@juspay/neurolink 11.23.2 → 11.23.3

package/dist/providers/openAI/client.js

@@ -12,6 +12,7 @@ import { assertSafeUrl } from "../../utils/ssrfGuard.js";
 import { createTimeoutController } from "../../utils/timeout.js";
 import { stripTrailingSlash } from "../openaiChatCompletionsClient.js";
 import { OpenAIChatCompletionsProvider } from "../openaiChatCompletionsBase.js";
+import { isOpenAIQuotaExhaustedError } from "../../utils/providerRetry.js";
 const OPENAI_DEFAULT_BASE_URL = "https://api.openai.com/v1";
 /**
  * Resolve the effective OpenAI base URL from optional credential / env
@@ -113,6 +114,23 @@ export class OpenAIProvider extends OpenAIChatCompletionsProvider {
                     ? ctx.message
                     : "Invalid OpenAI API key. Please check your OPENAI_API_KEY environment variable.",
             },
+            // MUST precede the 429 rule below. OpenAI returns 429 for two
+            // unrelated conditions: transient throttling, and a permanently
+            // exhausted quota (out of credit, or a spend cap reached). Matching on
+            // the status code alone collapses them, and the resulting advice —
+            // "try again later" — is the one thing that can never resolve the
+            // second. The signal is the error TYPE, not the wording.
+            {
+                // One definition of "permanent billing state", in the helper. The
+                // earlier cut also tested `errorType === "insufficient_quota"` here,
+                // which is precisely the helper's own first branch — two places to
+                // keep in step for no benefit.
+                match: (ctx) => isOpenAIQuotaExhaustedError(ctx.error),
+                errorClass: ProviderError,
+                message: "OpenAI quota exhausted — this will not resolve by retrying. " +
+                    "Check your plan and billing details at " +
+                    "https://platform.openai.com/account/billing",
+            },
             {
                 match: (ctx) => ctx.statusCode === 429 ||
                     errorType === "rate_limit_error" ||

package/dist/utils/providerRetry.d.ts

@@ -52,6 +52,24 @@ export declare function duckTypedStatusCode(error: unknown): number | undefined;
  * hand-rolled OpenAI-compatible client's record). Shared with baseProvider.
  */
 export declare function extractRetryAfterMsFromError(error: unknown): number | undefined;
+/**
+ * An OpenAI-wire `insufficient_quota` error: the account is out of credit or
+ * has hit a spend cap.
+ *
+ * Exported and shared with the OpenAI error table so the retry predicate and
+ * the message a caller sees cannot disagree about what a permanent billing
+ * state looks like. The type is NOT a field on the error — OpenAI puts it
+ * inside the JSON response body, which the SDK keeps as a string — so any
+ * caller that needs it has to parse, and none should do so on its own.
+ *
+ * Deliberately keyed on the error TYPE, never on the word "quota". OpenAI and
+ * xAI send 429 + `type: "insufficient_quota"` for a permanent billing state,
+ * but other providers use "quota" for ordinary throttling — Google's 429 reads
+ * "Quota exceeded for quota metric ...", which IS retryable and must stay that
+ * way. Matching free text here would make Gemini throttling non-retryable and
+ * quietly remove a layer of resilience.
+ */
+export declare function isOpenAIQuotaExhaustedError(error: unknown): boolean;
 export declare function isRetryableProviderError(error: unknown): boolean;
 /**
  * Extract the HTTP status code from an AI SDK error, if available.

package/dist/utils/providerRetry.js

@@ -105,7 +105,53 @@ export function extractRetryAfterMsFromError(error) {
     }
     return undefined;
 }
+/**
+ * An OpenAI-wire `insufficient_quota` error: the account is out of credit or
+ * has hit a spend cap.
+ *
+ * Exported and shared with the OpenAI error table so the retry predicate and
+ * the message a caller sees cannot disagree about what a permanent billing
+ * state looks like. The type is NOT a field on the error — OpenAI puts it
+ * inside the JSON response body, which the SDK keeps as a string — so any
+ * caller that needs it has to parse, and none should do so on its own.
+ *
+ * Deliberately keyed on the error TYPE, never on the word "quota". OpenAI and
+ * xAI send 429 + `type: "insufficient_quota"` for a permanent billing state,
+ * but other providers use "quota" for ordinary throttling — Google's 429 reads
+ * "Quota exceeded for quota metric ...", which IS retryable and must stay that
+ * way. Matching free text here would make Gemini throttling non-retryable and
+ * quietly remove a layer of resilience.
+ */
+export function isOpenAIQuotaExhaustedError(error) {
+    if (typeof error !== "object" || error === null) {
+        return false;
+    }
+    const err = error;
+    if (err.type === "insufficient_quota") {
+        return true;
+    }
+    // The SDK keeps the raw payload as a string; the type lives inside it.
+    if (typeof err.responseBody === "string") {
+        try {
+            const parsed = JSON.parse(err.responseBody);
+            const inner = parsed?.error;
+            return inner?.type === "insufficient_quota";
+        }
+        catch {
+            return false;
+        }
+    }
+    return false;
+}
 export function isRetryableProviderError(error) {
+    // Before every other branch, including the SDK's own flag: a 429 carrying
+    // `insufficient_quota` is marked retryable by the AI SDK because it only
+    // looks at the status code. Retrying it burns the full ladder — 3 attempts
+    // and ~20s of backoff, since no Retry-After accompanies these — on a state
+    // that cannot change until someone tops up the account.
+    if (isOpenAIQuotaExhaustedError(error)) {
+        return false;
+    }
     // Preferred path: use the AI SDK's own branded type check + isRetryable flag
     if (APICallError.isInstance(error)) {
         return error.isRetryable;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "11.23.2",
+  "version": "11.23.3",
   "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": {