@juspay/neurolink 9.59.4 → 9.59.6

package/dist/core/baseProvider.js

@@ -570,6 +570,19 @@ export class BaseProvider {
         if (!hasVideoFrames(messages)) {
             return null;
         }
+        // Bug 2 fix: callers requesting structured output (schema or explicit
+        // output.format) must NOT be hijacked into the prose-returning video
+        // analysis path. Without this gate, schema/format are silently dropped
+        // whenever messages contain >=3 image parts.
+        if (options.schema !== undefined || options.output?.format !== undefined) {
+            logger.info("[VideoFrameGen] Skipping video-frame analysis route; caller requested structured output", {
+                provider: this.providerName,
+                model: this.modelName,
+                hasSchema: options.schema !== undefined,
+                outputFormat: options.output?.format,
+            });
+            return null;
+        }
         const videoAnalysisResult = await executeVideoAnalysis(messages, {
             provider: options.provider,
             providerName: this.providerName,

package/dist/lib/core/baseProvider.js

@@ -570,6 +570,19 @@ export class BaseProvider {
         if (!hasVideoFrames(messages)) {
             return null;
         }
+        // Bug 2 fix: callers requesting structured output (schema or explicit
+        // output.format) must NOT be hijacked into the prose-returning video
+        // analysis path. Without this gate, schema/format are silently dropped
+        // whenever messages contain >=3 image parts.
+        if (options.schema !== undefined || options.output?.format !== undefined) {
+            logger.info("[VideoFrameGen] Skipping video-frame analysis route; caller requested structured output", {
+                provider: this.providerName,
+                model: this.modelName,
+                hasSchema: options.schema !== undefined,
+                outputFormat: options.output?.format,
+            });
+            return null;
+        }
         const videoAnalysisResult = await executeVideoAnalysis(messages, {
             provider: options.provider,
             providerName: this.providerName,

package/dist/lib/neurolink.js

@@ -7255,7 +7255,7 @@ Current user's request: ${currentInput}`;
         const finalOptions = {
             timeout: options?.timeout ??
                 toolInfo?.tool?.timeoutMs ??
-                TOOL_TIMEOUTS.EXECUTION_DEFAULT_MS,
+                TOOL_TIMEOUTS.EXECUTION_BATCH_MS,
             maxRetries: options?.maxRetries ??
                 toolInfo?.tool?.maxRetries ??
                 RETRY_ATTEMPTS.DEFAULT,

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

@@ -311,7 +311,14 @@ export declare class GoogleVertexProvider extends BaseProvider {
      */
     private buildImageGenerationParts;
     /**
-     * Parse the Vertex AI image generation REST API response and extract image data.
+     * Parse the Vertex AI image generation REST API response.
+     *
+     * Dual-mode image models (gemini-3.1-flash-image-preview, gemini-2.5-flash-image,
+     * gemini-3-pro-image-preview) decide per-request whether to emit an image or text.
+     * When the response contains text parts but no image part, surface the text via
+     * `textFallback` so the caller can return a normal text result instead of throwing
+     * "model returned text instead of image data" and burning retries on a query that
+     * the model has already answered.
      */
     private parseImageGenerationResponse;
     /**

package/dist/lib/providers/googleVertex.js

@@ -3015,7 +3015,14 @@ export class GoogleVertexProvider extends BaseProvider {
         return parts;
     }
     /**
-     * Parse the Vertex AI image generation REST API response and extract image data.
+     * Parse the Vertex AI image generation REST API response.
+     *
+     * Dual-mode image models (gemini-3.1-flash-image-preview, gemini-2.5-flash-image,
+     * gemini-3-pro-image-preview) decide per-request whether to emit an image or text.
+     * When the response contains text parts but no image part, surface the text via
+     * `textFallback` so the caller can return a normal text result instead of throwing
+     * "model returned text instead of image data" and burning retries on a query that
+     * the model has already answered.
      */
     parseImageGenerationResponse(data, imageModelName) {
         const candidate = data.candidates?.[0];
@@ -3030,10 +3037,15 @@ export class GoogleVertexProvider extends BaseProvider {
                 (part.inline_data &&
                     part.inline_data.mime_type?.startsWith("image/"))));
         if (!imagePart) {
-            const hasTextContent = candidate.content.parts.some((part) => part.text);
-            throw new ProviderError(hasTextContent
-                ? `Image generation completed but model returned text instead of image data. Model: ${imageModelName}`
-                : `Image generation completed but no image data was returned. Model: ${imageModelName}`, this.providerName);
+            // Filter out empty/whitespace-only text parts so an effectively empty
+            // response throws "no image data" instead of returning content: "".
+            const textParts = candidate.content.parts
+                .map((part) => (typeof part.text === "string" ? part.text : ""))
+                .filter((text) => text.trim().length > 0);
+            if (textParts.length > 0) {
+                return { textFallback: textParts.join("") };
+            }
+            throw new ProviderError(`Image generation completed but no image data was returned. Model: ${imageModelName}`, this.providerName);
         }
         const imageData = imagePart.inlineData?.data || imagePart.inline_data?.data;
         const mimeType = imagePart.inlineData?.mimeType ||
@@ -3127,7 +3139,31 @@ export class GoogleVertexProvider extends BaseProvider {
                 throw new ProviderError(`Vertex AI API error (${response.status}): ${errorText}`, this.providerName);
             }
             const data = (await response.json());
-            const { imageData, mimeType } = this.parseImageGenerationResponse(data, imageModelName);
+            const parsed = this.parseImageGenerationResponse(data, imageModelName);
+            // Dual-mode model decided to emit text instead of an image. Surface the
+            // text as a normal text result instead of throwing — the model already
+            // answered the user; failing here just burns retries.
+            if ("textFallback" in parsed) {
+                logger.info("Dual-mode image model returned text; returning as text result", {
+                    model: imageModelName,
+                    textLength: parsed.textFallback.length,
+                    responseTime: Date.now() - startTime,
+                });
+                const inputTokens = this.estimateTokenCount(prompt);
+                const outputTokens = this.estimateTokenCount(parsed.textFallback);
+                const textResult = {
+                    content: parsed.textFallback,
+                    provider: this.providerName,
+                    model: imageModelName,
+                    usage: {
+                        input: inputTokens,
+                        output: outputTokens,
+                        total: inputTokens + outputTokens,
+                    },
+                };
+                return await this.enhanceResult(textResult, options, startTime);
+            }
+            const { imageData, mimeType } = parsed;
             logger.info("Image generation successful", {
                 model: imageModelName,
                 mimeType,

package/dist/neurolink.js

@@ -7255,7 +7255,7 @@ Current user's request: ${currentInput}`;
         const finalOptions = {
             timeout: options?.timeout ??
                 toolInfo?.tool?.timeoutMs ??
-                TOOL_TIMEOUTS.EXECUTION_DEFAULT_MS,
+                TOOL_TIMEOUTS.EXECUTION_BATCH_MS,
             maxRetries: options?.maxRetries ??
                 toolInfo?.tool?.maxRetries ??
                 RETRY_ATTEMPTS.DEFAULT,

package/dist/providers/googleVertex.d.ts

@@ -311,7 +311,14 @@ export declare class GoogleVertexProvider extends BaseProvider {
      */
     private buildImageGenerationParts;
     /**
-     * Parse the Vertex AI image generation REST API response and extract image data.
+     * Parse the Vertex AI image generation REST API response.
+     *
+     * Dual-mode image models (gemini-3.1-flash-image-preview, gemini-2.5-flash-image,
+     * gemini-3-pro-image-preview) decide per-request whether to emit an image or text.
+     * When the response contains text parts but no image part, surface the text via
+     * `textFallback` so the caller can return a normal text result instead of throwing
+     * "model returned text instead of image data" and burning retries on a query that
+     * the model has already answered.
      */
     private parseImageGenerationResponse;
     /**

package/dist/providers/googleVertex.js

@@ -3015,7 +3015,14 @@ export class GoogleVertexProvider extends BaseProvider {
         return parts;
     }
     /**
-     * Parse the Vertex AI image generation REST API response and extract image data.
+     * Parse the Vertex AI image generation REST API response.
+     *
+     * Dual-mode image models (gemini-3.1-flash-image-preview, gemini-2.5-flash-image,
+     * gemini-3-pro-image-preview) decide per-request whether to emit an image or text.
+     * When the response contains text parts but no image part, surface the text via
+     * `textFallback` so the caller can return a normal text result instead of throwing
+     * "model returned text instead of image data" and burning retries on a query that
+     * the model has already answered.
      */
     parseImageGenerationResponse(data, imageModelName) {
         const candidate = data.candidates?.[0];
@@ -3030,10 +3037,15 @@ export class GoogleVertexProvider extends BaseProvider {
                 (part.inline_data &&
                     part.inline_data.mime_type?.startsWith("image/"))));
         if (!imagePart) {
-            const hasTextContent = candidate.content.parts.some((part) => part.text);
-            throw new ProviderError(hasTextContent
-                ? `Image generation completed but model returned text instead of image data. Model: ${imageModelName}`
-                : `Image generation completed but no image data was returned. Model: ${imageModelName}`, this.providerName);
+            // Filter out empty/whitespace-only text parts so an effectively empty
+            // response throws "no image data" instead of returning content: "".
+            const textParts = candidate.content.parts
+                .map((part) => (typeof part.text === "string" ? part.text : ""))
+                .filter((text) => text.trim().length > 0);
+            if (textParts.length > 0) {
+                return { textFallback: textParts.join("") };
+            }
+            throw new ProviderError(`Image generation completed but no image data was returned. Model: ${imageModelName}`, this.providerName);
         }
         const imageData = imagePart.inlineData?.data || imagePart.inline_data?.data;
         const mimeType = imagePart.inlineData?.mimeType ||
@@ -3127,7 +3139,31 @@ export class GoogleVertexProvider extends BaseProvider {
                 throw new ProviderError(`Vertex AI API error (${response.status}): ${errorText}`, this.providerName);
             }
             const data = (await response.json());
-            const { imageData, mimeType } = this.parseImageGenerationResponse(data, imageModelName);
+            const parsed = this.parseImageGenerationResponse(data, imageModelName);
+            // Dual-mode model decided to emit text instead of an image. Surface the
+            // text as a normal text result instead of throwing — the model already
+            // answered the user; failing here just burns retries.
+            if ("textFallback" in parsed) {
+                logger.info("Dual-mode image model returned text; returning as text result", {
+                    model: imageModelName,
+                    textLength: parsed.textFallback.length,
+                    responseTime: Date.now() - startTime,
+                });
+                const inputTokens = this.estimateTokenCount(prompt);
+                const outputTokens = this.estimateTokenCount(parsed.textFallback);
+                const textResult = {
+                    content: parsed.textFallback,
+                    provider: this.providerName,
+                    model: imageModelName,
+                    usage: {
+                        input: inputTokens,
+                        output: outputTokens,
+                        total: inputTokens + outputTokens,
+                    },
+                };
+                return await this.enhanceResult(textResult, options, startTime);
+            }
+            const { imageData, mimeType } = parsed;
             logger.info("Image generation successful", {
                 model: imageModelName,
                 mimeType,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.59.4",
+  "version": "9.59.6",
   "packageManager": "pnpm@10.15.1",
   "description": "Universal AI Development Platform with working MCP integration, multi-provider support, and professional CLI. Built-in tools operational, 58+ external MCP servers discoverable. Connect to filesystem, GitHub, database operations, and more. Build, test, and deploy AI applications with 13 providers: OpenAI, Anthropic, Google AI, AWS Bedrock, Azure, Hugging Face, Ollama, and Mistral AI.",
   "author": {