultVertexModel());
-        // Structured output (analysisSchema, JSON format, or schema) is incompatible with tools on Gemini.
-        // Compute once and reuse in both the native Gemini 3 gate and the streamText fallback path.
+        const modelName = this.resolveAlias(options.model || this.modelName || getDefaultVertexModel());
+        const nativeGemini3Result = await this.maybeExecuteNativeGemini3ToolStream(options, analysisSchema, modelName);
+        if (nativeGemini3Result) {
+            return nativeGemini3Result;
+        }
+        return this.executeAISDKStream(options, analysisSchema, modelName);
+    }
+    async maybeExecuteNativeGemini3ToolStream(options, analysisSchema, modelName) {
         const wantsStructuredOutput = analysisSchema || options.output?.format === "json" || options.schema;
-        // Check for tools from options AND from SDK (MCP tools)
-        // Need to check early if we should route to native SDK
-        const gemini3CheckShouldUseTools = !options.disableTools && this.supportsTools() && !wantsStructuredOutput;
+        const shouldUseTools = !options.disableTools && this.supportsTools() && !wantsStructuredOutput;
         const optionTools = options.tools || {};
-        const sdkTools = gemini3CheckShouldUseTools ? await this.getAllTools() : {};
+        const sdkTools = shouldUseTools ? await this.getAllTools() : {};
         const combinedToolCount = Object.keys(optionTools).length + Object.keys(sdkTools).length;
-        const hasTools = gemini3CheckShouldUseTools && combinedToolCount > 0;
-        if (isGemini3Model(gemini3CheckModelName) && hasTools) {
-            // Process CSV files before routing to native SDK (bypasses normal message builder)
-            const processedOptions = await this.processCSVFilesForNativeSDK(options);
-            // Merge SDK tools into options for native SDK path
-            const mergedOptions = {
-                ...processedOptions,
-                tools: { ...sdkTools, ...optionTools },
-            };
-            logger.info("[GoogleVertex] Routing Gemini 3 to native SDK for tool calling", {
-                model: gemini3CheckModelName,
-                optionToolCount: Object.keys(optionTools).length,
-                sdkToolCount: Object.keys(sdkTools).length,
-                totalToolCount: combinedToolCount,
-            });
-            return this.executeNativeGemini3Stream(mergedOptions);
+        const hasTools = shouldUseTools && combinedToolCount > 0;
+        if (!isGemini3Model(modelName) || !hasTools) {
+            return null;
         }
-        // Initialize stream execution tracking
+        const processedOptions = await this.processCSVFilesForNativeSDK(options);
+        const mergedOptions = {
+            ...processedOptions,
+            tools: { ...sdkTools, ...optionTools },
+        };
+        logger.info("[GoogleVertex] Routing Gemini 3 to native SDK for tool calling", {
+            model: modelName,
+            optionToolCount: Object.keys(optionTools).length,
+            sdkToolCount: Object.keys(sdkTools).length,
+            totalToolCount: combinedToolCount,
+        });
+        return this.executeNativeGemini3Stream(mergedOptions);
+    }
+    async executeAISDKStream(options, analysisSchema, modelName) {
         const functionTag = "GoogleVertexProvider.executeStream";
-        let chunkCount = 0;
-        // Setup timeout controller
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
+        const tracking = {
+            chunkCount: 0,
+            collectedToolCalls: [],
+            collectedToolResults: [],
+        };
+        const timeoutController = createTimeoutController(this.getTimeout(options), this.providerName, "stream");
         try {
-            // Validate stream options
             this.validateStreamOptionsOnly(options);
-            // Build message array from options with multimodal support
-            // Using protected helper from BaseProvider to eliminate code duplication
             const messages = await this.buildMessagesForStream(options);
-            const model = await this.getAISDKModelWithMiddleware(options); // This is where network connection happens!
-            // Get all available tools (direct + MCP + external + user-provided RAG tools) for streaming
-            const shouldUseTools = !options.disableTools && this.supportsTools();
-            const baseStreamTools = shouldUseTools ? await this.getAllTools() : {};
-            const rawTools = shouldUseTools
-                ? { ...baseStreamTools, ...(options.tools || {}) }
-                : {};
-            // Only sanitize for Gemini models (not Anthropic/Claude models routed through Vertex)
-            const isAnthropic = isAnthropicModel(gemini3CheckModelName);
-            let tools;
-            if (Object.keys(rawTools).length > 0 && !isAnthropic) {
-                const sanitized = sanitizeToolsForGemini(rawTools);
-                if (sanitized.dropped.length > 0) {
-                    logger.warn(`[GoogleVertex] Dropped ${sanitized.dropped.length} incompatible tool(s): ${sanitized.dropped.join(", ")}`);
-                }
-                tools =
-                    Object.keys(sanitized.tools).length > 0 ? sanitized.tools : undefined;
-            }
-            else if (isAnthropic && Object.keys(rawTools).length > 0) {
-                // Anthropic models don't need Gemini sanitization — pass tools through
-                tools = rawTools;
-            }
-            else {
-                tools = undefined;
-            }
-            logger.debug(`${functionTag}: Tools for streaming`, {
+            const model = await this.getAISDKModelWithMiddleware(options);
+            const { shouldUseTools, tools, isAnthropic } = await this.resolveAISDKStreamTools(options, modelName, functionTag);
+            const streamOptions = this.buildAISDKStreamOptions({
+                options,
+                analysisSchema,
+                functionTag,
+                modelName,
+                model,
+                messages,
+                tools,
                 shouldUseTools,
-                baseToolCount: Object.keys(baseStreamTools).length,
-                externalToolCount: Object.keys(options.tools || {}).length,
-                toolCount: Object.keys(tools ?? {}).length,
-                toolNames: Object.keys(tools ?? {}),
-            });
-            // Model-specific maxTokens handling
-            const modelName = this.resolveAlias(options.model || this.modelName || getDefaultVertexModel());
-            // Use cached model configuration to determine maxTokens handling for streaming performance
-            // This avoids hardcoded model-specific logic and repeated config lookups
-            const shouldSetMaxTokens = this.shouldSetMaxTokensCached(modelName);
-            const maxTokens = shouldSetMaxTokens
-                ? options.maxTokens // No default limit
-                : undefined;
-            // Build complete stream options with proper typing
-            let streamOptions = {
-                model: model,
-                messages: messages,
-                temperature: options.temperature,
-                ...(maxTokens && { maxTokens }),
-                maxRetries: 0, // NL11: Disable AI SDK's invisible internal retries; we handle retries with OTel instrumentation
-                ...(shouldUseTools &&
-                    tools &&
-                    Object.keys(tools).length > 0 && {
-                    tools,
-                    toolChoice: "auto",
-                    stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
-                }),
-                abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
-                // Gemini 3: use thinkingLevel via providerOptions (Vertex AI)
-                // Gemini 2.5: use thinkingBudget via providerOptions
-                ...(options.thinkingConfig?.enabled && {
-                    providerOptions: {
-                        vertex: {
-                            thinkingConfig: {
-                                ...(options.thinkingConfig.thinkingLevel && {
-                                    thinkingLevel: options.thinkingConfig.thinkingLevel,
-                                }),
-                                ...(options.thinkingConfig.budgetTokens &&
-                                    !options.thinkingConfig.thinkingLevel && {
-                                    thinkingBudget: options.thinkingConfig.budgetTokens,
-                                }),
-                                includeThoughts: true,
-                            },
-                        },
-                    },
-                }),
-                onError: (event) => {
-                    const error = event.error;
-                    const errorMessage = error instanceof Error ? error.message : String(error);
-                    logger.error(`${functionTag}: Stream error`, {
-                        provider: this.providerName,
-                        modelName: this.modelName,
-                        error: errorMessage,
-                        chunkCount,
-                    });
-                },
-                onFinish: (event) => {
-                    logger.debug(`${functionTag}: Stream finished`, {
-                        finishReason: event.finishReason,
-                        totalChunks: chunkCount,
-                    });
-                },
-                onChunk: () => {
-                    chunkCount++;
-                },
-                onStepFinish: ({ toolCalls, toolResults }) => {
-                    logger.info("Tool execution completed", { toolResults, toolCalls });
-                    // Handle tool execution storage
-                    this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
-                        logger.warn("[GoogleVertexProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: error instanceof Error ? error.message : String(error),
-                        });
-                    });
-                },
-            };
-            if (analysisSchema) {
-                try {
-                    // Gemini cannot use tools and JSON schema simultaneously
-                    if (!isAnthropic) {
-                        delete streamOptions.tools;
-                        delete streamOptions.toolChoice;
-                        delete streamOptions.stopWhen;
-                    }
-                    streamOptions = {
-                        ...streamOptions,
-                        experimental_output: Output.object({
-                            schema: analysisSchema,
-                        }),
-                    };
-                }
-                catch (error) {
-                    logger.warn("Schema application failed, continuing without schema", {
-                        error: String(error),
-                    });
-                }
-            }
-            // Wrap streamText in an OTel span to capture provider-level latency and token usage
-            const streamSpan = streamTracer.startSpan("neurolink.provider.streamText", {
-                kind: SpanKind.CLIENT,
-                attributes: {
-                    "gen_ai.system": "vertex",
-                    "gen_ai.request.model": getModelId(model, this.modelName || "unknown"),
-                },
+                isAnthropic,
+                timeoutController,
+                tracking,
             });
-            let result;
-            try {
-                result = streamText(streamOptions);
-            }
-            catch (err) {
-                streamSpan.recordException(err instanceof Error ? err : new Error(String(err)));
-                streamSpan.setStatus({
-                    code: SpanStatusCode.ERROR,
-                    message: err instanceof Error ? err.message : String(err),
-                });
-                streamSpan.end();
-                throw err;
-            }
-            // Collect token usage and finish reason asynchronously when the stream completes,
-            // then end the span. This avoids blocking the stream consumer.
-            Promise.resolve(result.usage)
-                .then((usage) => {
-                streamSpan.setAttribute("gen_ai.usage.input_tokens", usage.inputTokens || 0);
-                streamSpan.setAttribute("gen_ai.usage.output_tokens", usage.outputTokens || 0);
-                const effectiveModel = options.model ||
-                    getModelId(model, this.modelName || getDefaultVertexModel());
-                const cost = calculateCost(this.providerName, effectiveModel, {
-                    input: usage.inputTokens || 0,
-                    output: usage.outputTokens || 0,
-                    total: (usage.inputTokens || 0) + (usage.outputTokens || 0),
-                });
-                if (cost && cost > 0) {
-                    streamSpan.setAttribute("neurolink.cost", cost);
-                }
-            })
-                .catch(() => {
-                // Usage may not be available if the stream is aborted
-            });
-            Promise.resolve(result.finishReason)
-                .then((reason) => {
-                streamSpan.setAttribute("gen_ai.response.finish_reason", reason || "unknown");
-            })
-                .catch(() => {
-                // Finish reason may not be available if the stream is aborted
-            });
-            Promise.resolve(result.text)
-                .then(() => {
-                streamSpan.end();
-            })
-                .catch((err) => {
-                streamSpan.setStatus({
-                    code: SpanStatusCode.ERROR,
-                    message: err instanceof Error ? err.message : String(err),
-                });
-                streamSpan.end();
+            const result = this.startObservedAISDKStream(streamOptions, model, modelName, options);
+            this.observeAISDKStreamResult(result, {
+                model,
+                modelName,
+                options,
+                timeoutController,
             });
-            // Defer timeout cleanup until the stream completes or errors.
-            // Guard against NoOutputGeneratedError becoming an unhandled rejection.
-            Promise.resolve(result.text)
-                .catch((err) => {
-                logger.debug("Stream text promise rejected (expected for empty streams)", {
-                    error: err instanceof Error ? err.message : String(err),
-                });
-            })
-                .finally(() => timeoutController?.cleanup());
-            // Transform string stream to content object stream using BaseProvider method
-            const transformedStream = this.createTextStream(result);
-            // Track tool calls and results for streaming
-            const toolCalls = [];
-            const toolResults = [];
             return {
-                stream: transformedStream,
+                stream: this.createTextStream(result),
                 provider: this.providerName,
                 model: this.modelName,
                 ...(shouldUseTools && {
-                    toolCalls,
-                    toolResults,
+                    toolCalls: tracking.collectedToolCalls,
+                    toolResults: tracking.collectedToolResults,
                 }),
             };
         }
@@ -1030,11 +856,222 @@ export class GoogleVertexProvider extends BaseProvider {
                 provider: this.providerName,
                 modelName: this.modelName,
                 error: String(error),
-                chunkCount,
+                chunkCount: tracking.chunkCount,
             });
             throw this.handleProviderError(error);
         }
     }
+    async resolveAISDKStreamTools(options, modelName, functionTag) {
+        const shouldUseTools = !options.disableTools && this.supportsTools();
+        const baseStreamTools = shouldUseTools ? await this.getAllTools() : {};
+        const rawTools = shouldUseTools
+            ? { ...baseStreamTools, ...(options.tools || {}) }
+            : {};
+        const isAnthropic = isAnthropicModel(modelName);
+        let tools;
+        if (Object.keys(rawTools).length > 0 && !isAnthropic) {
+            const sanitized = sanitizeToolsForGemini(rawTools);
+            if (sanitized.dropped.length > 0) {
+                logger.warn(`[GoogleVertex] Dropped ${sanitized.dropped.length} incompatible tool(s): ${sanitized.dropped.join(", ")}`);
+            }
+            tools =
+                Object.keys(sanitized.tools).length > 0 ? sanitized.tools : undefined;
+        }
+        else if (isAnthropic && Object.keys(rawTools).length > 0) {
+            tools = rawTools;
+        }
+        else {
+            tools = undefined;
+        }
+        logger.debug(`${functionTag}: Tools for streaming`, {
+            shouldUseTools,
+            baseToolCount: Object.keys(baseStreamTools).length,
+            externalToolCount: Object.keys(options.tools || {}).length,
+            toolCount: Object.keys(tools ?? {}).length,
+            toolNames: Object.keys(tools ?? {}),
+        });
+        return {
+            shouldUseTools,
+            tools,
+            isAnthropic,
+            baseToolCount: Object.keys(baseStreamTools).length,
+        };
+    }
+    buildAISDKStreamOptions(params) {
+        const { options, analysisSchema, functionTag, modelName, model, messages, tools, shouldUseTools, isAnthropic, timeoutController, tracking, } = params;
+        const shouldSetMaxTokens = this.shouldSetMaxTokensCached(modelName);
+        const maxTokens = shouldSetMaxTokens ? options.maxTokens : undefined;
+        let streamOptions = {
+            model,
+            messages,
+            temperature: options.temperature,
+            ...(maxTokens && { maxTokens }),
+            maxRetries: 0,
+            ...(shouldUseTools &&
+                tools &&
+                Object.keys(tools).length > 0 && {
+                tools,
+                toolChoice: resolveToolChoice(options, tools, shouldUseTools),
+                stopWhen: stepCountIs(options.maxSteps || DEFAULT_MAX_STEPS),
+            }),
+            abortSignal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
+            experimental_telemetry: this.telemetryHandler.getTelemetryConfig(options),
+            ...(options.thinkingConfig?.enabled && {
+                providerOptions: {
+                    vertex: {
+                        thinkingConfig: {
+                            ...(options.thinkingConfig.thinkingLevel && {
+                                thinkingLevel: options.thinkingConfig.thinkingLevel,
+                            }),
+                            ...(options.thinkingConfig.budgetTokens &&
+                                !options.thinkingConfig.thinkingLevel && {
+                                thinkingBudget: options.thinkingConfig.budgetTokens,
+                            }),
+                            includeThoughts: true,
+                        },
+                    },
+                },
+            }),
+            onError: (event) => {
+                const errorMessage = event.error instanceof Error
+                    ? event.error.message
+                    : String(event.error);
+                logger.error(`${functionTag}: Stream error`, {
+                    provider: this.providerName,
+                    modelName: this.modelName,
+                    error: errorMessage,
+                    chunkCount: tracking.chunkCount,
+                });
+            },
+            onFinish: (event) => {
+                logger.debug(`${functionTag}: Stream finished`, {
+                    finishReason: event.finishReason,
+                    totalChunks: tracking.chunkCount,
+                });
+            },
+            onChunk: () => {
+                tracking.chunkCount++;
+            },
+            onStepFinish: ({ toolCalls, toolResults }) => {
+                this.captureAISDKStreamToolStep(options, toolCalls, toolResults, tracking);
+            },
+        };
+        if (!analysisSchema) {
+            return streamOptions;
+        }
+        try {
+            if (!isAnthropic) {
+                delete streamOptions.tools;
+                delete streamOptions.toolChoice;
+                delete streamOptions.stopWhen;
+            }
+            streamOptions = {
+                ...streamOptions,
+                experimental_output: Output.object({ schema: analysisSchema }),
+            };
+        }
+        catch (error) {
+            logger.warn("Schema application failed, continuing without schema", {
+                error: String(error),
+            });
+        }
+        return streamOptions;
+    }
+    captureAISDKStreamToolStep(options, toolCalls, toolResults, tracking) {
+        logger.info("Tool execution completed", { toolResults, toolCalls });
+        for (const toolCall of toolCalls) {
+            tracking.collectedToolCalls.push({
+                toolCallId: toolCall.toolCallId,
+                toolName: toolCall.toolName,
+                args: toolCall.args ?? toolCall.input ?? toolCall.parameters ?? {},
+            });
+        }
+        for (const toolResult of toolResults) {
+            tracking.collectedToolResults.push({
+                toolName: toolResult.toolName,
+                status: toolResult.error ? "failure" : "success",
+                output: (toolResult.output ?? toolResult.result) ??
+                    undefined,
+                error: toolResult.error,
+                id: toolResult.toolCallId ?? toolResult.toolName,
+            });
+        }
+        this.handleToolExecutionStorage(toolCalls, toolResults, options, new Date()).catch((error) => {
+            logger.warn("[GoogleVertexProvider] Failed to store tool executions", {
+                provider: this.providerName,
+                error: error instanceof Error ? error.message : String(error),
+            });
+        });
+    }
+    startObservedAISDKStream(streamOptions, model, modelName, options) {
+        const streamSpan = streamTracer.startSpan("neurolink.provider.streamText", {
+            kind: SpanKind.CLIENT,
+            attributes: {
+                "gen_ai.system": "vertex",
+                "gen_ai.request.model": getModelId(model, this.modelName || "unknown"),
+            },
+        });
+        try {
+            const result = streamText(streamOptions);
+            this.attachAISDKStreamObservers(result, streamSpan, model, modelName, options);
+            return result;
+        }
+        catch (error) {
+            streamSpan.recordException(error instanceof Error ? error : new Error(String(error)));
+            streamSpan.setStatus({
+                code: SpanStatusCode.ERROR,
+                message: error instanceof Error ? error.message : String(error),
+            });
+            streamSpan.end();
+            throw error;
+        }
+    }
+    attachAISDKStreamObservers(result, streamSpan, model, modelName, options) {
+        Promise.resolve(result.usage)
+            .then((usage) => {
+            streamSpan.setAttribute("gen_ai.usage.input_tokens", usage.inputTokens || 0);
+            streamSpan.setAttribute("gen_ai.usage.output_tokens", usage.outputTokens || 0);
+            const effectiveModel = options.model ||
+                getModelId(model, modelName || getDefaultVertexModel());
+            const cost = calculateCost(this.providerName, effectiveModel, {
+                input: usage.inputTokens || 0,
+                output: usage.outputTokens || 0,
+                total: (usage.inputTokens || 0) + (usage.outputTokens || 0),
+            });
+            if (cost && cost > 0) {
+                streamSpan.setAttribute("neurolink.cost", cost);
+            }
+        })
+            .catch(() => undefined);
+        Promise.resolve(result.finishReason)
+            .then((reason) => {
+            streamSpan.setAttribute("gen_ai.response.finish_reason", reason || "unknown");
+        })
+            .catch(() => undefined);
+        Promise.resolve(result.text)
+            .then(() => {
+            streamSpan.end();
+        })
+            .catch((error) => {
+            streamSpan.setStatus({
+                code: SpanStatusCode.ERROR,
+                message: error instanceof Error ? error.message : String(error),
+            });
+            streamSpan.end();
+        });
+    }
+    observeAISDKStreamResult(result, params) {
+        void params.model;
+        void params.modelName;
+        void params.options;
+        Promise.resolve(result.text)
+            .catch((error) => {
+            logger.debug("Stream text promise rejected (expected for empty streams)", {
+                error: error instanceof Error ? error.message : String(error),
+            });
+        })
+            .finally(() => params.timeoutController?.cleanup());
+    }
     /**
      * Create @google/genai client configured for Vertex AI
      */
@@ -1189,210 +1226,187 @@ export class GoogleVertexProvider extends BaseProvider {
                 [ATTR.GEN_AI_OPERATION]: "stream",
                 [ATTR.NL_PROVIDER]: this.providerName,
             },
-        }, async (span) => {
-            const client = await this.createVertexGenAIClient(options.region);
-            const effectiveLocation = options.region || this.location || getVertexLocation();
-            logger.debug("[GoogleVertex] Using native @google/genai for Gemini 3", {
-                model: modelName,
-                hasTools: !!options.tools && Object.keys(options.tools).length > 0,
-                project: this.projectId,
-                location: effectiveLocation,
+        }, (span) => this.executeNativeGemini3StreamWithSpan(options, modelName, span));
+    }
+    async executeNativeGemini3StreamWithSpan(options, modelName, span) {
+        const client = await this.createVertexGenAIClient(options.region);
+        const effectiveLocation = options.region || this.location || getVertexLocation();
+        logger.debug("[GoogleVertex] Using native @google/genai for Gemini 3", {
+            model: modelName,
+            hasTools: !!options.tools && Object.keys(options.tools).length > 0,
+            project: this.projectId,
+            location: effectiveLocation,
+        });
+        const multimodalInput = options.input;
+        const contents = this.buildNativeContentParts(options.input.text, multimodalInput, "native stream");
+        let hasToolsInput = !!options.tools &&
+            Object.keys(options.tools).length > 0 &&
+            !options.disableTools;
+        const streamOptions = options;
+        const wantsJsonOutput = streamOptions.output?.format === "json" || streamOptions.schema;
+        if (wantsJsonOutput && hasToolsInput) {
+            logger.warn("[GoogleVertex] Gemini does not support tools and JSON schema output simultaneously. Disabling tools for this request.");
+            hasToolsInput = false;
+        }
+        let toolsConfig;
+        let executeMap = new Map();
+        if (hasToolsInput) {
+            const toolDeclarationResult = buildNativeToolDeclarations(options.tools);
+            toolsConfig = toolDeclarationResult.toolsConfig;
+            executeMap = toolDeclarationResult.executeMap;
+            logger.debug("[GoogleVertex] Converted tools for native SDK", {
+                toolCount: toolsConfig[0].functionDeclarations.length,
+                toolNames: toolsConfig[0].functionDeclarations.map((tool) => tool.name),
             });
-            // Build contents from input with multimodal support
-            const multimodalInput = options.input;
-            const contents = this.buildNativeContentParts(options.input.text, multimodalInput, "native stream");
-            // Convert tools to native format
-            let hasToolsInput = options.tools &&
-                Object.keys(options.tools).length > 0 &&
-                !options.disableTools;
-            // Guard: Gemini cannot use tools + JSON schema simultaneously
-            const streamOptions = options;
-            const wantsJsonOutput = streamOptions.output?.format === "json" || streamOptions.schema;
-            if (wantsJsonOutput && hasToolsInput) {
-                logger.warn("[GoogleVertex] Gemini does not support tools and JSON schema output simultaneously. Disabling tools for this request.");
-                hasToolsInput = false;
-            }
-            let toolsConfig;
-            let executeMap = new Map();
-            if (hasToolsInput) {
-                const result = buildNativeToolDeclarations(options.tools);
-                toolsConfig = result.toolsConfig;
-                executeMap = result.executeMap;
-                logger.debug("[GoogleVertex] Converted tools for native SDK", {
-                    toolCount: toolsConfig[0].functionDeclarations.length,
-                    toolNames: toolsConfig[0].functionDeclarations.map((t) => t.name),
+        }
+        const config = buildNativeConfig(options, toolsConfig);
+        if (wantsJsonOutput) {
+            config.responseMimeType = "application/json";
+            if (streamOptions.schema) {
+                const rawSchema = convertZodToJsonSchema(streamOptions.schema);
+                const inlinedSchema = inlineJsonSchema(rawSchema);
+                if (inlinedSchema.$schema) {
+                    delete inlinedSchema.$schema;
+                }
+                config.responseSchema = inlinedSchema;
+                logger.debug("[GoogleVertex] Added responseSchema for JSON output (stream)", {
+                    schemaKeys: Object.keys(inlinedSchema),
                 });
             }
-            // Build config — systemInstruction stays in config for Gemini 3.x.
-            // The @google/genai SDK maps config.systemInstruction to the HTTP-level
-            // system_instruction field, which is the correct mechanism for all
-            // Gemini 3.x models (including global endpoint).  Older workaround
-            // that moved systemInstruction into user/model content messages caused
-            // "Please use a valid role: user, model" on Gemini 3.1+ preview models.
-            const config = buildNativeConfig(options, toolsConfig);
-            // Add JSON output format support for native SDK stream
-            if (streamOptions.output?.format === "json" || streamOptions.schema) {
-                config.responseMimeType = "application/json";
-                if (streamOptions.schema) {
-                    const rawSchema = convertZodToJsonSchema(streamOptions.schema);
-                    const inlinedSchema = inlineJsonSchema(rawSchema);
-                    if (inlinedSchema.$schema) {
-                        delete inlinedSchema.$schema;
-                    }
-                    config.responseSchema = inlinedSchema;
-                    logger.debug("[GoogleVertex] Added responseSchema for JSON output (stream)", {
-                        schemaKeys: Object.keys(inlinedSchema),
-                    });
+        }
+        const startTime = Date.now();
+        const timeoutController = createTimeoutController(this.getTimeout(options), this.providerName, "stream");
+        const composedSignal = composeAbortSignals(options.abortSignal, timeoutController?.controller.signal);
+        const maxSteps = computeMaxStepsShared(options.maxSteps);
+        const currentContents = this.prependConversationHistory([...contents], options.conversationMessages);
+        const channel = createTextChannel();
+        const allToolCalls = [];
+        const metadata = {
+            streamId: `native-vertex-${Date.now()}`,
+            startTime,
+            responseTime: 0,
+            totalToolExecutions: 0,
+        };
+        let analyticsResolve;
+        let analyticsReject;
+        const analyticsPromise = new Promise((resolve, reject) => {
+            analyticsResolve = resolve;
+            analyticsReject = reject;
+        });
+        const loopPromise = this.runNativeGemini3StreamLoop({
+            client,
+            modelName,
+            span,
+            config,
+            currentContents,
+            executeMap,
+            channel,
+            allToolCalls,
+            metadata,
+            analyticsResolve,
+            analyticsReject,
+            startTime,
+            timeoutController,
+            composedSignal,
+            maxSteps,
+        });
+        loopPromise.catch(() => undefined);
+        return {
+            stream: channel.iterable,
+            provider: this.providerName,
+            model: modelName,
+            toolCalls: allToolCalls,
+            analytics: analyticsPromise,
+            metadata,
+        };
+    }
+    async runNativeGemini3StreamLoop(params) {
+        let lastStepText = "";
+        let totalInputTokens = 0;
+        let totalOutputTokens = 0;
+        let step = 0;
+        let completedWithFinalAnswer = false;
+        const failedTools = new Map();
+        try {
+            while (step < params.maxSteps) {
+                if (params.composedSignal?.aborted) {
+                    throw params.composedSignal.reason instanceof Error
+                        ? params.composedSignal.reason
+                        : new Error("Request aborted");
                 }
-            }
-            const startTime = Date.now();
-            const timeout = this.getTimeout(options);
-            const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-            const composedSignal = composeAbortSignals(options.abortSignal, timeoutController?.controller.signal);
-            const maxSteps = computeMaxStepsShared(options.maxSteps);
-            // Inject conversation history so the native path has multi-turn context
-            const currentContents = this.prependConversationHistory([...contents], options.conversationMessages);
-            // Create a push-based text channel so the caller receives tokens as
-            // they arrive from the network rather than after full buffering.
-            const channel = createTextChannel();
-            // Shared mutable state updated by the background agentic loop.
-            const allToolCalls = [];
-            // Shared metadata object mutated by the background loop so that
-            // responseTime and totalToolExecutions reflect final values.
-            const metadata = {
-                streamId: `native-vertex-${Date.now()}`,
-                startTime,
-                responseTime: 0,
-                totalToolExecutions: 0,
-            };
-            // analyticsResolvers lets the background loop settle the analytics
-            // promise once token counts are known (after the loop completes).
-            let analyticsResolve;
-            let analyticsReject;
-            const analyticsPromise = new Promise((res, rej) => {
-                analyticsResolve = res;
-                analyticsReject = rej;
-            });
-            // Run the agentic loop in the background without awaiting it here,
-            // so we can return the StreamResult (with channel.iterable) immediately.
-            const loopPromise = (async () => {
-                let lastStepText = "";
-                let totalInputTokens = 0;
-                let totalOutputTokens = 0;
-                let step = 0;
-                let completedWithFinalAnswer = false;
-                const failedTools = new Map();
+                step++;
+                logger.debug(`[GoogleVertex] Native SDK step ${step}/${params.maxSteps}`);
                 try {
-                    // Agentic loop for tool calling
-                    while (step < maxSteps) {
-                        if (composedSignal?.aborted) {
-                            throw composedSignal.reason instanceof Error
-                                ? composedSignal.reason
-                                : new Error("Request aborted");
-                        }
-                        step++;
-                        logger.debug(`[GoogleVertex] Native SDK step ${step}/${maxSteps}`);
-                        try {
-                            const rawStream = await client.models.generateContentStream({
-                                model: modelName,
-                                contents: currentContents,
-                                config,
-                                ...(composedSignal
-                                    ? { httpOptions: { signal: composedSignal } }
-                                    : {}),
-                            });
-                            // For every step, use incremental collection so text parts
-                            // are pushed to the channel as they arrive.  For intermediate
-                            // steps (those that produce function calls) we still need the
-                            // complete rawResponseParts for pushModelResponseToHistory,
-                            // which collectStreamChunksIncremental provides at stream end.
-                            const chunkResult = await collectStreamChunksIncremental(rawStream, channel);
-                            totalInputTokens += chunkResult.inputTokens;
-                            totalOutputTokens += chunkResult.outputTokens;
-                            const stepText = extractTextFromParts(chunkResult.rawResponseParts);
-                            // If no function calls, this was the final step — channel
-                            // already received all text parts incrementally.
-                            if (chunkResult.stepFunctionCalls.length === 0) {
-                                completedWithFinalAnswer = true;
-                                break;
-                            }
-                            lastStepText = stepText;
-                            // Record tool call events on the span
-                            for (const fc of chunkResult.stepFunctionCalls) {
-                                span.addEvent("gen_ai.tool_call", {
-                                    "tool.name": fc.name,
-                                    "tool.step": step,
-                                });
-                            }
-                            logger.debug(`[GoogleVertex] Executing ${chunkResult.stepFunctionCalls.length} function calls`);
-                            pushModelResponseToHistory(currentContents, chunkResult.rawResponseParts, chunkResult.stepFunctionCalls);
-                            const functionResponses = await executeNativeToolCalls("[GoogleVertex]", chunkResult.stepFunctionCalls, executeMap, failedTools, allToolCalls, { abortSignal: composedSignal });
-                            // Function/tool responses must use role: "user" — the
-                            // @google/genai SDK's validateHistory() only accepts "user"
-                            // and "model" roles (matching automaticFunctionCalling).
-                            currentContents.push({
-                                role: "user",
-                                parts: functionResponses,
-                            });
-                        }
-                        catch (error) {
-                            logger.error("[GoogleVertex] Native SDK error", error);
-                            throw this.handleProviderError(error);
-                        }
+                    const rawStream = await params.client.models.generateContentStream({
+                        model: params.modelName,
+                        contents: params.currentContents,
+                        config: params.config,
+                        ...(params.composedSignal
+                            ? { httpOptions: { signal: params.composedSignal } }
+                            : {}),
+                    });
+                    const chunkResult = await collectStreamChunksIncremental(rawStream, params.channel);
+                    totalInputTokens += chunkResult.inputTokens;
+                    totalOutputTokens += chunkResult.outputTokens;
+                    const stepText = extractTextFromParts(chunkResult.rawResponseParts);
+                    if (chunkResult.stepFunctionCalls.length === 0) {
+                        completedWithFinalAnswer = true;
+                        break;
                     }
-                    // Handle max-steps termination: if the model was still calling
-                    // tools when we hit the limit, push a synthetic final message.
-                    if (step >= maxSteps && !completedWithFinalAnswer) {
-                        const fallback = handleMaxStepsTermination("[GoogleVertex]", step, maxSteps, "", // finalText is empty — model didn't stop on its own
-                        lastStepText);
-                        if (fallback) {
-                            channel.push(fallback);
-                        }
+                    lastStepText = stepText;
+                    for (const functionCall of chunkResult.stepFunctionCalls) {
+                        params.span.addEvent("gen_ai.tool_call", {
+                            "tool.name": functionCall.name,
+                            "tool.step": step,
+                        });
                     }
-                    const responseTime = Date.now() - startTime;
-                    // Propagate final values to the shared metadata object so that
-                    // the already-returned StreamResult reflects accurate telemetry.
-                    metadata.responseTime = responseTime;
-                    metadata.totalToolExecutions = allToolCalls.length;
-                    // Set token usage and finish reason on the span
-                    span.setAttribute(ATTR.GEN_AI_INPUT_TOKENS, totalInputTokens);
-                    span.setAttribute(ATTR.GEN_AI_OUTPUT_TOKENS, totalOutputTokens);
-                    span.setAttribute(ATTR.GEN_AI_FINISH_REASON, step >= maxSteps && !completedWithFinalAnswer
-                        ? "max_steps"
-                        : "stop");
-                    analyticsResolve({
-                        provider: this.providerName,
-                        model: modelName,
-                        tokenUsage: {
-                            input: totalInputTokens,
-                            output: totalOutputTokens,
-                            total: totalInputTokens + totalOutputTokens,
-                        },
-                        requestDuration: responseTime,
-                        timestamp: new Date().toISOString(),
+                    logger.debug(`[GoogleVertex] Executing ${chunkResult.stepFunctionCalls.length} function calls`);
+                    pushModelResponseToHistory(params.currentContents, chunkResult.rawResponseParts, chunkResult.stepFunctionCalls);
+                    const functionResponses = await executeNativeToolCalls("[GoogleVertex]", chunkResult.stepFunctionCalls, params.executeMap, failedTools, params.allToolCalls, { abortSignal: params.composedSignal });
+                    params.currentContents.push({
+                        role: "user",
+                        parts: functionResponses,
                     });
-                    channel.close();
                 }
-                catch (err) {
-                    channel.error(err);
-                    analyticsReject(err);
+                catch (error) {
+                    logger.error("[GoogleVertex] Native SDK error", error);
+                    throw this.handleProviderError(error);
                 }
-                finally {
-                    timeoutController?.cleanup();
+            }
+            if (step >= params.maxSteps && !completedWithFinalAnswer) {
+                const fallback = handleMaxStepsTermination("[GoogleVertex]", step, params.maxSteps, "", lastStepText);
+                if (fallback) {
+                    params.channel.push(fallback);
                 }
-            })();
-            // Suppress unhandled-rejection warnings on loopPromise — errors are
-            // forwarded to the channel and will surface when the caller iterates.
-            loopPromise.catch(() => undefined);
-            return {
-                stream: channel.iterable,
+            }
+            const responseTime = Date.now() - params.startTime;
+            params.metadata.responseTime = responseTime;
+            params.metadata.totalToolExecutions = params.allToolCalls.length;
+            params.span.setAttribute(ATTR.GEN_AI_INPUT_TOKENS, totalInputTokens);
+            params.span.setAttribute(ATTR.GEN_AI_OUTPUT_TOKENS, totalOutputTokens);
+            params.span.setAttribute(ATTR.GEN_AI_FINISH_REASON, step >= params.maxSteps && !completedWithFinalAnswer
+                ? "max_steps"
+                : "stop");
+            params.analyticsResolve({
                 provider: this.providerName,
-                model: modelName,
-                toolCalls: allToolCalls,
-                analytics: analyticsPromise,
-                metadata,
-            };
-        });
+                model: params.modelName,
+                tokenUsage: {
+                    input: totalInputTokens,
+                    output: totalOutputTokens,
+                    total: totalInputTokens + totalOutputTokens,
+                },
+                requestDuration: responseTime,
+                timestamp: new Date().toISOString(),
+            });
+            params.channel.close();
+        }
+        catch (error) {
+            params.channel.error(error);
+            params.analyticsReject(error);
+        }
+        finally {
+            params.timeoutController?.cleanup();
+        }
     }
     /**
      * Execute generate using native @google/genai SDK for Gemini 3 models on Vertex AI