@juspay/neurolink 12.9.0 → 12.9.2

package/dist/providers/catalog/schema.d.ts

@@ -47,8 +47,8 @@ export declare const providerCatalogJsonSchema: z.ZodObject<{
             vision: z.ZodBoolean;
             status: z.ZodEnum<{
                 production: "production";
-                preview: "preview";
                 retired: "retired";
+                preview: "preview";
             }>;
             description: z.ZodString;
             enumMember: z.ZodOptional<z.ZodString>;
@@ -81,6 +81,7 @@ export declare const providerCatalogJsonSchema: z.ZodObject<{
     }, z.core.$strict>>;
     quirks: z.ZodOptional<z.ZodObject<{
         timeoutErrorClass: z.ZodOptional<z.ZodLiteral<"provider">>;
+        messageContentFormat: z.ZodOptional<z.ZodLiteral<"string">>;
         registryDefaultIgnoresModelEnvVar: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strict>>;
     setup: z.ZodObject<{

package/dist/providers/catalog/schema.js

@@ -136,6 +136,7 @@ const catalogErrorRuleJsonSchema = z
 });
 const catalogQuirksSchema = z.strictObject({
     timeoutErrorClass: z.literal("provider").optional(),
+    messageContentFormat: z.literal("string").optional(),
     registryDefaultIgnoresModelEnvVar: z.boolean().optional(),
 });
 const catalogBillingPolicySchema = z.enum([

package/dist/providers/catalog/xai.json

@@ -8,23 +8,38 @@
         "baseURL": "https://api.x.ai/v1"
     },
     "models": {
-        "default": "grok-3",
-        "fallbacks": [
-            "grok-3",
-            "grok-3-mini",
-            "grok-2-latest",
-            "grok-2-vision-latest",
-            "grok-beta"
-        ],
+        "default": "grok-4.6",
+        "fallbacks": ["grok-4.6", "grok-4.5", "grok-4.3", "grok-3", "grok-3-mini"],
         "defaultContextWindow": 131072,
         "defaultMaxOutputTokens": 4096,
         "catalog": {
+            "grok-4.6": {
+                "enumMember": "GROK_4_6",
+                "contextWindow": 131072,
+                "vision": true,
+                "status": "production",
+                "description": "Recommended - Grok 4.6, current flagship; vision-capable"
+            },
+            "grok-4.5": {
+                "enumMember": "GROK_4_5",
+                "contextWindow": 131072,
+                "vision": true,
+                "status": "production",
+                "description": "Grok 4.5 — vision-capable"
+            },
+            "grok-4.3": {
+                "enumMember": "GROK_4_3",
+                "contextWindow": 131072,
+                "vision": true,
+                "status": "production",
+                "description": "Grok 4.3 — vision-capable"
+            },
             "grok-3": {
                 "contextWindow": 131072,
                 "pricingPerMTok": { "input": 3.0, "output": 15.0 },
                 "vision": false,
                 "status": "production",
-                "description": "Recommended - Latest flagship Grok"
+                "description": "Grok 3 — still served, though absent from the /models roster"
             },
             "grok-3-mini": {
                 "contextWindow": 131072,
@@ -37,31 +52,26 @@
                 "contextWindow": 131072,
                 "pricingPerMTok": { "input": 2.0, "output": 10.0 },
                 "vision": false,
-                "status": "production",
-                "description": "Previous flagship"
+                "status": "retired",
+                "description": "Retired 2026-08 — xAI rejects it as an invalid model"
             },
             "grok-2-vision-latest": {
                 "contextWindow": 32768,
                 "pricingPerMTok": { "input": 2.0, "output": 10.0 },
-                "vision": true,
-                "status": "production",
-                "description": "Multimodal (text + images)"
+                "vision": false,
+                "status": "retired",
+                "description": "Retired 2026-08 — xAI rejects it as an invalid model; vision moved to the grok-4.x line"
             },
             "grok-beta": {
                 "contextWindow": 131072,
                 "pricingPerMTok": { "input": 5.0, "output": 15.0 },
                 "vision": false,
-                "status": "preview",
-                "description": "Pre-release / experimental"
+                "status": "retired",
+                "description": "Retired 2026-08 — xAI rejects it as an invalid model"
             }
         },
-        "topModels": [
-            "grok-3",
-            "grok-3-mini",
-            "grok-2-vision-latest",
-            "grok-2-latest",
-            "grok-beta"
-        ]
+        "visionModel": "grok-4.6",
+        "topModels": ["grok-4.6", "grok-4.5", "grok-4.3", "grok-3"]
     },
     "capabilities": {
         "text": true,
@@ -99,10 +109,13 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-08-28",
-            "method": "transcribed from pre-migration TS catalog"
+            "date": "2026-08-30",
+            "method": "GET /models against the live account, cross-checked with a real chat call per id"
+        },
+        "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."
         },
-        "liveMatrix": null,
         "addedInPR": "https://github.com/juspay/neurolink/pull/1587"
     }
 }

package/dist/providers/configuredOpenAICompat.d.ts

@@ -1,5 +1,5 @@
 import type { AIProviderName } from "../constants/enums.js";
-import type { OpenAICompatCatalogEntry, OpenAICompatCredentials } from "../types/index.js";
+import type { OpenAICompatCatalogEntry, OpenAICompatChatRequest, OpenAICompatCredentials } from "../types/index.js";
 import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
 /**
  * Generic OpenAI-compatible provider driven entirely by an
@@ -12,6 +12,22 @@ import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
  * adjustBodyAfter400, getChatCompletionsURL, getAuthHeaders,
  * suppressResponseFormatWithTools, ...) it does NOT belong in the catalog —
  * write a dedicated subclass instead (see deepseek.ts, azureOpenai.ts).
+ *
+ * The exception is a WIRE DIALECT: a vendor that speaks OpenAI for ordinary
+ * chat but encodes one part of the request differently. Expressing that as
+ * data (the catalog's `messageContentFormat`) keeps the provider a
+ * one-JSON-file entry instead of promoting it to a hand-written subclass
+ * over a single incompatibility.
+ *
+ * Today that is Cloudflare Workers AI, whose OpenAI-compatible endpoint
+ * accepts `messages[].content` only as a plain string — never the
+ * content-parts array OpenAI allows, and never the `null` that OpenAI uses
+ * on an assistant message carrying tool_calls. Its schema rejects both with
+ * HTTP 400 ("Type mismatch of '/messages/N/content'"). That bites precisely
+ * on the second turn of a tool call, so single-turn chat looks healthy while
+ * every tool round-trip fails. Normalizing content to a string below is the
+ * whole fix: tools, tool_choice, the `tool` role and `tool_calls` are all
+ * accepted as-is.
  */
 export declare class ConfiguredOpenAICompatProvider extends OpenAIChatCompletionsProvider {
     private readonly entry;
@@ -20,5 +36,6 @@ export declare class ConfiguredOpenAICompatProvider extends OpenAIChatCompletion
     protected getDefaultModel(): string;
     protected getFallbackModelName(): string;
     protected getFallbackModels(): string[];
+    protected adjustRequestBody(body: OpenAICompatChatRequest, modelId: string): OpenAICompatChatRequest;
     protected formatProviderError(error: unknown): Error;
 }

package/dist/providers/configuredOpenAICompat.js

@@ -4,6 +4,25 @@ import { getProviderModel, resolveOpenAICompatConfig, } from "../utils/providerC
 import { classifyProviderError } from "../utils/errorClassifier.js";
 import { TimeoutError } from "../utils/timeout.js";
 import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
+/**
+ * Collapse OpenAI's `content` union down to the plain string that
+ * string-only vendors accept. Image parts carry no string representation and
+ * are dropped — a provider that cannot accept a content array cannot accept
+ * inline images either.
+ */
+const flattenMessageContent = (content) => {
+    if (typeof content === "string") {
+        return content;
+    }
+    // An assistant message with tool_calls legitimately has null content;
+    // the empty string is its string-only equivalent.
+    if (content === null || content === undefined) {
+        return "";
+    }
+    return content
+        .map((part) => (part.type === "text" ? part.text : ""))
+        .join("");
+};
 /**
  * Generic OpenAI-compatible provider driven entirely by an
  * OpenAICompatCatalogEntry. Replaces a hand-written subclass for any
@@ -15,6 +34,22 @@ import { OpenAIChatCompletionsProvider } from "./openaiChatCompletionsBase.js";
  * adjustBodyAfter400, getChatCompletionsURL, getAuthHeaders,
  * suppressResponseFormatWithTools, ...) it does NOT belong in the catalog —
  * write a dedicated subclass instead (see deepseek.ts, azureOpenai.ts).
+ *
+ * The exception is a WIRE DIALECT: a vendor that speaks OpenAI for ordinary
+ * chat but encodes one part of the request differently. Expressing that as
+ * data (the catalog's `messageContentFormat`) keeps the provider a
+ * one-JSON-file entry instead of promoting it to a hand-written subclass
+ * over a single incompatibility.
+ *
+ * Today that is Cloudflare Workers AI, whose OpenAI-compatible endpoint
+ * accepts `messages[].content` only as a plain string — never the
+ * content-parts array OpenAI allows, and never the `null` that OpenAI uses
+ * on an assistant message carrying tool_calls. Its schema rejects both with
+ * HTTP 400 ("Type mismatch of '/messages/N/content'"). That bites precisely
+ * on the second turn of a tool call, so single-turn chat looks healthy while
+ * every tool round-trip fails. Normalizing content to a string below is the
+ * whole fix: tools, tool_choice, the `tool` role and `tool_calls` are all
+ * accepted as-is.
  */
 export class ConfiguredOpenAICompatProvider extends OpenAIChatCompletionsProvider {
     entry;
@@ -51,6 +86,19 @@ export class ConfiguredOpenAICompatProvider extends OpenAIChatCompletionsProvide
     getFallbackModels() {
         return this.entry.fallbackModels;
     }
+    adjustRequestBody(body, modelId) {
+        const adjusted = super.adjustRequestBody(body, modelId);
+        if (this.entry.messageContentFormat !== "string") {
+            return adjusted;
+        }
+        return {
+            ...adjusted,
+            messages: adjusted.messages.map((message) => ({
+                ...message,
+                content: flattenMessageContent(message.content),
+            })),
+        };
+    }
     formatProviderError(error) {
         // classifyProviderError hard-codes TimeoutError -> NetworkError ahead
         // of any rule table and does not allow a per-provider override. An

package/dist/providers/openaiChatCompletionsBase.d.ts

@@ -45,6 +45,21 @@ export declare abstract class OpenAIChatCompletionsProvider extends BaseProvider
      * Hardcoded model names returned from `getAvailableModels()` when the
      * remote `/models` endpoint can't be reached. Default empty.
      */
+    /**
+     * Feed the catalog's `fallbacks` to BaseProvider's invalid-model retry, so
+     * a default the vendor has retired degrades to the next live model in the
+     * entry instead of failing the call outright.
+     */
+    protected getModelFallbacks(): string[];
+    /**
+     * `resolvedModel` memoizes the first id this provider resolved, and
+     * getAISDKModel() builds its wire model from that memo rather than from
+     * `modelName`. Leaving it stale here silently defeats the invalid-model
+     * fallback: modelName advances to the next candidate while every request
+     * still carries the retired id, so each retry fails for the same reason
+     * the first attempt did.
+     */
+    protected refreshHandlersForModel(model: string): void;
     protected getFallbackModels(): string[];
     /**
      * Hook to mutate the `buildBody` options before the wire body is

package/dist/providers/openaiChatCompletionsBase.js

@@ -68,6 +68,26 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
      * Hardcoded model names returned from `getAvailableModels()` when the
      * remote `/models` endpoint can't be reached. Default empty.
      */
+    /**
+     * Feed the catalog's `fallbacks` to BaseProvider's invalid-model retry, so
+     * a default the vendor has retired degrades to the next live model in the
+     * entry instead of failing the call outright.
+     */
+    getModelFallbacks() {
+        return this.getFallbackModels();
+    }
+    /**
+     * `resolvedModel` memoizes the first id this provider resolved, and
+     * getAISDKModel() builds its wire model from that memo rather than from
+     * `modelName`. Leaving it stale here silently defeats the invalid-model
+     * fallback: modelName advances to the next candidate while every request
+     * still carries the retired id, so each retry fails for the same reason
+     * the first attempt did.
+     */
+    refreshHandlersForModel(model) {
+        this.resolvedModel = model;
+        super.refreshHandlersForModel(model);
+    }
     getFallbackModels() {
         return [];
     }

package/dist/proxy/proxyActivity.d.ts

@@ -1,4 +1,6 @@
 import type { ProxyActivitySnapshot, ProxyResponseTrackingObserver } from "../types/index.js";
+export declare function registerProxyResponseObserver(metadata: object, observer: ProxyResponseTrackingObserver): void;
+export declare function takeProxyResponseObservers(metadata: object): ProxyResponseTrackingObserver[];
 /** Track one client-facing proxy request until its response body settles. */
 export declare function beginProxyRequest(): () => void;
 export declare function getProxyActivitySnapshot(): ProxyActivitySnapshot;

package/dist/proxy/proxyActivity.js

@@ -3,6 +3,24 @@ import { logger } from "../utils/logger.js";
 const PROXY_RESPONSE_CANCEL_TIMEOUT_MS = 1_000;
 let activeRequests = 0;
 let lastActivityAtMs = null;
+// Route handlers can attach terminal observers to their request context without
+// wrapping the response body a second time. The HTTP runtime drains every
+// response through one tracker, which fans these observers out at the point
+// where bytes actually leave the proxy.
+const responseObserversByMetadata = new WeakMap();
+export function registerProxyResponseObserver(metadata, observer) {
+    const existing = responseObserversByMetadata.get(metadata);
+    if (existing) {
+        existing.push(observer);
+        return;
+    }
+    responseObserversByMetadata.set(metadata, [observer]);
+}
+export function takeProxyResponseObservers(metadata) {
+    const observers = responseObserversByMetadata.get(metadata) ?? [];
+    responseObserversByMetadata.delete(metadata);
+    return observers;
+}
 function touchActivity() {
     lastActivityAtMs = Date.now();
 }

package/dist/server/routes/codexProxyRoutes.js

@@ -24,7 +24,7 @@ import { loadAccountQuotas, saveAccountQuota, } from "../../proxy/accountQuota.j
 import { createCodexUsageTap } from "../../proxy/codexUsage.js";
 import { CODEX_ACCOUNT_PREFIX, parseCodexRateLimitHeaders, } from "../../proxy/codexAccountUsage.js";
 import { buildClientAttribution } from "../../proxy/clientAttribution.js";
-import { trackProxyResponse } from "../../proxy/proxyActivity.js";
+import { registerProxyResponseObserver } from "../../proxy/proxyActivity.js";
 import { logRequest, logRequestAttempt } from "../../proxy/requestLogger.js";
 import { parseRetryAfterMs } from "../../proxy/routingPolicy.js";
 import { recordAttempt, recordAttemptError, recordFinalError, recordFinalSuccess, } from "../../proxy/usageStats.js";
@@ -459,7 +459,7 @@ export async function handleCodexResponsesRequest(ctx) {
                     status: upstream.status,
                     headers,
                 });
-                return trackProxyResponse(relay, () => undefined, {
+                registerProxyResponseObserver(ctx.metadata, {
                     onTerminal: ({ outcome }) => {
                         void usageSeen
                             .then((usage) => {
@@ -491,6 +491,7 @@ export async function handleCodexResponsesRequest(ctx) {
                             .catch(() => undefined);
                     },
                 });
+                return relay;
             }
             const errText = await upstream.text().catch(() => "");
             // 401/403 → try a forced token refresh once, then rotate.

package/dist/types/providerCatalog.d.ts

@@ -44,6 +44,12 @@ export type CatalogErrorRuleJson = {
 };
 export type CatalogQuirks = {
     timeoutErrorClass?: "provider";
+    /** Vendor speaks OpenAI for chat but restricts how message content is
+     *  encoded. "string": `messages[].content` must be a plain string —
+     *  the content-parts array and the `null` OpenAI uses on an assistant
+     *  message with tool_calls are both rejected with HTTP 400. Normalized by
+     *  ConfiguredOpenAICompatProvider so tool round-trips work. */
+    messageContentFormat?: "string";
     registryDefaultIgnoresModelEnvVar?: boolean;
 };
 export type CatalogBillingPolicy = "free-tier" | "free-with-card" | "no-free-tier";

package/dist/types/providers.d.ts

@@ -693,6 +693,9 @@ export type OpenAICompatCatalogEntry = {
      * the classifier's default (six of the seven catalog entries).
      */
     timeoutErrorClass?: new (message: string, provider?: string) => ProviderError;
+    /** See CatalogQuirks.messageContentFormat — a vendor that accepts
+     *  `messages[].content` only as a plain string. */
+    messageContentFormat?: "string";
 };
 /** The subset of OpenAICompatCatalogEntry that resolveOpenAICompatConfig()
  *  needs — lets call sites pass a minimal object without the full catalog

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.9.0",
+  "version": "12.9.2",
   "packageManager": "pnpm@10.15.1",
   "description": "TypeScript AI SDK with 24+ LLM providers behind one consistent API. MCP-native (connect any MCP server), voice TTS/STT/realtime, RAG, agents, memory, context compaction. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · DeepSeek · NVIDIA NIM and more.",
   "author": {
@@ -41,6 +41,7 @@
     "prepack": "svelte-kit sync && svelte-package && pnpm run build:react-hooks && pnpm run build:cli && pnpm run build:browser && publint",
     "build:react-hooks": "pnpm exec tsc --jsx react-jsx --module nodenext --moduleResolution nodenext --target esnext --esModuleInterop --skipLibCheck --outDir dist --declaration false src/lib/client/reactHooks.tsx",
     "check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json && tsc --noEmit --strict",
+    "check:models": "tsx tools/check-model-liveness.ts",
     "check:ci-scripts": "svelte-kit sync && tsc -p tsconfig.ci-scripts.json",
     "check:test-parse": "node scripts/parse-check-tests.mjs",
     "check:tools-tests": "svelte-kit sync && tsc -p tsconfig.tools-tests.json",