fix(memory): route Qwen thinking through chat template kwargs

Memory/src/model/llm.ts

@@ -303,6 +303,7 @@ class HttpLlmClient implements LlmClient {
     const thinkingBudget = thinking.enabled && thinkingUsesEnableThinking(this.config.vendor ?? "", base, model)
       ? this.config.thinkingBudget
       : undefined;
+    const thinkingPayload = openAiCompatibleThinkingPayload(thinking, this.config.extraBody);
     const agentRegion = resolveMemoryAgentRegion(this.config.sourceProvider);
     const response = await postJsonWithRetry<OpenAiChatResponse>({
       actualModelContext: this.config.actualModelContext,
@@ -323,10 +324,10 @@ class HttpLlmClient implements LlmClient {
         ...(!omitTemperature ? { temperature: options.temperature ?? this.config.temperature } : {}),
         max_tokens: options.maxTokens ?? this.config.maxTokens,
         stream: false,
-        ...thinking.fields,
+        ...thinkingPayload.fields,
         ...(thinkingBudget !== undefined ? { thinking_budget: thinkingBudget } : {}),
         ...(options.jsonMode && !omitJsonMode ? { response_format: { type: "json_object" } } : {}),
-        ...(this.config.extraBody ?? {})
+        ...thinkingPayload.extraBody
       }
     });
     const choice = response.choices?.[0];
@@ -561,6 +562,39 @@ class HttpLlmClient implements LlmClient {
   }
 }
 
+function openAiCompatibleThinkingPayload(
+  thinking: ThinkingControl,
+  configuredExtraBody: Record<string, unknown> | undefined
+): { fields: Record<string, unknown>; extraBody: Record<string, unknown> } {
+  const extraBody = configuredExtraBody ?? {};
+  const configuredChatTemplateKwargs = extraBody.chat_template_kwargs;
+  if (
+    !("enable_thinking" in thinking.fields) ||
+    !isUnknownRecord(configuredChatTemplateKwargs)
+  ) {
+    return { fields: thinking.fields, extraBody };
+  }
+
+  const fields = { ...thinking.fields };
+  delete fields.enable_thinking;
+  return {
+    fields,
+    extraBody: {
+      ...extraBody,
+      chat_template_kwargs: {
+        ...configuredChatTemplateKwargs,
+        ...(!Object.hasOwn(configuredChatTemplateKwargs, "enable_thinking")
+          ? { enable_thinking: thinking.enabled }
+          : {})
+      }
+    }
+  };
+}
+
+function isUnknownRecord(value: unknown): value is Record<string, unknown> {
+  return Boolean(value) && typeof value === "object" && !Array.isArray(value);
+}
+
 function openAiCompatibleThinkingControl(input: {
   vendor: string;
   endpoint: string;

Memory/tests/llm-thinking.test.ts

@@ -76,6 +76,39 @@ describe("memory LLM thinking configuration", () => {
     expect(requestBody(fetchMock)).not.toHaveProperty("thinking_budget");
   });
 
+  it("routes Qwen thinking through chat_template_kwargs when the endpoint opts in", async () => {
+    const fetchMock = openAiFetch();
+    vi.stubGlobal("fetch", fetchMock);
+    const client = createLlmClient(llmConfig({
+      vendor: "qwen",
+      endpoint: "https://sglang.example/v1",
+      model: "qwen3.8-27b",
+      enableThinking: true,
+      extraBody: { chat_template_kwargs: { tokenizer_option: "preserved" } }
+    }));
+
+    await client.complete([{ role: "user", content: "filter" }], {
+      operation: "retrieval.filter",
+      thinkingMode: "disabled"
+    });
+    expect(requestBody(fetchMock)).toMatchObject({
+      chat_template_kwargs: {
+        enable_thinking: false,
+        tokenizer_option: "preserved"
+      }
+    });
+    expect(requestBody(fetchMock)).not.toHaveProperty("enable_thinking");
+
+    fetchMock.mockClear();
+    await client.complete([{ role: "user", content: "evolve" }], {
+      operation: "evolution.induction",
+      thinkingMode: "enabled"
+    });
+    expect(requestBody(fetchMock)).toMatchObject({
+      chat_template_kwargs: { enable_thinking: true }
+    });
+  });
+
   it.each([
     ["DeepSeek", "deepseek", "https://api.deepseek.com", "deepseek-v4-pro"],
     ["Zhipu", "zhipu", "https://open.bigmodel.cn/api/paas/v4", "glm-5.1"],