@juspay/neurolink 12.33.0 → 12.34.1

package/dist/constants/contextWindows.js

@@ -83,6 +83,10 @@ export const MODEL_CONTEXT_WINDOWS = {
     voyage: {
         // Voyage embeddings: max input tokens vary 16K-32K per model
         _default: 32_000,
+        "voyage-4": 32_000,
+        "voyage-4-large": 32_000,
+        "voyage-4-lite": 32_000,
+        "voyage-code-4": 32_000,
         "voyage-3.5": 32_000,
         "voyage-3.5-lite": 32_000,
         "voyage-3-large": 32_000,
@@ -111,6 +115,10 @@ export const MODEL_CONTEXT_WINDOWS = {
     },
     anthropic: {
         _default: 200_000,
+        // Claude 5.5 / 5.1 (Sep 2026) — 1M context window, confirmed on
+        // platform.claude.com/docs/en/about-claude/models/overview
+        "claude-opus-5-5": 1_000_000,
+        "claude-fable-5-1": 1_000_000,
         // Claude 5 (mid 2026) — 1M context window
         "claude-sonnet-5": 1_000_000,
         // Claude 4.6 (Feb 2026) — 1M context window
@@ -134,6 +142,11 @@ export const MODEL_CONTEXT_WINDOWS = {
     },
     openai: {
         _default: 128_000,
+        // GPT-6 family (Sep 2026) — 1.05M context, per
+        // developers.openai.com/api/docs/models
+        "gpt-6-astra": 1_050_000,
+        "gpt-6-sol": 1_050_000,
+        "gpt-6-luna": 1_050_000,
         // GPT-5.4 family — 1.05M context
         "gpt-5.4": 1_050_000,
         "gpt-5.4-mini": 400_000,
@@ -202,6 +215,8 @@ export const MODEL_CONTEXT_WINDOWS = {
     vertex: {
         _default: 1_048_576,
         // Claude on Vertex
+        "claude-opus-5-5": 1_000_000,
+        "claude-fable-5-1": 1_000_000,
         "claude-sonnet-5": 1_000_000,
         "claude-opus-4-6": 1_000_000,
         "claude-sonnet-4-6": 1_000_000,
@@ -238,6 +253,9 @@ export const MODEL_CONTEXT_WINDOWS = {
     },
     bedrock: {
         _default: 200_000,
+        // Claude 5.5 / 5.1 (Sep 2026)
+        "anthropic.claude-opus-5-5": 1_000_000,
+        "anthropic.claude-fable-5-1": 1_000_000,
         // Claude 4.6
         "anthropic.claude-opus-4-6-v1": 1_000_000,
         "anthropic.claude-sonnet-4-6": 1_000_000,
@@ -263,6 +281,16 @@ export const MODEL_CONTEXT_WINDOWS = {
     },
     azure: {
         _default: 128_000,
+        // GPT-6 family (Sep 2026) — 1.05M context, per
+        // learn.microsoft.com/azure/ai-foundry/openai/concepts/models
+        "gpt-6-astra": 1_050_000,
+        "gpt-6-sol": 1_050_000,
+        "gpt-6-luna": 1_050_000,
+        // GPT-5.6 / GPT-5.5 — Azure-only as of Sep 2026, same 1.05M context
+        "gpt-5.6-sol": 1_050_000,
+        "gpt-5.6-terra": 1_050_000,
+        "gpt-5.6-luna": 1_050_000,
+        "gpt-5.5": 1_050_000,
         // GPT-5.4
         "gpt-5.4": 1_050_000,
         "gpt-5.4-mini": 400_000,

package/dist/constants/enums.d.ts

@@ -180,6 +180,9 @@ export declare enum BedrockModels {
  * Supported Models for OpenAI
  */
 export declare enum OpenAIModels {
+    GPT_6_ASTRA = "gpt-6-astra",
+    GPT_6_SOL = "gpt-6-sol",
+    GPT_6_LUNA = "gpt-6-luna",
     GPT_5_3_CODEX = "gpt-5.3-codex",
     GPT_5_4 = "gpt-5.4",
     GPT_5_4_MINI = "gpt-5.4-mini",
@@ -228,6 +231,13 @@ export declare enum OpenAIModels {
  * Note: Azure uses deployment names, these are model identifiers
  */
 export declare enum AzureOpenAIModels {
+    GPT_6_ASTRA = "gpt-6-astra",
+    GPT_6_SOL = "gpt-6-sol",
+    GPT_6_LUNA = "gpt-6-luna",
+    GPT_5_6_SOL = "gpt-5.6-sol",
+    GPT_5_6_TERRA = "gpt-5.6-terra",
+    GPT_5_6_LUNA = "gpt-5.6-luna",
+    GPT_5_5 = "gpt-5.5",
     GPT_5_2 = "gpt-5.2",
     GPT_5_2_CHAT = "gpt-5.2-chat",
     GPT_5_2_PRO = "gpt-5.2-pro",
@@ -355,6 +365,8 @@ export declare enum AnthropicModels {
     CLAUDE_OPUS_5 = "claude-opus-5",
     CLAUDE_SONNET_5 = "claude-sonnet-5",
     CLAUDE_FABLE_5 = "claude-fable-5",
+    CLAUDE_OPUS_5_5 = "claude-opus-5-5",
+    CLAUDE_FABLE_5_1 = "claude-fable-5-1",
     CLAUDE_OPUS_4_8 = "claude-opus-4-8",
     CLAUDE_OPUS_4_7 = "claude-opus-4-7",
     CLAUDE_OPUS_4_6 = "claude-opus-4-6",
@@ -650,6 +662,14 @@ export declare enum CohereModels {
  * @see https://docs.voyageai.com/docs/embeddings
  */
 export declare enum VoyageModels {
+    /** Voyage 4 — latest general-purpose flagship */
+    VOYAGE_4 = "voyage-4",
+    /** Voyage 4 Large — largest / highest quality */
+    VOYAGE_4_LARGE = "voyage-4-large",
+    /** Voyage 4 Lite — smaller / cheaper */
+    VOYAGE_4_LITE = "voyage-4-lite",
+    /** Voyage Code 4 — code-tuned */
+    VOYAGE_CODE_4 = "voyage-code-4",
     /** Voyage 3.5 — latest general-purpose (default) */
     VOYAGE_3_5 = "voyage-3.5",
     /** Voyage 3.5 Lite — smaller / cheaper */
@@ -1042,6 +1062,7 @@ export declare enum UpstageModels {
     SYN_PRO_251021 = "syn-pro-251021"
 }
 export declare enum XaiModels {
+    GROK_4_7 = "grok-4.7",
     GROK_4_6 = "grok-4.6",
     GROK_4_5 = "grok-4.5",
     GROK_4_3 = "grok-4.3",

package/dist/constants/enums.js

@@ -257,6 +257,11 @@ export var BedrockModels;
  */
 export var OpenAIModels;
 (function (OpenAIModels) {
+    // GPT-6 Series (Released September 2026) - current flagship family, per
+    // developers.openai.com/api/docs/models
+    OpenAIModels["GPT_6_ASTRA"] = "gpt-6-astra";
+    OpenAIModels["GPT_6_SOL"] = "gpt-6-sol";
+    OpenAIModels["GPT_6_LUNA"] = "gpt-6-luna";
     // GPT-5.3 Series (Released February 2026) - Latest coding models
     OpenAIModels["GPT_5_3_CODEX"] = "gpt-5.3-codex";
     // GPT-5.4 Series (Released March 2026) - Latest flagship models
@@ -318,6 +323,18 @@ export var OpenAIModels;
  */
 export var AzureOpenAIModels;
 (function (AzureOpenAIModels) {
+    // GPT-6 Series (September 2026) - current flagship family, confirmed on
+    // Azure OpenAI (learn.microsoft.com/azure/ai-foundry/openai/concepts/models)
+    AzureOpenAIModels["GPT_6_ASTRA"] = "gpt-6-astra";
+    AzureOpenAIModels["GPT_6_SOL"] = "gpt-6-sol";
+    AzureOpenAIModels["GPT_6_LUNA"] = "gpt-6-luna";
+    // GPT-5.6 Series (Azure-only as of Sep 2026 — retired from the OpenAI-direct
+    // catalog, still documented on Azure OpenAI)
+    AzureOpenAIModels["GPT_5_6_SOL"] = "gpt-5.6-sol";
+    AzureOpenAIModels["GPT_5_6_TERRA"] = "gpt-5.6-terra";
+    AzureOpenAIModels["GPT_5_6_LUNA"] = "gpt-5.6-luna";
+    // GPT-5.5 (Azure-only as of Sep 2026 — same situation as GPT-5.6 above)
+    AzureOpenAIModels["GPT_5_5"] = "gpt-5.5";
     // GPT-5.2 Series (Latest - December 2025)
     AzureOpenAIModels["GPT_5_2"] = "gpt-5.2";
     AzureOpenAIModels["GPT_5_2_CHAT"] = "gpt-5.2-chat";
@@ -482,6 +499,10 @@ export var AnthropicModels;
     AnthropicModels["CLAUDE_OPUS_5"] = "claude-opus-5";
     AnthropicModels["CLAUDE_SONNET_5"] = "claude-sonnet-5";
     AnthropicModels["CLAUDE_FABLE_5"] = "claude-fable-5";
+    // Claude 5.5 / 5.1 Series (September 2026) — confirmed on
+    // platform.claude.com/docs/en/about-claude/models/overview
+    AnthropicModels["CLAUDE_OPUS_5_5"] = "claude-opus-5-5";
+    AnthropicModels["CLAUDE_FABLE_5_1"] = "claude-fable-5-1";
     // Claude 4.7 / 4.8 Series
     AnthropicModels["CLAUDE_OPUS_4_8"] = "claude-opus-4-8";
     AnthropicModels["CLAUDE_OPUS_4_7"] = "claude-opus-4-7";
@@ -868,6 +889,14 @@ export var CohereModels;
  */
 export var VoyageModels;
 (function (VoyageModels) {
+    /** Voyage 4 — latest general-purpose flagship */
+    VoyageModels["VOYAGE_4"] = "voyage-4";
+    /** Voyage 4 Large — largest / highest quality */
+    VoyageModels["VOYAGE_4_LARGE"] = "voyage-4-large";
+    /** Voyage 4 Lite — smaller / cheaper */
+    VoyageModels["VOYAGE_4_LITE"] = "voyage-4-lite";
+    /** Voyage Code 4 — code-tuned */
+    VoyageModels["VOYAGE_CODE_4"] = "voyage-code-4";
     /** Voyage 3.5 — latest general-purpose (default) */
     VoyageModels["VOYAGE_3_5"] = "voyage-3.5";
     /** Voyage 3.5 Lite — smaller / cheaper */
@@ -1290,6 +1319,7 @@ export var UpstageModels;
 })(UpstageModels || (UpstageModels = {}));
 export var XaiModels;
 (function (XaiModels) {
+    XaiModels["GROK_4_7"] = "grok-4.7";
     XaiModels["GROK_4_6"] = "grok-4.6";
     XaiModels["GROK_4_5"] = "grok-4.5";
     XaiModels["GROK_4_3"] = "grok-4.3";

package/dist/constants/tokens.d.ts

@@ -71,6 +71,9 @@ export declare const PROVIDER_TOKEN_LIMITS: {
     };
     /** OpenAI model limits */
     readonly OPENAI: {
+        readonly "gpt-6-astra": 128000;
+        readonly "gpt-6-sol": 128000;
+        readonly "gpt-6-luna": 128000;
         readonly "gpt-5.4": 128000;
         readonly "gpt-5.4-mini": 128000;
         readonly "gpt-5.4-nano": 128000;

package/dist/constants/tokens.js

@@ -74,6 +74,9 @@ export const PROVIDER_TOKEN_LIMITS = {
     },
     /** OpenAI model limits */
     OPENAI: {
+        "gpt-6-astra": 128_000,
+        "gpt-6-sol": 128_000,
+        "gpt-6-luna": 128_000,
         "gpt-5.4": 128_000,
         "gpt-5.4-mini": 128_000,
         "gpt-5.4-nano": 128_000,

package/dist/evaluation/EvaluatorFactory.js

@@ -57,7 +57,7 @@ export class EvaluatorFactory extends BaseFactory {
             const mergedConfig = {
                 threshold: 7,
                 evaluationStrategy: "ragas",
-                evaluationModel: process.env.NEUROLINK_RAGAS_EVALUATION_MODEL || "gemini-1.5-flash",
+                evaluationModel: process.env.NEUROLINK_RAGAS_EVALUATION_MODEL || "gemini-2.5-flash",
                 provider: process.env.NEUROLINK_RAGAS_EVALUATION_PROVIDER || "vertex",
                 ...config,
             };
@@ -95,7 +95,7 @@ export class EvaluatorFactory extends BaseFactory {
             const mergedConfig = {
                 threshold: 5,
                 evaluationStrategy: "ragas",
-                evaluationModel: process.env.NEUROLINK_RAGAS_EVALUATION_MODEL || "gemini-1.5-flash",
+                evaluationModel: process.env.NEUROLINK_RAGAS_EVALUATION_MODEL || "gemini-2.5-flash",
                 provider: process.env.NEUROLINK_RAGAS_EVALUATION_PROVIDER || "vertex",
                 ...config,
             };
@@ -114,7 +114,7 @@ export class EvaluatorFactory extends BaseFactory {
             const mergedConfig = {
                 threshold: 6,
                 evaluationStrategy: "ragas",
-                evaluationModel: "gemini-1.5-flash",
+                evaluationModel: "gemini-2.5-flash",
                 provider: "vertex",
                 ...config,
             };

package/dist/evaluation/EvaluatorRegistry.js

@@ -73,7 +73,7 @@ export class EvaluatorRegistry extends BaseRegistry {
             name: "RAGAS Evaluator",
             description: "RAGAS-style LLM-as-judge evaluation with relevance, accuracy, and completeness metrics",
             requiresLLM: true,
-            defaultModel: "gemini-1.5-flash",
+            defaultModel: "gemini-2.5-flash",
             defaultProvider: "vertex",
             version: "1.0.0",
             features: [

package/dist/evaluation/ragasEvaluator.js

@@ -18,7 +18,7 @@ export class RAGASEvaluator {
         this.evaluationModel =
             evaluationModel ||
                 process.env.NEUROLINK_RAGAS_EVALUATION_MODEL ||
-                "gemini-1.5-flash";
+                "gemini-2.5-flash";
         this.providerName =
             providerName ||
                 process.env.NEUROLINK_RAGAS_EVALUATION_PROVIDER ||

package/dist/mcp/externalServerManager.js

@@ -7,7 +7,6 @@
  * - Tool discovery and registration
  */
 import { EventEmitter } from "events";
-import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
 import { mcpLogger } from "../utils/logger.js";
 import { MCPClientFactory } from "./mcpClientFactory.js";
 import { ToolDiscoveryService } from "./toolDiscoveryService.js";
@@ -1063,6 +1062,7 @@ export class ExternalServerManager extends EventEmitter {
             instance.client = clientResult.client;
             instance.transportInstance = clientResult.transport;
             instance.process = clientResult.process || null;
+            const { StdioClientTransport } = await import("@modelcontextprotocol/sdk/client/stdio.js");
             instance.pid =
                 clientResult.transport instanceof StdioClientTransport
                     ? (clientResult.transport.pid ?? undefined)

package/dist/mcp/mcpClientFactory.d.ts

@@ -4,7 +4,7 @@
  * Supports stdio, SSE, WebSocket, and HTTP transports
  * Enhanced with retry, rate limiting, and OAuth 2.1 support
  */
-import { Client } from "@modelcontextprotocol/sdk/client/index.js";
+import type { Client } from "@modelcontextprotocol/sdk/client/index.js";
 import type { Transport } from "@modelcontextprotocol/sdk/shared/transport.js";
 import type { ClientCapabilities } from "@modelcontextprotocol/sdk/types.js";
 import type { ChildProcess } from "child_process";

package/dist/mcp/mcpClientFactory.js

@@ -4,11 +4,6 @@
  * Supports stdio, SSE, WebSocket, and HTTP transports
  * Enhanced with retry, rate limiting, and OAuth 2.1 support
  */
-import { Client } from "@modelcontextprotocol/sdk/client/index.js";
-import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
-import { SSEClientTransport } from "@modelcontextprotocol/sdk/client/sse.js";
-import { WebSocketClientTransport } from "@modelcontextprotocol/sdk/client/websocket.js";
-import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";
 import { mcpLogger } from "../utils/logger.js";
 import { globalCircuitBreakerManager } from "./mcpCircuitBreaker.js";
 import { CircuitBreakerOpenError } from "../types/index.js";
@@ -209,6 +204,7 @@ export class MCPClientFactory {
         const transport = transportResult.transport;
         try {
             // Create client
+            const { Client } = await import("@modelcontextprotocol/sdk/client/index.js");
             const client = new Client(this.NEUROLINK_IMPLEMENTATION, {
                 capabilities: this.DEFAULT_CAPABILITIES,
             });
@@ -296,6 +292,7 @@ export class MCPClientFactory {
         if (!config.command) {
             throw new Error(`Command is required for stdio transport`);
         }
+        const { StdioClientTransport } = await import("@modelcontextprotocol/sdk/client/stdio.js");
         const transport = new StdioClientTransport({
             command: config.command,
             args: config.args || [],
@@ -333,6 +330,7 @@ export class MCPClientFactory {
         });
         try {
             const url = new URL(config.url);
+            const { SSEClientTransport } = await import("@modelcontextprotocol/sdk/client/sse.js");
             const transport = new SSEClientTransport(url);
             return { transport };
         }
@@ -352,6 +350,7 @@ export class MCPClientFactory {
         });
         try {
             const url = new URL(config.url);
+            const { WebSocketClientTransport } = await import("@modelcontextprotocol/sdk/client/websocket.js");
             const transport = new WebSocketClientTransport(url);
             return { transport };
         }
@@ -402,6 +401,7 @@ export class MCPClientFactory {
             const requestInit = {
                 headers: Object.keys(headers).length > 0 ? headers : undefined,
             };
+            const { StreamableHTTPClientTransport } = await import("@modelcontextprotocol/sdk/client/streamableHttp.js");
             const transport = new StreamableHTTPClientTransport(url, {
                 requestInit,
                 fetch: fetchWithEnhancements,

package/dist/middleware/utils/guardrailsUtils.js

@@ -113,7 +113,7 @@ params, config) {
  */
 export async function performPrecallEvaluation(config, userInput) {
     try {
-        const provider = await AIProviderFactory.createProvider(config.provider || "google-ai", config.evaluationModel || "gemini-1.5-flash");
+        const provider = await AIProviderFactory.createProvider(config.provider || "google-ai", config.evaluationModel || "gemini-2.5-flash");
         const evaluationPrompt = config.evaluationPrompt || DEFAULT_EVALUATION_PROMPT;
         const prompt = evaluationPrompt.replace("{USER_INPUT}", userInput);
         const result = await provider.generate({

package/dist/models/manifests/anthropic.js

@@ -23,11 +23,43 @@ export const anthropicManifest = {
         },
     ],
     models: {
+        "claude-opus-5-5": {
+            aliases: ["opus-5.5"],
+            displayName: "Claude Opus 5.5",
+            contextWindow: 1_000_000,
+            maxOutputTokens: 128_000,
+            // No pricingPerMTok DELIBERATELY: PRICING.anthropic already carries
+            // real rates for this id (input 4.0/output 20.0 per MTok) — see the
+            // claude-sonnet-5 comment below for why this defers to that table.
+            vision: true,
+            functionCalling: true,
+            reasoning: true,
+            jsonMode: true,
+            samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /opus[-_.]?5(?![0-9])/i
+        },
+        "claude-fable-5-1": {
+            aliases: ["fable-5.1"],
+            displayName: "Claude Fable 5.1",
+            contextWindow: 1_000_000,
+            maxOutputTokens: 128_000,
+            // No pricingPerMTok DELIBERATELY: PRICING.anthropic already carries
+            // real rates for this id (input 10.0/output 50.0 per MTok) — see the
+            // claude-sonnet-5 comment below for why this defers to that table.
+            vision: true,
+            functionCalling: true,
+            reasoning: true,
+            jsonMode: true,
+            samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /fable/i
+        },
         "claude-sonnet-5": {
             aliases: ["sonnet-5", "claude-sonnet"],
             displayName: "Claude Sonnet 5",
             contextWindow: 1_000_000,
-            maxOutputTokens: 64_000,
+            // Was 64_000; corrected to 128_000 to match the official spec
+            // (platform.claude.com/docs/en/about-claude/models/overview) and the
+            // matching fix in getClaudeMaxOutputTokens (tokenLimits.ts), which this
+            // manifest's own header comment names as the authoritative source.
+            maxOutputTokens: 128_000,
             // No pricingPerMTok DELIBERATELY: PRICING.anthropic carries real
             // rates for this id, and findRates() is manifest-first with a legacy
             // fallback — omitting the rate here defers to that table instead of

package/dist/neurolink.js

@@ -2804,11 +2804,15 @@ Current user's request: ${currentInput}`;
         try {
             const langfuseConfig = this.observabilityConfig?.langfuse;
             if (langfuseConfig?.enabled) {
-                void initializeOpenTelemetry(langfuseConfig).catch((err) => {
+                try {
+                    await initializeOpenTelemetry(langfuseConfig);
+                }
+                catch (err) {
                     logger.error("[NeuroLink] OpenTelemetry initialization failed", {
                         error: err instanceof Error ? err.message : String(err),
                     });
-                });
+                    return;
+                }
                 logger.debug("[NeuroLink] Langfuse observability initialized via public method");
             }
             else {

package/dist/providers/catalog/xai.json

@@ -9,10 +9,24 @@
     },
     "models": {
         "default": "grok-4.6",
-        "fallbacks": ["grok-4.6", "grok-4.5", "grok-4.3", "grok-3", "grok-3-mini"],
+        "fallbacks": [
+            "grok-4.6",
+            "grok-4.5",
+            "grok-4.3",
+            "grok-3",
+            "grok-3-mini",
+            "grok-4.7"
+        ],
         "defaultContextWindow": 131072,
         "defaultMaxOutputTokens": 4096,
         "catalog": {
+            "grok-4.7": {
+                "enumMember": "GROK_4_7",
+                "contextWindow": 500000,
+                "vision": true,
+                "status": "production",
+                "description": "Grok 4.7 — xAI's newest flagship as of 2026-09-27 (text + image input); not yet the SDK default, see docs.x.ai/developers/grok-4-7"
+            },
             "grok-4.6": {
                 "enumMember": "GROK_4_6",
                 "contextWindow": 131072,
@@ -71,7 +85,7 @@
             }
         },
         "visionModel": "grok-4.6",
-        "topModels": ["grok-4.6", "grok-4.5", "grok-4.3", "grok-3"]
+        "topModels": ["grok-4.6", "grok-4.5", "grok-4.3", "grok-3", "grok-4.7"]
     },
     "capabilities": {
         "text": true,
@@ -114,7 +128,7 @@
         },
         "liveMatrix": {
             "date": "2026-08-30",
-            "result": "grok-2-latest, grok-2-vision-latest and grok-beta are rejected as invalid models. The grok-4.x line replaces them: 4.3/4.5/4.6 all answer, and all three read a 64x64 PNG correctly, so vision moved there from the retired grok-2-vision-latest. An earlier probe with an 8x8 PNG reported no vision — that was xAI answering 'Invalid PNG image', a complaint about the file rather than the capability. grok-3 and grok-3-mini are kept because they still serve, even though xAI's /models roster omits them: roster absence is not evidence of death here."
+            "result": "grok-2-latest, grok-2-vision-latest and grok-beta are rejected as invalid models. The grok-4.x line replaces them: 4.3/4.5/4.6 all answer, and all three read a 64x64 PNG correctly, so vision moved there from the retired grok-2-vision-latest. An earlier probe with an 8x8 PNG reported no vision — that was xAI answering 'Invalid PNG image', a complaint about the file rather than the capability. grok-3 and grok-3-mini are kept because they still serve, even though xAI's /models roster omits them: roster absence is not evidence of death here. grok-4.7 added 2026-09-27 from docs.x.ai/developers/grok-4-7 and docs.x.ai/developers/pricing (official docs review only — no xAI API key was available in this environment, so unlike the rest of this file it has not been roster/chat-probed live; context 500,000 tokens and text+image input are as published, and it is deliberately left with no pricingPerMTok, matching grok-4.6/4.5/4.3, because its real rate is tiered by prompt length ($2/$0.50/$6 per 1M below 200k tokens, $4/$1/$12 at or above) and the schema's pricingPerMTok field can only hold one flat rate)."
         },
         "addedInPR": "https://github.com/juspay/neurolink/pull/1587"
     }

package/dist/providers/googleVertex/client.js

@@ -6480,8 +6480,6 @@ export class GoogleVertexProvider extends BaseProvider {
                 "gemini-2.5-flash-lite",
                 "gemini-2.0-flash-001",
                 "gemini-2.0-flash-lite",
-                "gemini-1.5-pro",
-                "gemini-1.5-flash",
             ],
             claude: [
                 "claude-sonnet-4-5@20250929",

package/dist/providers/litellm/client.js

@@ -13,7 +13,7 @@ import { createTimeoutController } from "../../utils/timeout.js";
 import { stripTrailingSlash } from "../openaiChatCompletionsClient.js";
 import { OpenAIChatCompletionsProvider } from "../openaiChatCompletionsBase.js";
 const streamTracer = trace.getTracer("neurolink.provider.litellm");
-const FALLBACK_LITELLM_MODEL = "openai/gpt-4o-mini";
+const FALLBACK_LITELLM_MODEL = "openai/gpt-5.4-mini";
 const getLiteLLMConfig = () => ({
     baseURL: process.env.LITELLM_BASE_URL || "http://localhost:4000",
     apiKey: process.env.LITELLM_API_KEY || "sk-anything",
@@ -243,7 +243,7 @@ export class LiteLLMProvider extends OpenAIChatCompletionsProvider {
             .map((m) => m.trim())
             .filter((m) => m.length > 0) || [
             "openai/gpt-4o",
-            "anthropic/claude-3-haiku",
+            "anthropic/claude-haiku-4-5-20251001",
             "meta-llama/llama-3.1-8b-instruct",
             "google/gemini-2.5-flash",
         ]);

package/dist/providers/openAI/client.js

@@ -44,7 +44,7 @@ const resolveOpenAIBaseURL = (credentialBaseURL, envBaseURL) => {
     return resolved;
 };
 const getOpenAIApiKey = () => validateApiKey(createOpenAIConfig());
-const getOpenAIModel = () => getProviderModel("OPENAI_MODEL", "gpt-4o");
+const getOpenAIModel = () => getProviderModel("OPENAI_MODEL", "gpt-5.4");
 const streamTracer = trace.getTracer("neurolink.provider.openai");
 /**
  * OpenAI Provider — direct HTTP, no AI SDK.

package/dist/providers/openRouter/client.js

@@ -278,7 +278,7 @@ export class OpenRouterProvider extends OpenAIChatCompletionsProvider {
             "openai/gpt-4-turbo",
             // Google models
             "google/gemini-2.0-flash",
-            "google/gemini-1.5-pro",
+            "google/gemini-2.5-pro",
             // Meta Llama models
             "meta-llama/llama-3.1-70b-instruct",
             "meta-llama/llama-3.1-8b-instruct",

package/dist/proxy/vertexAnthropicFallback.d.ts

@@ -12,6 +12,7 @@
  * `tool_use`, `tool_result`, `cache_control` and thinking blocks all survive
  * here because nothing reshapes them.
  */
+import type { GoogleAuth } from "google-auth-library";
 import type { VertexAccessTokenProvider, UsageContext, VertexPassthroughTerminal, VertexAnthropicPassthroughRequest, VertexAnthropicPassthroughTarget } from "../types/index.js";
 /**
  * Translate `output_config` for Vertex.
@@ -40,6 +41,13 @@ export declare function buildVertexAnthropicPayload(body: Readonly<Record<string
  * caller's `stream` flag selects the verb rather than changing the body.
  */
 export declare function buildVertexAnthropicUrl(target: VertexAnthropicPassthroughTarget): string;
+/** Exported only for continuous-test-suite-vertex-anthropic-fallback.ts,
+ * which proves concurrent callers share one instance. Not part of the
+ * package's public surface — this module isn't re-exported from index.ts. */
+export declare function auth(): Promise<GoogleAuth>;
+/** Test-only: isolates the concurrency proof from cache state any earlier
+ * test in the same process may have left behind. */
+export declare function resetAuthCacheForTests(): void;
 /**
  * Replace the Google credential lookup, for tests only.
  *

package/dist/proxy/vertexAnthropicFallback.js

@@ -12,7 +12,6 @@
  * `tool_use`, `tool_result`, `cache_control` and thinking blocks all survive
  * here because nothing reshapes them.
  */
-import { GoogleAuth } from "google-auth-library";
 import { logger } from "../utils/logger.js";
 import { isTransientNetworkError } from "./proxyFetch.js";
 /** Vertex pins the Anthropic API version in the body, not a header. */
@@ -379,16 +378,41 @@ export function buildVertexAnthropicUrl(target) {
     return (`https://${host}/v1/projects/${target.projectId}/locations/` +
         `${target.location}/publishers/anthropic/models/${target.model}:${verb}`);
 }
-let cachedAuth;
-function auth() {
-    cachedAuth ??= new GoogleAuth({ scopes: [CLOUD_PLATFORM_SCOPE] });
-    return cachedAuth;
+// google-auth-library is only needed once a Vertex fallback request actually
+// needs a Google access token, so it's loaded with `await import(...)` instead
+// of a static import. The module promise is cached at module scope, and the
+// resulting client instance is cached separately, exactly as before.
+let googleAuthModulePromise;
+function loadGoogleAuth() {
+    if (!googleAuthModulePromise) {
+        googleAuthModulePromise = import("google-auth-library");
+    }
+    return googleAuthModulePromise;
+}
+// Cache the PROMISE, not the resolved client: the guard below runs to
+// completion before any await, so two calls entering before the first
+// resolves both see the same in-flight promise instead of each racing to
+// construct — and return — its own GoogleAuth instance.
+let cachedAuthPromise;
+/** Exported only for continuous-test-suite-vertex-anthropic-fallback.ts,
+ * which proves concurrent callers share one instance. Not part of the
+ * package's public surface — this module isn't re-exported from index.ts. */
+export function auth() {
+    if (!cachedAuthPromise) {
+        cachedAuthPromise = loadGoogleAuth().then(({ GoogleAuth }) => new GoogleAuth({ scopes: [CLOUD_PLATFORM_SCOPE] }));
+    }
+    return cachedAuthPromise;
+}
+/** Test-only: isolates the concurrency proof from cache state any earlier
+ * test in the same process may have left behind. */
+export function resetAuthCacheForTests() {
+    cachedAuthPromise = undefined;
 }
 let accessTokenProviderForTests;
 async function vertexAccessToken() {
     return accessTokenProviderForTests
         ? accessTokenProviderForTests()
-        : auth().getAccessToken();
+        : (await auth()).getAccessToken();
 }
 /**
  * Replace the Google credential lookup, for tests only.

package/dist/services/server/ai/observability/instrumentation.d.ts

@@ -7,9 +7,9 @@
  * Flow: Vercel AI SDK → OpenTelemetry Spans → LangfuseSpanProcessor → Langfuse Platform
  */
 import { trace, type Span as ApiSpan } from "@opentelemetry/api";
-import { LoggerProvider } from "@opentelemetry/sdk-logs";
-import { type SpanProcessor } from "@opentelemetry/sdk-trace-base";
-import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
+import type { LoggerProvider } from "@opentelemetry/sdk-logs";
+import type { SpanProcessor } from "@opentelemetry/sdk-trace-base";
+import type { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
 import type { LangfuseConfig, LangfuseContext } from "../../../../types/index.js";
 /**
  * True when a span is an internal NeuroLink wrapper that should NOT be sent to
@@ -38,18 +38,11 @@ export declare function langfuseShouldExportSpan({ otelSpan, }: {
     };
 }): boolean;
 /**
- * Initialize OpenTelemetry with Langfuse span processor
+ * Initialize OpenTelemetry once and let concurrent callers share the work.
  *
- * This connects Vercel AI SDK's experimental_telemetry to Langfuse by:
- * 1. Creating LangfuseSpanProcessor with Langfuse credentials
- * 2. Creating a NodeTracerProvider with service metadata and span processor
- * 3. Registering the provider globally for AI SDK to use
- *
- * NEW: If useExternalTracerProvider is true or autoDetectExternalProvider detects
- * an existing provider, steps 2 and 3 are skipped. The span processors are still
- * created and can be retrieved via getSpanProcessors().
- *
- * @param config - Langfuse configuration passed from parent application
+ * NeuroLink constructors intentionally start this asynchronously. Keeping the
+ * in-flight promise here lets another constructor join it and lets an immediate
+ * flush or shutdown wait for the same initialization before touching providers.
  */
 export declare function initializeOpenTelemetry(config: LangfuseConfig): Promise<void>;
 /**