diff --git a/src/gateway/openresponses-http.test.ts b/src/gateway/openresponses-http.test.ts
index 5078f2f4fffd..21c7f18d5ae3 100644
--- a/src/gateway/openresponses-http.test.ts
+++ b/src/gateway/openresponses-http.test.ts
@@ -8,6 +8,7 @@ import { afterAll, beforeAll, beforeEach, describe, expect, it, vi } from "vites
 import { createDeferred } from "../../test/helpers/promise.js";
 import { createClientToolNameConflictError } from "../agents/agent-tool-definition-adapter.js";
 import { FailoverError } from "../agents/failover-error.js";
+import { DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS } from "../agents/tool-result-limits.js";
 import { HISTORY_CONTEXT_MARKER } from "../auto-reply/reply/history.js";
 import { CURRENT_MESSAGE_MARKER } from "../auto-reply/reply/mentions.js";
 import { resetConfigRuntimeState } from "../config/config.js";
@@ -1875,6 +1876,33 @@ describe("OpenResponses HTTP API (e2e)", () => {
     expect(agentCommandMock).toHaveBeenCalledTimes(1);
   });
 
+  it("rejects structured tool output that exceeds the model-visible prompt budget", async () => {
+    agentCommandMock.mockClear();
+    const res = await postResponses(enabledPort, {
+      model: "openclaw",
+      input: [
+        {
+          type: "function_call_output",
+          call_id: "call_oversized",
+          output: [
+            {
+              type: "input_text",
+              text: "x".repeat(DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS + 1),
+            },
+          ],
+        },
+      ],
+    });
+
+    const body = (await res.json()) as { error?: { message?: string; type?: string } };
+    expect(res.status).toBe(400);
+    expect(body.error?.type).toBe("invalid_request_error");
+    expect(body.error?.message).toContain(
+      `${DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS}-character live tool-result limit`,
+    );
+    expect(agentCommandMock).not.toHaveBeenCalled();
+  });
+
   it.each([
     {
       name: "an unknown reasoning summary discriminant",
diff --git a/src/gateway/openresponses-http.ts b/src/gateway/openresponses-http.ts
index 64b2f178f411..ee9ed470d3cf 100644
--- a/src/gateway/openresponses-http.ts
+++ b/src/gateway/openresponses-http.ts
@@ -95,7 +95,7 @@ import {
   type ToolChoiceConstraint,
 } from "./openai-tool-choice.js";
 import { wrapUntrustedFileContent } from "./openresponses-file-content.js";
-import { buildAgentPrompt } from "./openresponses-prompt.js";
+import { buildAgentPrompt, OpenResponsesToolOutputTooLargeError } from "./openresponses-prompt.js";
 import { createAssistantOutputItem, createFunctionCallOutputItem } from "./openresponses-shape.js";
 
 type OpenResponsesHttpOptions = {
@@ -509,7 +509,18 @@ export async function handleOpenResponsesHttpRequest(
     return true;
   }
 
-  const prompt = buildAgentPrompt(payload.input);
+  let prompt: ReturnType<typeof buildAgentPrompt>;
+  try {
+    prompt = buildAgentPrompt(payload.input);
+  } catch (err) {
+    if (err instanceof OpenResponsesToolOutputTooLargeError) {
+      sendJson(res, 400, {
+        error: { message: err.message, type: "invalid_request_error" },
+      });
+      return true;
+    }
+    throw err;
+  }
 
   // Count URL sources request-wide, but replay media only from the current user turn.
   let images: ImageContent[] = [];
diff --git a/src/gateway/openresponses-parity.test.ts b/src/gateway/openresponses-parity.test.ts
index dadf295cee44..42a81c6614be 100644
--- a/src/gateway/openresponses-parity.test.ts
+++ b/src/gateway/openresponses-parity.test.ts
@@ -1,4 +1,5 @@
 import { describe, expect, it } from "vitest";
+import { DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS } from "../agents/tool-result-limits.js";
 import { IMAGE_ONLY_USER_MESSAGE } from "./agent-prompt.js";
 import { CreateResponseBodySchema } from "./open-responses.schema.js";
 import { wrapUntrustedFileContent } from "./openresponses-file-content.js";
@@ -88,6 +89,28 @@ describe("OpenResponses aggregate behavior", () => {
     expect(result.message).toContain("Summarize it");
   });
 
+  it("enforces the canonical live tool-result prompt budget", () => {
+    expect(
+      buildAgentPrompt([
+        {
+          type: "function_call_output",
+          call_id: "call-max",
+          output: "x".repeat(DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS),
+        },
+      ]).message,
+    ).toHaveLength(DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS);
+
+    expect(() =>
+      buildAgentPrompt([
+        {
+          type: "function_call_output",
+          call_id: "call-oversized",
+          output: "x".repeat(DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS + 1),
+        },
+      ]),
+    ).toThrow(`${DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS}-character live tool-result limit`);
+  });
+
   it("preserves attachment-only turn placeholders", () => {
     expect(
       buildAgentPrompt([
diff --git a/src/gateway/openresponses-prompt.ts b/src/gateway/openresponses-prompt.ts
index f1fc37d8fc0f..903c053d9996 100644
--- a/src/gateway/openresponses-prompt.ts
+++ b/src/gateway/openresponses-prompt.ts
@@ -1,4 +1,6 @@
 // Prompt adapter from OpenAI Responses input items to OpenClaw agent messages.
+import { estimateToolResultTextChars } from "../agents/embedded-agent-runner/tool-result-text-budget.js";
+import { DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS } from "../agents/tool-result-limits.js";
 import {
   buildAgentMessageFromConversationEntries,
   type ConversationEntry,
@@ -9,6 +11,8 @@ import type { ContentPart, ItemParam } from "./open-responses.schema.js";
 const FILE_ONLY_USER_MESSAGE = "User sent file(s) with no text.";
 type ResponseMessageItem = Extract<ItemParam, { type: "message" }>;
 
+export class OpenResponsesToolOutputTooLargeError extends Error {}
+
 function extractTextContent(content: string | ContentPart[]): string {
   if (typeof content === "string") {
     return content;
@@ -101,12 +105,19 @@ export function buildAgentPrompt(input: string | ItemParam[]): {
       });
     } else if (item.type === "function_call_output") {
       // Structured SDK tool output remains transport data, not permission to
-      // fetch its file or image URLs; serialize every accepted part losslessly.
+      // fetch its file or image URLs. The HTTP boundary does not yet know the
+      // effective model context, so use the canonical low-context live-result cap.
+      const body = typeof item.output === "string" ? item.output : JSON.stringify(item.output);
+      if (estimateToolResultTextChars(body) > DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS) {
+        throw new OpenResponsesToolOutputTooLargeError(
+          `Function call output exceeds the ${DEFAULT_MAX_LIVE_TOOL_RESULT_CHARS}-character live tool-result limit.`,
+        );
+      }
       conversationEntries.push({
         role: "tool",
         entry: {
           sender: `Tool:${item.call_id}`,
-          body: typeof item.output === "string" ? item.output : JSON.stringify(item.output),
+          body,
         },
       });
     }