fix(ai): keep known throttles out of overflow matching (#58866)

Author: rekram1-node

packages/ai/src/provider-error.ts

@@ -262,7 +262,7 @@ export function classifyProviderFailure(input: ProviderFailure): AIError["reason
     (codes.includes("context_length_exceeded") ||
       codes.includes("model_context_window_exceeded") ||
       codes.includes("request_too_large") ||
-      isContextOverflow(text))
+      (!codes.includes("throttlingexception") && isContextOverflow(text)))
   )
     return new InvalidRequestError({ ...details, classification: "context-overflow" })
   if (input.status === 413 || isPayloadTooLarge(text))

packages/ai/test/provider-error.test.ts

@@ -1,6 +1,6 @@
 import { describe, expect, test } from "bun:test"
 import { isContextOverflow, isContextOverflowFailure, isRetryable } from "../src/effect.js"
-import { AIError } from "../src/schema/index.js"
+import { AIError, HttpContext } from "../src/schema/index.js"
 import { classifyProviderFailure, providerErrorMessage } from "../src/provider-error.js"
 
 describe("provider error classification", () => {
@@ -449,6 +449,37 @@ describe("provider error classification", () => {
     expect(isContextOverflow("Too many tokens")).toBe(true)
   })
 
+  test("does not infer context overflow from known throttling exceptions", () => {
+    const cases = [
+      { status: 429, data: { code: "ThrottlingException" } },
+      { status: 429, rawBody: '{"__type":"com.amazonaws.bedrock#ThrottlingException"}' },
+      { rawBody: '{"exception":{"type":"throttlingException"}}' },
+      {
+        status: 429,
+        http: new HttpContext({
+          url: "https://bedrock.test/model/example/converse-stream",
+          status: 429,
+          headers: { "x-amzn-errortype": "ThrottlingException:http://internal.amazon.com/coral/" },
+        }),
+      },
+    ]
+    for (const input of cases) {
+      for (const message of ["Token limit exceeded", "Too many tokens, please try again later"]) {
+        const error = new AIError({ reason: classifyProviderFailure({ ...input, message }) })
+        expect(error.reason._tag).toBe("RateLimit")
+        expect(isContextOverflowFailure(error)).toBe(false)
+        expect(isRetryable(error)).toBe(true)
+      }
+    }
+    expect(
+      classifyProviderFailure({
+        message: "prompt is too long",
+        status: 400,
+        data: { code: "ValidationException" },
+      }),
+    ).toMatchObject({ _tag: "InvalidRequest", classification: "context-overflow" })
+  })
+
   test("does not let substituted server codes make a 4xx retryable", () => {
     const openai = { error: { type: "server_error", message: "Upstream request failed: Model is unavailable." } }
     const anthropic = {