From 8809590841d6f32d5c886e334ecd8f8eaf89a8ed Mon Sep 17 00:00:00 2001
From: skogaby <skogabyskogaby@gmail.com>
Date: Thu, 30 Jul 2026 00:35:19 -0500
Subject: [PATCH] fix(llm): stop classifying Bedrock throttling as context
 overflow

Bedrock's ThrottlingException message "Too many tokens, please wait
before trying again." matched the /too many tokens/i overflow pattern
while the throttling exclusion only matched a literal "throttling
error:" prefix the AI SDK never emits. The 429 was converted into a
ContextOverflowError, which skipped retry and triggered a spurious
auto-compaction instead of backing off.

Broaden the exclusion list to match throttle vocabulary and transient
retry language anywhere in the message, and gate parseAPICallError so
a 429 status is never classified as context overflow.
---
 packages/llm/src/provider-error.ts           | 11 ++++++++++-
 packages/llm/test/provider-error.test.ts     |  5 +++++
 packages/opencode/src/provider/error.ts      |  7 ++++++-
 packages/opencode/test/session/retry.test.ts | 18 ++++++++++++++++++
 4 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/packages/llm/src/provider-error.ts b/packages/llm/src/provider-error.ts
index f8b8a5c013e1..da743018658b 100644
--- a/packages/llm/src/provider-error.ts
+++ b/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)) &&
diff --git a/packages/llm/test/provider-error.test.ts b/packages/llm/test/provider-error.test.ts
index c4185969640a..7fd937349a89 100644
--- a/packages/llm/test/provider-error.test.ts
+++ b/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.",
     ]
diff --git a/packages/opencode/src/provider/error.ts b/packages/opencode/src/provider/error.ts
index 21149a2cf389..71540b1cf1fb 100644
--- a/packages/opencode/src/provider/error.ts
+++ b/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,
diff --git a/packages/opencode/test/session/retry.test.ts b/packages/opencode/test/session/retry.test.ts
index 30ac879a6a9d..07f28fdc648e 100644
--- a/packages/opencode/test/session/retry.test.ts
+++ b/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(
       {