fix(providers): strip cache-boundary marker from non-Anthropic prompts (#89716)

src/llm/providers/google-shared.test.ts

@@ -1,6 +1,7 @@
 // Google shared provider tests cover response conversion and finish reasons.
 import { FinishReason, type GenerateContentResponse } from "@google/genai";
 import { describe, expect, it } from "vitest";
+import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "../../agents/system-prompt-cache-boundary.js";
 import type { AssistantMessage, Model } from "../types.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import {
@@ -146,4 +147,14 @@ describe("buildGoogleGenerateContentParams", () => {
 
     expect(params.config?.stopSequences).toEqual(["STOP"]);
   });
+
+  it("strips the internal cache boundary marker from systemInstruction", () => {
+    const params = buildGoogleGenerateContentParams(model, {
+      systemPrompt: `Stable${SYSTEM_PROMPT_CACHE_BOUNDARY}Dynamic`,
+      messages: [{ role: "user", content: "hello", timestamp: 0 }],
+    });
+
+    expect(params.config?.systemInstruction).toBe("Stable\nDynamic");
+    expect(JSON.stringify(params)).not.toContain("OPENCLAW_CACHE_BOUNDARY");
+  });
 });

src/llm/providers/google-shared.ts

@@ -12,6 +12,7 @@ import {
   type Part,
   type ThinkingConfig,
 } from "@google/genai";
+import { stripSystemPromptCacheBoundary } from "../../agents/system-prompt-cache-boundary.js";
 import { calculateCost, clampThinkingLevel } from "../model-utils.js";
 import type {
   Api,
@@ -500,7 +501,9 @@ export function buildGoogleGenerateContentParams<T extends GoogleApiType>(
 
   const config: GenerateContentConfig = {
     ...(Object.keys(generationConfig).length > 0 && generationConfig),
-    ...(context.systemPrompt && { systemInstruction: sanitizeSurrogates(context.systemPrompt) }),
+    ...(context.systemPrompt && {
+      systemInstruction: sanitizeSurrogates(stripSystemPromptCacheBoundary(context.systemPrompt)),
+    }),
     ...(context.tools && context.tools.length > 0 && { tools: convertTools(context.tools) }),
   };
 

src/llm/providers/mistral.test.ts

@@ -1,5 +1,6 @@
 // Mistral provider tests cover request mapping and stream conversion.
 import { beforeEach, describe, expect, it, vi } from "vitest";
+import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "../../agents/system-prompt-cache-boundary.js";
 import type { Context, Model } from "../types.js";
 
 const mistralMockState = vi.hoisted(() => ({
@@ -214,4 +215,24 @@ describe("Mistral provider", () => {
       function: { name: "healthy_tool" },
     });
   });
+
+  it("strips the internal cache boundary marker from the system message", async () => {
+    const stream = streamSimpleMistral(
+      makeMistralModel(),
+      {
+        systemPrompt: `Stable${SYSTEM_PROMPT_CACHE_BOUNDARY}Dynamic`,
+        messages: [{ role: "user", content: "hello", timestamp: 0 }],
+      },
+      { apiKey: "sk-mistral-provider" },
+    );
+
+    await stream.result();
+
+    const payload = mistralMockState.payloads[0] as {
+      messages: Array<{ role: string; content: string }>;
+    };
+    const systemMessage = payload.messages.find((message) => message.role === "system");
+    expect(systemMessage?.content).toBe("Stable\nDynamic");
+    expect(JSON.stringify(payload)).not.toContain("OPENCLAW_CACHE_BOUNDARY");
+  });
 });

src/llm/providers/mistral.ts

@@ -7,6 +7,7 @@ import type {
   ContentChunk,
   FunctionTool,
 } from "@mistralai/mistralai/models/components";
+import { stripSystemPromptCacheBoundary } from "../../agents/system-prompt-cache-boundary.js";
 import { getEnvApiKey } from "../env-api-keys.js";
 import { calculateCost, clampThinkingLevel } from "../model-utils.js";
 import type {
@@ -309,7 +310,7 @@ function buildChatPayload(
   if (context.systemPrompt) {
     payload.messages.unshift({
       role: "system",
-      content: sanitizeSurrogates(context.systemPrompt),
+      content: sanitizeSurrogates(stripSystemPromptCacheBoundary(context.systemPrompt)),
     });
   }
 

src/llm/providers/openai-chatgpt-responses.test.ts

@@ -1,6 +1,7 @@
 // ChatGPT Responses provider tests cover stream handling and timeout behavior.
 import { MAX_TIMER_TIMEOUT_MS } from "@openclaw/normalization-core/number-coercion";
 import { afterEach, describe, expect, it, vi } from "vitest";
+import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "../../agents/system-prompt-cache-boundary.js";
 import type { Context, Model } from "../types.js";
 import {
   extractOpenAICodexAccountId,
@@ -403,6 +404,60 @@ describe("streamOpenAICodexResponses transport", () => {
     expect(result.errorMessage).toContain("Request timed out after 5ms");
   });
 
+  it("strips the internal cache boundary marker from request instructions", async () => {
+    let capturedPayload: { instructions?: string } | undefined;
+    const stream = streamOpenAICodexResponses(
+      model,
+      {
+        systemPrompt: `Stable${SYSTEM_PROMPT_CACHE_BOUNDARY}Dynamic`,
+        messages: [{ role: "user", content: "hi", timestamp: 1 }],
+      },
+      {
+        apiKey: createJwt({
+          "https://api.openai.com/auth": {
+            chatgpt_account_id: "acct-1",
+          },
+        }),
+        transport: "sse",
+        onPayload: (payload) => {
+          capturedPayload = payload as typeof capturedPayload;
+          throw new Error("stop after payload");
+        },
+      },
+    );
+
+    const result = await stream.result();
+
+    expect(result.stopReason).toBe("error");
+    expect(capturedPayload?.instructions).toBe("Stable\nDynamic");
+    expect(JSON.stringify(capturedPayload)).not.toContain("OPENCLAW_CACHE_BOUNDARY");
+  });
+
+  it("falls back to the default instructions when no system prompt is set", async () => {
+    let capturedPayload: { instructions?: string } | undefined;
+    const stream = streamOpenAICodexResponses(
+      model,
+      { messages: [{ role: "user", content: "hi", timestamp: 1 }] },
+      {
+        apiKey: createJwt({
+          "https://api.openai.com/auth": {
+            chatgpt_account_id: "acct-1",
+          },
+        }),
+        transport: "sse",
+        onPayload: (payload) => {
+          capturedPayload = payload as typeof capturedPayload;
+          throw new Error("stop after payload");
+        },
+      },
+    );
+
+    const result = await stream.result();
+
+    expect(result.stopReason).toBe("error");
+    expect(capturedPayload?.instructions).toBe("You are a helpful assistant.");
+  });
+
   it("prefers promptCacheKey over sessionId for request cache affinity", async () => {
     let payload: unknown;
     vi.stubGlobal(

src/llm/providers/openai-chatgpt-responses.ts

@@ -25,6 +25,7 @@ import {
   resolveTimerTimeoutMs,
   clampTimerTimeoutMs,
 } from "@openclaw/normalization-core/number-coercion";
+import { stripSystemPromptCacheBoundary } from "../../agents/system-prompt-cache-boundary.js";
 import { getEnvApiKey } from "../env-api-keys.js";
 import { clampThinkingLevel } from "../model-utils.js";
 import { registerSessionResourceCleanup } from "../session-resources.js";
@@ -488,7 +489,8 @@ function buildRequestBody(
     model: model.id,
     store: false,
     stream: true,
-    instructions: context.systemPrompt || "You are a helpful assistant.",
+    instructions:
+      stripSystemPromptCacheBoundary(context.systemPrompt ?? "") || "You are a helpful assistant.",
     input: messages,
     text: { verbosity: options?.textVerbosity || "low" },
     include: ["reasoning.encrypted_content"],

src/llm/providers/openai-responses-shared.test.ts

@@ -1,6 +1,7 @@
 // OpenAI Responses shared tests cover tool conversion and response item mapping.
 import type { Tool as OpenAIResponsesTool } from "openai/resources/responses/responses.js";
 import { describe, expect, it } from "vitest";
+import { SYSTEM_PROMPT_CACHE_BOUNDARY } from "../../agents/system-prompt-cache-boundary.js";
 import type { AssistantMessage, AssistantMessageEvent, Context, Model, Tool } from "../types.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import {
@@ -262,6 +263,24 @@ describe("convertResponsesMessages", () => {
     });
   });
 
+  it("strips the internal cache boundary marker from the system prompt message", () => {
+    const input = convertResponsesMessages(
+      nativeOpenAIModel,
+      {
+        systemPrompt: `Stable${SYSTEM_PROMPT_CACHE_BOUNDARY}Dynamic`,
+        messages: [],
+      } satisfies Context,
+      allowedToolCallProviders,
+    );
+
+    expect(input[0]).toMatchObject({
+      type: "message",
+      role: "developer",
+      content: [{ type: "input_text", text: "Stable\nDynamic" }],
+    });
+    expect(JSON.stringify(input)).not.toContain("OPENCLAW_CACHE_BOUNDARY");
+  });
+
   it("omits phase-tagged assistant replay ids without reasoning", () => {
     const input = convertResponsesMessages(
       nativeOpenAIModel,

src/llm/providers/openai-responses-shared.ts

@@ -13,6 +13,7 @@ import type {
   ResponseReasoningItem,
   ResponseStreamEvent,
 } from "openai/resources/responses/responses.js";
+import { stripSystemPromptCacheBoundary } from "../../agents/system-prompt-cache-boundary.js";
 import {
   AZURE_RESPONSES_TEXT_CONTENT_PART_TYPE,
   OPENAI_RESPONSES_OUTPUT_TEXT_CONTENT_PART_TYPE,
@@ -254,7 +255,12 @@ export function convertResponsesMessages<TApi extends Api>(
     messages.push({
       type: "message",
       role,
-      content: [{ type: "input_text", text: sanitizeSurrogates(context.systemPrompt) }],
+      content: [
+        {
+          type: "input_text",
+          text: sanitizeSurrogates(stripSystemPromptCacheBoundary(context.systemPrompt)),
+        },
+      ],
     });
   }