diff --git a/packages/agent/src/adapters/error-classification.test.ts b/packages/agent/src/adapters/error-classification.test.ts
index d15bb4851f..af20ac5b3f 100644
--- a/packages/agent/src/adapters/error-classification.test.ts
+++ b/packages/agent/src/adapters/error-classification.test.ts
@@ -1,5 +1,8 @@
 import { describe, expect, it } from "vitest";
-import { classifyAgentError } from "./error-classification";
+import {
+  classifyAgentError,
+  isPromptTooLongError,
+} from "./error-classification";
 
 describe("classifyAgentError", () => {
   it.each([
@@ -22,6 +25,8 @@ describe("classifyAgentError", () => {
     ["API Error: 429 rate limited", "upstream_provider_failure"],
     ["API Error: 529 overloaded", "upstream_provider_failure"],
     ["API Error: 400 invalid request", "agent_error"],
+    // 413 is a hard client rejection, never a transient upstream failure.
+    ["API Error: 413 Payload Too Large", "agent_error"],
     [
       "Connection closed mid-response without the API Error prefix",
       "agent_error",
@@ -32,3 +37,40 @@ describe("classifyAgentError", () => {
     expect(classifyAgentError(message)).toBe(expected);
   });
 });
+
+describe("isPromptTooLongError", () => {
+  it.each([
+    [
+      'API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"prompt is too long: 214431 tokens > 204698 maximum"}}',
+      true,
+    ],
+    [
+      'API Error: 413 {"error":{"message":"litellm.ContextWindowExceededError: The estimated number of input and maximum output tokens (262334) exceeded this model context window limit (262144)","code":"5021"}}',
+      true,
+    ],
+    // Must match without the "API Error: 413" prefix.
+    [
+      "litellm.ContextWindowExceededError: The estimated number of input and maximum output tokens (262334) exceeded this model context window limit (262144)",
+      true,
+    ],
+    // The ACP-wrapped shape the agent-server catch actually sees.
+    [
+      'Internal error: API Error: 413 {"error":{"message":"exceeded this model context window limit (262144)"}}',
+      true,
+    ],
+    // Any gateway 413 means an oversized payload, whatever the body text.
+    ["API Error: 413 Payload Too Large", true],
+    // Pins the 413 matcher's i flag.
+    ["api error: 413 payload too large", true],
+    ["API Error: 429 rate limited", false],
+    ["API Error: 400 invalid request", false],
+    ["some unrelated failure", false],
+  ] as const)("detects %j as %s", (message, expected) => {
+    expect(isPromptTooLongError(new Error(message))).toBe(expected);
+  });
+
+  it("handles non-Error inputs", () => {
+    expect(isPromptTooLongError({ message: "prompt is too long" })).toBe(true);
+    expect(isPromptTooLongError(undefined)).toBe(false);
+  });
+});
diff --git a/packages/agent/src/adapters/error-classification.ts b/packages/agent/src/adapters/error-classification.ts
index 83a88f78fe..56e5268585 100644
--- a/packages/agent/src/adapters/error-classification.ts
+++ b/packages/agent/src/adapters/error-classification.ts
@@ -48,7 +48,16 @@ export function classifyAgentError(
   return "agent_error";
 }
 
-/** Hard API rejection: the assembled prompt exceeds the model's context window. */
+/**
+ * Hard API rejection: the prompt exceeds the model's context window
+ * (Anthropic phrasing, or the LLM gateway's HTTP 413). Retrying the same
+ * transcript can never succeed; callers must shrink the prompt.
+ */
 export function isPromptTooLongError(error: unknown): boolean {
-  return /prompt is too long/i.test(getErrorMessage(error));
+  const message = getErrorMessage(error);
+  return (
+    /prompt is too long/i.test(message) ||
+    /exceeded this model context window limit/i.test(message) ||
+    /API Error:\s*413\b/i.test(message)
+  );
 }
diff --git a/packages/agent/src/server/agent-server.test.ts b/packages/agent/src/server/agent-server.test.ts
index 49cd5ca9d2..171e953ba0 100644
--- a/packages/agent/src/server/agent-server.test.ts
+++ b/packages/agent/src/server/agent-server.test.ts
@@ -3406,11 +3406,27 @@ describe("AgentServer HTTP Mode", () => {
     });
 
     it.each([
-      { retryOutcome: "succeeds", retryFails: false },
-      { retryOutcome: "fails", retryFails: true },
+      {
+        retryOutcome: "succeeds",
+        retryFails: false,
+        oversizedError: "Internal error: Prompt is too long",
+      },
+      {
+        retryOutcome: "fails",
+        retryFails: true,
+        oversizedError: "Internal error: Prompt is too long",
+      },
+      // The LLM gateway phrases oversized rejections as HTTP 413; the
+      // fresh-session retry must trigger on that shape too.
+      {
+        retryOutcome: "succeeds after a gateway 413",
+        retryFails: false,
+        oversizedError:
+          'Internal error: API Error: 413 {"error":{"message":"litellm.ContextWindowExceededError: The estimated number of input and maximum output tokens (262334) exceeded this model context window limit (262144)","code":"5021"}}',
+      },
     ])(
       "clears resume state when the fresh-session retry $retryOutcome",
-      async ({ retryFails }) => {
+      async ({ retryFails, oversizedError }) => {
         const s = createServer();
         await s.start();
 
@@ -3418,7 +3434,7 @@ describe("AgentServer HTTP Mode", () => {
         const prompt = vi.fn(async (params: { prompt: ContentBlock[] }) => {
           prompts.push(params.prompt);
           if (prompts.length === 1) {
-            throw new Error("Internal error: Prompt is too long");
+            throw new Error(oversizedError);
           }
           if (retryFails) {
             throw new Error("Fresh-session retry failed");