fix(llm): stop classifying Bedrock throttling as context overflow

packages/llm/src/provider-error.ts

@@ -31,7 +31,16 @@ const patterns = [
   /token limit exceeded/i,
 ]
 
-const exclusions = [/^(throttling error|service unavailable):/i, /rate limit/i, /too many requests/i]
+// Throttling messages can mention tokens (e.g. Bedrock's ThrottlingException
+// "Too many tokens, please wait before trying again.") without any prefix, so
+// exclude throttle vocabulary and transient retry language anywhere in the message.
+const exclusions = [
+  /^service unavailable:/i,
+  /throttl/i,
+  /rate limit/i,
+  /too many requests/i,
+  /please (wait|try again)/i,
+]
 
 export const isContextOverflow = (message: string) =>
   !exclusions.some((pattern) => pattern.test(message)) &&

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

@@ -21,6 +21,11 @@ describe("provider error classification", () => {
   test("does not classify rate limits as context overflow", () => {
     const messages = [
       "Throttling error: Too many tokens, please wait before trying again.",
+      // Bedrock ThrottlingException via @ai-sdk/amazon-bedrock (no prefix)
+      "Too many tokens, please wait before trying again.",
+      "ThrottlingException: Too many tokens, please wait before trying again.",
+      "undefined: Too many tokens, please wait before trying again.",
+      "Too many tokens, please try again later.",
       "Rate limit exceeded, please retry after 30 seconds.",
       "Too many requests. Please slow down.",
     ]

packages/opencode/src/provider/error.ts

@@ -165,7 +165,12 @@ export type ParsedAPICallError =
 export function parseAPICallError(input: { providerID: ProviderV2.ID; error: APICallError }): ParsedAPICallError {
   const m = message(input.providerID, input.error)
   const body = json(input.error.responseBody)
-  if (isContextOverflow(m) || input.error.statusCode === 413 || body?.error?.code === "context_length_exceeded") {
+  // 429 is throttling, never a context overflow, even when the message mentions
+  // tokens (e.g. Bedrock's "Too many tokens, please wait before trying again.")
+  const overflow =
+    input.error.statusCode !== 429 &&
+    (isContextOverflow(m) || input.error.statusCode === 413 || body?.error?.code === "context_length_exceeded")
+  if (overflow) {
     return {
       type: "context_overflow",
       message: m,

packages/opencode/test/session/retry.test.ts

@@ -413,6 +413,24 @@ describe("session.message-v2.fromError", () => {
     expect(result.data.isRetryable).toBe(true)
   })
 
+  test("classifies Bedrock throttling as retryable APIError, not context overflow", () => {
+    const error = new APICallError({
+      message: "Too many tokens, please wait before trying again.",
+      url: "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-sonnet/converse-stream",
+      requestBodyValues: {},
+      statusCode: 429,
+      responseHeaders: { "content-type": "application/json" },
+      responseBody: '{"message":"Too many tokens, please wait before trying again."}',
+      isRetryable: true,
+    })
+    const result = MessageV2.fromError(error, { providerID: ProviderV2.ID.make("amazon-bedrock") })
+    expect(SessionV1.ContextOverflowError.isInstance(result)).toBe(false)
+    if (!SessionV1.APIError.isInstance(result)) throw new Error("expected APIError")
+    expect(result.data.isRetryable).toBe(true)
+    expect(result.data.statusCode).toBe(429)
+    expect(SessionRetry.retryable(result, retryProvider)).toBeDefined()
+  })
+
   test("converts OpenAI server_error stream chunks to retryable APIError", () => {
     const result = MessageV2.fromError(
       {