@juspay/neurolink 12.9.1 → 12.9.3

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/accountLedger.d.ts

@@ -29,10 +29,13 @@ import type { CliAccountUsageTotals } from "../types/index.js";
 /** UTC date stamp of the log file the totals cover. */
 export declare function currentUsageDate(now?: Date): string;
 /**
- * Token and cost totals for one UTC day, keyed by bare account label.
+ * Token and cost totals for one UTC day, keyed by provider-qualified account
+ * key ("anthropic:<label>" / "codex:<label>").
  *
- * Only Anthropic-pool rows are attributed, so a Codex account sharing a label
- * with an Anthropic one cannot contribute its tokens to the wrong row.
+ * Both engines' rows are attributed, each under its own key, so a Codex login
+ * sharing an email with an Anthropic one contributes to its own row and never
+ * to the other's. Keying by bare label used to force a choice between merging
+ * them and dropping Codex outright; it dropped Codex.
  */
 export declare function readAccountUsage(date?: string): Promise<Map<string, CliAccountUsageTotals>>;
 /** Drop all cached cursors. Exported for tests, which vary HOME per case. */

package/dist/proxy/accountLedger.js

@@ -33,9 +33,29 @@ const REQUEST_FILE_PATTERN = /^proxy-\d{4}-\d{2}-\d{2}\.jsonl$/;
 /**
  * Account types written by the Anthropic pool. The request log is shared with
  * the Codex engine, and an operator can use the same email for both, so rows
- * must be filtered by engine before being attributed to an Anthropic account.
+ * are attributed by provider-qualified key, never by bare label.
  */
 const ANTHROPIC_ACCOUNT_TYPES = new Set(["oauth", "api_key"]);
+const CODEX_ACCOUNT_TYPE = "codex-oauth";
+/**
+ * The key a log row is attributed under. Current builds write `accountKey`;
+ * rows from before that field existed carry only the label and the engine's
+ * account type, which is enough to rebuild it — the two engines never share a
+ * type string.
+ */
+function ledgerAccountKey(record, account) {
+    if (typeof record.accountKey === "string" && record.accountKey.length > 0) {
+        return record.accountKey;
+    }
+    const type = typeof record.accountType === "string" ? record.accountType : "";
+    if (ANTHROPIC_ACCOUNT_TYPES.has(type)) {
+        return `anthropic:${account.trim().toLowerCase()}`;
+    }
+    if (type === CODEX_ACCOUNT_TYPE) {
+        return `codex:${account}`;
+    }
+    return account;
+}
 const cursors = new Map();
 function getLogsDir() {
     return join(homedir(), ".neurolink", "logs");
@@ -134,9 +154,11 @@ async function advanceCursor(fileName, cursor) {
         // on the triple means two distinct calls that collide on id but differ in
         // account or model are still counted separately; a true re-log of the same
         // request keeps the same triple and merges.
-        const entryKey = `${requestId}\u0000${account}\u0000${typeof record.model === "string" ? record.model : ""}`;
+        const accountKey = ledgerAccountKey(record, account);
+        const entryKey = `${requestId}\u0000${accountKey}\u0000${typeof record.model === "string" ? record.model : ""}`;
         const next = {
             account,
+            accountKey,
             accountType: typeof record.accountType === "string" ? record.accountType : "",
             model: typeof record.model === "string" ? record.model : "",
             provider: typeof record.provider === "string" ? record.provider : undefined,
@@ -254,10 +276,13 @@ function emptyTotals() {
     };
 }
 /**
- * Token and cost totals for one UTC day, keyed by bare account label.
+ * Token and cost totals for one UTC day, keyed by provider-qualified account
+ * key ("anthropic:<label>" / "codex:<label>").
  *
- * Only Anthropic-pool rows are attributed, so a Codex account sharing a label
- * with an Anthropic one cannot contribute its tokens to the wrong row.
+ * Both engines' rows are attributed, each under its own key, so a Codex login
+ * sharing an email with an Anthropic one contributes to its own row and never
+ * to the other's. Keying by bare label used to force a choice between merging
+ * them and dropping Codex outright; it dropped Codex.
  */
 export async function readAccountUsage(date = currentUsageDate()) {
     const fileName = `proxy-${date}.jsonl`;
@@ -281,10 +306,12 @@ export async function readAccountUsage(date = currentUsageDate()) {
     const totals = new Map();
     const unpriced = new Map();
     for (const entry of cursor.entries.values()) {
-        if (!ANTHROPIC_ACCOUNT_TYPES.has(entry.accountType)) {
+        if (!ANTHROPIC_ACCOUNT_TYPES.has(entry.accountType) &&
+            entry.accountType !== CODEX_ACCOUNT_TYPE) {
+            // Translation and internal rows are proxy plumbing, not a login's spend.
             continue;
         }
-        const row = totals.get(entry.account) ?? emptyTotals();
+        const row = totals.get(entry.accountKey) ?? emptyTotals();
         row.requests += 1;
         row.inputTokens += entry.inputTokens;
         row.outputTokens += entry.outputTokens;
@@ -318,12 +345,12 @@ export async function readAccountUsage(date = currentUsageDate()) {
             }
             else {
                 row.unpricedRequests += 1;
-                const seen = unpriced.get(entry.account) ?? new Set();
+                const seen = unpriced.get(entry.accountKey) ?? new Set();
                 seen.add(entry.model);
-                unpriced.set(entry.account, seen);
+                unpriced.set(entry.accountKey, seen);
             }
         }
-        totals.set(entry.account, row);
+        totals.set(entry.accountKey, row);
     }
     for (const [account, row] of totals) {
         row.costUsd = Number(row.costUsd.toFixed(6));

package/dist/server/routes/claudeProxyRoutes.d.ts

@@ -12,6 +12,7 @@
 import { buildTranslationOptions } from "../../proxy/proxyTranslationEngine.js";
 import { ProxyTracer } from "../../proxy/proxyTracer.js";
 import { isPermanentRefreshFailure } from "../../proxy/tokenRefresh.js";
+import type { ProxyAccountDirectoryOverride } from "../../types/index.js";
 import type { AccountAllowlist, AccountAdmissionLease, JsonObject, AccountCooldownPlan, AccountQuota, AccountQuotaWindow, AccountUsageFetchResult, AnthropicAttemptLogger, AnthropicAuthRetryResult, AnthropicEntitlementFailure, AnthropicInvalidRequestFailure, AnthropicLoopState, AnthropicScopedExhaustion, AnthropicNonOkResult, AnthropicSuccessResult, AnthropicUpstreamFetchResult, ClaudeFinalRequestLogger, ClaudeLoggedErrorBuilder, ClaudeRequest, ClaudeProxyRouteRuntimeOptions, ModelRouterInterface, ParsedClaudeError, ProxyAccountRoutingDecision, ProxyAccountSortMetrics, ProxyBodyCaptureLogger, ProxyLimitsRefreshResponse, ProxyQuotaCooldownUpdate, ProxyOveragePolicy, ProxyPassthroughAccount, QueuedAccountAdmission, RouteGroup, RuntimeAccountState, ServerContext, StreamTerminalOutcome } from "../../types/index.js";
 declare function tryAcquireAccountAdmission(accountKey: string, capacity: number | undefined): AccountAdmissionLease | undefined;
 declare function enqueueAccountAdmission(accountKey: string, capacity: number): QueuedAccountAdmission;
@@ -467,6 +468,7 @@ export declare const __testHooks: {
     refreshAccountLimits: typeof refreshAccountLimits;
     applyAccountUsageResult: typeof applyAccountUsageResult;
     clearLimitsRefreshStateForTests: () => void;
+    setAccountDirectoryForTests: (override: ProxyAccountDirectoryOverride | null) => void;
     isRetryableNetworkError: typeof isRetryableNetworkError;
     isPermanentRefreshFailure: typeof isPermanentRefreshFailure;
     getStreamFailureDetails: typeof getStreamFailureDetails;