diff --git a/packages/ai/src/provider-error.ts b/packages/ai/src/provider-error.ts index c0ecbee2b4a..7408f609c03 100644 --- a/packages/ai/src/provider-error.ts +++ b/packages/ai/src/provider-error.ts @@ -65,6 +65,8 @@ const exclusions = [ /too many requests/i, // Cohere reports an output limit above the model maximum as "too many tokens"; compaction cannot fix it. /max[_ ]tokens must be less than/i, + // Bedrock phrases throttling as "Too many tokens[, per day], please wait..."; it is a 429 to retry, not overflow. + /too many tokens(?: per day)?, please wait/i, ] export const isContextOverflow = (message: string) => diff --git a/packages/ai/test/provider-error.test.ts b/packages/ai/test/provider-error.test.ts index 3222e2f1edf..09c637b70f3 100644 --- a/packages/ai/test/provider-error.test.ts +++ b/packages/ai/test/provider-error.test.ts @@ -1,5 +1,6 @@ import { describe, expect, test } from "bun:test" -import { isContextOverflow } from "../src/index.js" +import { isContextOverflow, isContextOverflowFailure, isRetryable } from "../src/index.js" +import { AIError } from "../src/schema/index.js" import { classifyProviderFailure, providerErrorMessage } from "../src/provider-error.js" describe("provider error classification", () => { @@ -423,6 +424,31 @@ describe("provider error classification", () => { ).toEqual(["RateLimit", "RateLimit"]) }) + test("classifies Bedrock throttling as a retryable rate limit, not context overflow", () => { + // Bedrock throttles with 429 ThrottlingException and a "Too many tokens" message; the token + // phrasing used to classify it as overflow and compact the session instead of retrying. + const bodies = [ + '{"message":"Too many tokens, please wait before trying again."}', + '{"message":"Too many tokens per day, please wait before trying again."}', + ] + for (const body of bodies) { + const error = new AIError({ + reason: classifyProviderFailure({ + message: providerErrorMessage(body), + status: 429, + rawBody: body, + }), + }) + expect(error.reason._tag).toBe("RateLimit") + expect(isContextOverflowFailure(error)).toBe(false) + expect(isRetryable(error)).toBe(true) + } + expect(isContextOverflow("Too many tokens, please wait before trying again.")).toBe(false) + expect(isContextOverflow("Too many tokens per day, please wait before trying again.")).toBe(false) + // Plain overflow phrasing without the throttle cue still compacts. + expect(isContextOverflow("Too many tokens")).toBe(true) + }) + 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 = {