@juspay/neurolink 9.49.0 → 9.50.1

package/dist/server/routes/claudeProxyRoutes.js

@@ -15,7 +15,7 @@ import { join } from "node:path";
 import { buildStableClaudeCodeBillingHeader, CLAUDE_CLI_USER_AGENT, CLAUDE_CODE_OAUTH_BETAS, getOrCreateClaudeCodeIdentity, parseClaudeCodeUserId, } from "../../auth/anthropicOAuth.js";
 import { parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
 import { buildClaudeError, ClaudeStreamSerializer, generateToolUseId, parseClaudeRequest, serializeClaudeResponse, } from "../../proxy/claudeFormat.js";
-import { ProxyTracer } from "../../proxy/proxyTracer.js";
+import { ProxyTracer, recordFallbackAttempt } from "../../proxy/proxyTracer.js";
 import { createRawStreamCapture } from "../../proxy/rawStreamCapture.js";
 import { logBodyCapture, logRequest, logRequestAttempt, logStreamError, } from "../../proxy/requestLogger.js";
 import { createSSEInterceptor } from "../../proxy/sseInterceptor.js";
@@ -1246,43 +1246,64 @@ async function executeClaudeFallbackTranslation(args) {
     if (body.stream) {
         const streamResult = await ctx.neurolink.stream(options);
         const serializer = new ClaudeStreamSerializer(body.model, 0);
-        async function* sseGenerator() {
-            for (const frame of serializer.start()) {
-                yield frame;
-            }
-            let collectedText = "";
-            for await (const chunk of streamResult.stream) {
-                const text = extractText(chunk);
-                if (text) {
-                    collectedText += text;
-                    for (const frame of serializer.pushDelta(text)) {
-                        yield frame;
-                    }
+        // Eagerly consume stream so errors fire synchronously and the
+        // fallback loop in tryConfiguredClaudeFallbackChain can catch them.
+        const frames = [];
+        let collectedText = "";
+        for (const frame of serializer.start()) {
+            frames.push(frame);
+        }
+        for await (const chunk of streamResult.stream) {
+            const text = extractText(chunk);
+            if (text) {
+                collectedText += text;
+                for (const frame of serializer.pushDelta(text)) {
+                    frames.push(frame);
                 }
             }
-            const toolCalls = streamResult.toolCalls ?? [];
-            if (!hasTranslatedOutput(collectedText, toolCalls)) {
-                throw new Error(`Translated provider ${providerLabel} returned no content or tool calls`);
-            }
-            if (toolCalls.length) {
-                for (const toolCall of toolCalls) {
-                    const toolName = toolCall.toolName ??
-                        toolCall.name ??
-                        "unknown";
-                    for (const frame of serializer.pushToolUse(generateToolUseId(), toolName, extractToolArgs(toolCall))) {
-                        yield frame;
-                    }
+        }
+        const toolCalls = streamResult.toolCalls ?? [];
+        if (!hasTranslatedOutput(collectedText, toolCalls)) {
+            throw new Error(`Translated provider ${providerLabel} returned no content or tool calls`);
+        }
+        if (toolCalls.length) {
+            for (const toolCall of toolCalls) {
+                const toolName = toolCall.toolName ??
+                    toolCall.name ??
+                    "unknown";
+                for (const frame of serializer.pushToolUse(generateToolUseId(), toolName, extractToolArgs(toolCall))) {
+                    frames.push(frame);
                 }
             }
-            const reason = streamResult.finishReason ?? "end_turn";
-            const resolvedUsage = extractUsageFromStreamResult(streamResult.usage);
-            for (const frame of serializer.finish(resolvedUsage.output, reason)) {
-                yield frame;
-            }
         }
+        const reason = streamResult.finishReason ?? "end_turn";
+        const resolvedUsage = extractUsageFromStreamResult(streamResult.usage);
+        for (const frame of serializer.finish(resolvedUsage.output, reason)) {
+            frames.push(frame);
+        }
+        // Telemetry AFTER validation — not before like the old lazy path
         tracer?.end(200, Date.now() - requestStartTime);
         recordFinalSuccess();
-        logFinalRequest(200, "", providerLabel);
+        logFinalRequest(200, "", providerLabel, undefined, undefined, {
+            inputTokens: resolvedUsage.input,
+            outputTokens: resolvedUsage.output,
+        });
+        const bufferedBody = frames.join("");
+        logProxyBody({
+            phase: "client_response",
+            headers: { "content-type": "text/event-stream" },
+            body: bufferedBody,
+            bodySize: Buffer.byteLength(bufferedBody, "utf8"),
+            contentType: "text/event-stream",
+            responseStatus: 200,
+            durationMs: Date.now() - requestStartTime,
+        });
+        // Return generator that yields pre-buffered frames
+        async function* sseGenerator() {
+            for (const frame of frames) {
+                yield frame;
+            }
+        }
         return sseGenerator();
     }
     const streamResult = await ctx.neurolink.stream(options);
@@ -1346,6 +1367,11 @@ async function tryConfiguredClaudeFallbackChain(args) {
             : "auto-provider";
         logger.always(`[proxy] skipping fallback ${label}: ${skipped.reason}`);
     }
+    tracer?.setFallbackInfo({
+        triggered: true,
+        attemptCount: fallbackPlan.attempts.slice(1).length,
+        reason: fallbackPolicyReason ?? "all_anthropic_accounts_exhausted",
+    });
     for (const fallback of fallbackPlan.attempts.slice(1)) {
         if (!fallback.provider || !fallback.model) {
             continue;
@@ -1354,6 +1380,7 @@ async function tryConfiguredClaudeFallbackChain(args) {
         if (!availability.available) {
             logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} health-check failed (${availability.reason ?? "provider unavailable"}), attempting anyway`);
         }
+        const fallbackStart = Date.now();
         try {
             logger.always(`[proxy] fallback → ${fallback.provider}/${fallback.model}`);
             const options = buildProxyFallbackOptions(parsedFallbackRequest, {
@@ -1370,13 +1397,57 @@ async function tryConfiguredClaudeFallbackChain(args) {
                 options: options,
                 providerLabel: fallback.provider,
             });
+            recordFallbackAttempt({
+                provider: fallback.provider,
+                model: fallback.model,
+                status: "success",
+                durationMs: Date.now() - fallbackStart,
+            });
+            tracer?.setFallbackInfo({
+                triggered: true,
+                provider: fallback.provider,
+                model: fallback.model,
+                attemptCount: fallbackPlan.attempts.slice(1).length,
+                reason: "fallback_success",
+            });
             return {
                 response,
                 fallbackPolicyReason,
             };
         }
         catch (fallbackErr) {
-            logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} failed: ${fallbackErr instanceof Error ? fallbackErr.message : String(fallbackErr)}`);
+            const errMsg = fallbackErr instanceof Error
+                ? fallbackErr.message
+                : String(fallbackErr);
+            let errorClass = "unknown";
+            if (errMsg.includes("Rate limit") ||
+                errMsg.includes("rate_limit") ||
+                errMsg.includes("max_parallel_requests")) {
+                errorClass = "rate_limit";
+            }
+            else if (errMsg.includes("context length") ||
+                errMsg.includes("ContextWindowExceeded")) {
+                errorClass = "context_overflow";
+            }
+            else if (errMsg.includes("no content or tool calls") ||
+                errMsg.includes("NoOutputGenerated")) {
+                errorClass = "empty_response";
+            }
+            else if (errMsg.includes("thinking_level") ||
+                errMsg.includes("Field required")) {
+                errorClass = "schema_mismatch";
+            }
+            else if (errMsg.includes("Resource exhausted")) {
+                errorClass = "provider_quota";
+            }
+            logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} failed [${errorClass}]: ${errMsg}`);
+            recordFallbackAttempt({
+                provider: fallback.provider,
+                model: fallback.model,
+                status: "failure",
+                errorMessage: `[${errorClass}] ${errMsg}`,
+                durationMs: Date.now() - fallbackStart,
+            });
         }
     }
     return {
@@ -3541,7 +3612,15 @@ function shouldOmitImagesForTarget(provider, model) {
     return provider === "litellm" && model === "open-large";
 }
 function shouldOmitThinkingConfigForTarget(provider, model) {
-    return provider === "vertex" && model === "gemini-2.5-flash";
+    if (provider === "litellm") {
+        return true;
+    }
+    if (provider !== "vertex") {
+        return false;
+    }
+    // Only Gemini 2.5+ and 3.x support thinking_level on Vertex.
+    const m = model?.toLowerCase() ?? "";
+    return !/gemini-(2\.5|3)/.test(m);
 }
 function extractToolArgs(toolCall) {
     return (toolCall.args ??

package/dist/utils/imageCompressor.d.ts

@@ -0,0 +1,45 @@
+import type { ProviderName } from "../types/providers.js";
+declare const SUPPORTED_FORMATS: readonly ["jpeg", "png", "webp"];
+type SupportedFormat = (typeof SUPPORTED_FORMATS)[number];
+/**
+ * Provider-specific image size limits in bytes
+ */
+export declare const PROVIDER_IMAGE_LIMITS: Record<ProviderName, number>;
+export interface CompressionOptions {
+    provider: ProviderName;
+    quality?: number;
+    maxDimension?: number;
+    format?: SupportedFormat;
+}
+export interface CompressionResult {
+    buffer: Buffer;
+    originalSize: number;
+    compressedSize: number;
+    compressionRatio: number;
+    metadata: {
+        width: number;
+        height: number;
+        format: string;
+    };
+}
+/**
+ * Compress an image to meet provider-specific size limits
+ * @param imageBuffer - Input image buffer
+ * @param options - Compression options including provider name
+ * @returns Compressed image buffer with metadata
+ */
+export declare function compressImage(imageBuffer: Buffer, options: CompressionOptions): Promise<CompressionResult>;
+/**
+ * Check if an image needs compression for a specific provider
+ * @param imageBuffer - Input image buffer
+ * @param provider - AI provider name
+ * @returns True if compression is needed
+ */
+export declare function needsCompression(imageBuffer: Buffer, provider: ProviderName): boolean;
+/**
+ * Get the size limit for a specific provider
+ * @param provider - AI provider name
+ * @returns Size limit in bytes
+ */
+export declare function getProviderSizeLimit(provider: ProviderName): number;
+export {};

package/dist/utils/imageCompressor.js

@@ -0,0 +1,136 @@
+import sharp from "sharp";
+import { withTimeout } from "./async/index.js";
+const SUPPORTED_FORMATS = ["jpeg", "png", "webp"];
+const IMAGE_COMPRESSION_TIMEOUT_MS = 30_000;
+/**
+ * Provider-specific image size limits in bytes
+ */
+export const PROVIDER_IMAGE_LIMITS = {
+    openai: 20 * 1024 * 1024, // 20MB
+    "openai-compatible": 20 * 1024 * 1024, // 20MB (same as OpenAI)
+    anthropic: 5 * 1024 * 1024, // 5MB
+    "google-ai": 4 * 1024 * 1024, // 4MB
+    vertex: 4 * 1024 * 1024, // 4MB
+    bedrock: 5 * 1024 * 1024, // 5MB
+    azure: 20 * 1024 * 1024, // 20MB
+    mistral: 5 * 1024 * 1024, // 5MB
+    huggingface: 10 * 1024 * 1024, // 10MB
+    ollama: 100 * 1024 * 1024, // 100MB (local, no strict limit)
+    openrouter: 20 * 1024 * 1024, // 20MB
+    sagemaker: 5 * 1024 * 1024, // 5MB
+    litellm: 20 * 1024 * 1024, // 20MB (proxy, use OpenAI default)
+    auto: 5 * 1024 * 1024, // 5MB (conservative fallback)
+};
+/**
+ * Compress an image to meet provider-specific size limits
+ * @param imageBuffer - Input image buffer
+ * @param options - Compression options including provider name
+ * @returns Compressed image buffer with metadata
+ */
+export async function compressImage(imageBuffer, options) {
+    const { provider, quality = 80, maxDimension, format } = options;
+    const sizeLimit = PROVIDER_IMAGE_LIMITS[provider];
+    const originalSize = imageBuffer.length;
+    // Get original metadata
+    const image = sharp(imageBuffer);
+    const metadata = await withTimeout(image.metadata(), IMAGE_COMPRESSION_TIMEOUT_MS, "Timed out reading image metadata");
+    if (!metadata.width || !metadata.height) {
+        throw new Error("Unable to read image dimensions");
+    }
+    // If image is already under limit and no format conversion needed, return as-is
+    if (originalSize <= sizeLimit && !format && !maxDimension) {
+        return {
+            buffer: imageBuffer,
+            originalSize,
+            compressedSize: originalSize,
+            compressionRatio: 1,
+            metadata: {
+                width: metadata.width,
+                height: metadata.height,
+                format: metadata.format ?? "unknown",
+            },
+        };
+    }
+    // Prepare compression pipeline
+    let pipeline = sharp(imageBuffer);
+    // Resize if needed
+    if (maxDimension) {
+        const needsResize = metadata.width > maxDimension || metadata.height > maxDimension;
+        if (needsResize) {
+            pipeline = pipeline.resize(maxDimension, maxDimension, {
+                fit: "inside",
+                withoutEnlargement: true,
+            });
+        }
+    }
+    // Resolve target format — validate metadata.format against supported set
+    const rawFormat = metadata.format;
+    const targetFormat = format ??
+        (SUPPORTED_FORMATS.includes(rawFormat)
+            ? rawFormat
+            : "jpeg");
+    const applyFormat = (p, q) => {
+        switch (targetFormat) {
+            case "jpeg":
+                return p.jpeg({ quality: q, mozjpeg: true });
+            case "png":
+                return p.png({ quality: q, compressionLevel: 9 });
+            case "webp":
+                return p.webp({ quality: q });
+        }
+    };
+    // Compress
+    let compressedBuffer = await withTimeout(applyFormat(pipeline, quality).toBuffer(), IMAGE_COMPRESSION_TIMEOUT_MS, "Timed out compressing image");
+    let currentQuality = quality;
+    // Iteratively reduce quality if still over limit
+    // Note: the sharp pipeline must be rebuilt on each iteration because
+    // sharp does not support modifying quality settings after creation.
+    while (compressedBuffer.length > sizeLimit && currentQuality > 10) {
+        currentQuality -= 10;
+        let p = sharp(imageBuffer);
+        if (maxDimension) {
+            p = p.resize(maxDimension, maxDimension, {
+                fit: "inside",
+                withoutEnlargement: true,
+            });
+        }
+        compressedBuffer = await withTimeout(applyFormat(p, currentQuality).toBuffer(), IMAGE_COMPRESSION_TIMEOUT_MS, "Timed out compressing image");
+    }
+    // Final check
+    if (compressedBuffer.length > sizeLimit) {
+        throw new Error(`Unable to compress image to ${sizeLimit} bytes for provider ${provider}. ` +
+            `Final size: ${compressedBuffer.length} bytes. ` +
+            `Try using a smaller image or lower maxDimension.`);
+    }
+    // Get final metadata
+    const finalMetadata = await withTimeout(sharp(compressedBuffer).metadata(), IMAGE_COMPRESSION_TIMEOUT_MS, "Timed out reading compressed image metadata");
+    return {
+        buffer: compressedBuffer,
+        originalSize,
+        compressedSize: compressedBuffer.length,
+        compressionRatio: originalSize / compressedBuffer.length,
+        metadata: {
+            width: finalMetadata.width ?? 0,
+            height: finalMetadata.height ?? 0,
+            format: targetFormat,
+        },
+    };
+}
+/**
+ * Check if an image needs compression for a specific provider
+ * @param imageBuffer - Input image buffer
+ * @param provider - AI provider name
+ * @returns True if compression is needed
+ */
+export function needsCompression(imageBuffer, provider) {
+    const sizeLimit = PROVIDER_IMAGE_LIMITS[provider];
+    return imageBuffer.length > sizeLimit;
+}
+/**
+ * Get the size limit for a specific provider
+ * @param provider - AI provider name
+ * @returns Size limit in bytes
+ */
+export function getProviderSizeLimit(provider) {
+    return PROVIDER_IMAGE_LIMITS[provider];
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.49.0",
+  "version": "9.50.1",
   "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": {