fix(memory): add dedicated SGLang and vLLM thinking transports

Memory/src/config/index.ts

@@ -15,6 +15,8 @@ export type LlmProviderName =
   | ""
   | "local_only"
   | "openai_compatible"
+  | "sglang"
+  | "vllm"
   | "gemini"
   | "anthropic"
   | "bedrock"
@@ -749,6 +751,10 @@ function unavailableLlm(defaults: LlmConfig): Record<string, unknown> {
 
 function memoryLlmProvider(provider: string): LlmProviderName {
   switch (provider) {
+    case "sglang":
+      return "sglang";
+    case "vllm":
+      return "vllm";
     case "anthropic":
       return "anthropic";
     case "google":
@@ -1137,6 +1143,8 @@ function llmProvider(value: unknown, fallback: LlmProviderName): LlmProviderName
     provider === "" ||
     provider === "local_only" ||
     provider === "openai_compatible" ||
+    provider === "sglang" ||
+    provider === "vllm" ||
     provider === "gemini" ||
     provider === "anthropic" ||
     provider === "bedrock" ||

Memory/src/model/llm.ts

@@ -270,6 +270,8 @@ class HttpLlmClient implements LlmClient {
   private completeOnce(messages: LlmMessage[], options: Required<Pick<LlmCompletionOptions, "operation">> & LlmCompletionOptions): Promise<LlmCallResult> {
     switch (this.config.provider) {
       case "openai_compatible":
+      case "sglang":
+      case "vllm":
         return this.completeOpenAiCompatible(messages, options);
       case "gemini":
         return this.completeGemini(messages, options);
@@ -303,6 +305,11 @@ class HttpLlmClient implements LlmClient {
     const thinkingBudget = thinking.enabled && thinkingUsesEnableThinking(this.config.vendor ?? "", base, model)
       ? this.config.thinkingBudget
       : undefined;
+    const thinkingPayload = openAiCompatibleThinkingPayload(
+      this.config.provider,
+      thinking,
+      this.config.extraBody
+    );
     const agentRegion = resolveMemoryAgentRegion(this.config.sourceProvider);
     const response = await postJsonWithRetry<OpenAiChatResponse>({
       actualModelContext: this.config.actualModelContext,
@@ -323,10 +330,10 @@ 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 +568,43 @@ class HttpLlmClient implements LlmClient {
   }
 }
 
+function openAiCompatibleThinkingPayload(
+  provider: LlmConfig["provider"],
+  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 (
+    (provider !== "sglang" && provider !== "vllm") ||
+    !("enable_thinking" in thinking.fields) ||
+    (configuredChatTemplateKwargs !== undefined &&
+      !isUnknownRecord(configuredChatTemplateKwargs))
+  ) {
+    return { fields: thinking.fields, extraBody };
+  }
+  const chatTemplateKwargs = configuredChatTemplateKwargs ?? {};
+
+  const fields = { ...thinking.fields };
+  delete fields.enable_thinking;
+  return {
+    fields,
+    extraBody: {
+      ...extraBody,
+      chat_template_kwargs: {
+        ...chatTemplateKwargs,
+        ...(!Object.hasOwn(chatTemplateKwargs, "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 +76,65 @@ describe("memory LLM thinking configuration", () => {
     expect(requestBody(fetchMock)).not.toHaveProperty("thinking_budget");
   });
 
+  it.each(["sglang", "vllm"] as const)(
+    "routes Qwen thinking through chat_template_kwargs for the %s provider",
+    async (provider) => {
+      const fetchMock = openAiFetch();
+      vi.stubGlobal("fetch", fetchMock);
+      const client = createLlmClient(llmConfig({
+        provider,
+        vendor: "qwen",
+        endpoint: `https://${provider}.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("keeps chat_template_kwargs opt-in data isolated on the generic OpenAI-compatible provider", async () => {
+    const fetchMock = openAiFetch();
+    vi.stubGlobal("fetch", fetchMock);
+    const client = createLlmClient(llmConfig({
+      provider: "openai_compatible",
+      vendor: "qwen",
+      endpoint: "https://cloud-provider.example/v1",
+      model: "qwen3.8-27b",
+      extraBody: { chat_template_kwargs: { tokenizer_option: "preserved" } }
+    }));
+
+    await client.complete([{ role: "user", content: "filter" }], {
+      operation: "retrieval.filter",
+      thinkingMode: "disabled"
+    });
+
+    expect(requestBody(fetchMock)).toMatchObject({
+      enable_thinking: false,
+      chat_template_kwargs: { tokenizer_option: "preserved" }
+    });
+  });
+
   it.each([
     ["DeepSeek", "deepseek", "https://api.deepseek.com", "deepseek-v4-pro"],
     ["Zhipu", "zhipu", "https://open.bigmodel.cn/api/paas/v4", "glm-5.1"],