@juspay/neurolink 9.50.0 → 9.50.2

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/services/server/ai/observability/instrumentation.js

@@ -445,7 +445,45 @@ function initializeExternalOpenTelemetryMode(config, resource, otlpEndpoint, ser
             const provider = globalProvider;
             if (globalProvider && typeof provider.addSpanProcessor === "function") {
                 provider.addSpanProcessor(new ContextEnricher());
-                const skipLangfuse = config.skipLangfuseSpanProcessor === true || !langfuseProcessor;
+                // Auto-detect: skip if consumer already registered a LangfuseSpanProcessor.
+                //
+                // Detection strategy (ordered by robustness):
+                // 1. `instanceof LangfuseSpanProcessor` — reliable when both sides use
+                //    the same @langfuse/otel package instance (same module identity).
+                // 2. Duck-type check for Langfuse-specific public member
+                //    (`langfuseClient` property) — survives minification.
+                // 3. `constructor.name === "LangfuseSpanProcessor"` — last resort,
+                //    brittle under minification or bundler renaming.
+                //
+                // NOTE: `_registeredSpanProcessors` is an internal OpenTelemetry field.
+                // If the OTel SDK removes or renames it, the array defaults to [] and
+                // `hasExistingLangfuse` is false — NeuroLink registers its own processor
+                // (same behavior as before this check). Consumers can always force skip
+                // via `skipLangfuseSpanProcessor: true`.
+                const existingProcessors = provider
+                    ._registeredSpanProcessors ?? [];
+                const hasExistingLangfuse = existingProcessors.some((p) => {
+                    if (p === null || p === undefined || typeof p !== "object") {
+                        return false;
+                    }
+                    // Prefer instanceof — works when same @langfuse/otel package is shared
+                    if (p instanceof LangfuseSpanProcessor) {
+                        return true;
+                    }
+                    // Duck-type: Langfuse processor exposes a langfuseClient property
+                    if ("langfuseClient" in p) {
+                        return true;
+                    }
+                    // Fallback: constructor name (brittle under minification)
+                    return (p.constructor?.name ===
+                        "LangfuseSpanProcessor");
+                });
+                const skipLangfuse = config.skipLangfuseSpanProcessor === true ||
+                    !langfuseProcessor ||
+                    hasExistingLangfuse;
+                if (hasExistingLangfuse && !config.skipLangfuseSpanProcessor) {
+                    logger.info(`${LOG_PREFIX} Auto-detected existing LangfuseSpanProcessor — skipping SDK registration to avoid duplicates`);
+                }
                 if (!skipLangfuse && langfuseProcessor) {
                     provider.addSpanProcessor(langfuseProcessor);
                 }

package/dist/types/externalMcp.d.ts

@@ -228,6 +228,7 @@ export type ExternalMCPServerEvents = {
     /** Server status changed */
     statusChanged: {
         serverId: string;
+        serverName: string;
         oldStatus: ExternalMCPServerStatus;
         newStatus: ExternalMCPServerStatus;
         timestamp: Date;
@@ -235,24 +236,28 @@ export type ExternalMCPServerEvents = {
     /** Server connected successfully */
     connected: {
         serverId: string;
+        serverName: string;
         toolCount: number;
         timestamp: Date;
     };
     /** Server disconnected */
     disconnected: {
         serverId: string;
+        serverName: string;
         reason?: string;
         timestamp: Date;
     };
     /** Server failed */
     failed: {
         serverId: string;
+        serverName: string;
         error: string;
         timestamp: Date;
     };
     /** Tool discovered */
     toolDiscovered: {
         serverId: string;
+        serverName: string;
         toolName: string;
         toolInfo: ExternalMCPToolInfo;
         timestamp: Date;
@@ -260,12 +265,14 @@ export type ExternalMCPServerEvents = {
     /** Tool removed */
     toolRemoved: {
         serverId: string;
+        serverName: string;
         toolName: string;
         timestamp: Date;
     };
     /** Health check completed */
     healthCheck: {
         serverId: string;
+        serverName: string;
         health: ExternalMCPServerHealth;
         timestamp: Date;
     };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.50.0",
+  "version": "9.50.2",
   "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": {