fix: 按 DeepSeek 文档为每种调用指定是否思考，修复分类器拿不到结果

deepseek-flash 默认先思考（reasoning_effort high），思考也占输出 token，思考模式下 temperature 无效。
分类器只给 10 个 token，全被思考用掉，正文为空，切到新后端后的提问都没有层级。

- 生成代码：开思考；引导：关思考，回得快；分类：关思考，temperature 0 才生效
- thinking 只发给 DeepSeek，豆包请求不变
- 实测 deepseek-flash：分类约 1.2s，引导约 1.5s；固定 68 条样本两次一致 65 条

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

server/src/services/classifier.ts

@@ -37,8 +37,8 @@ async function callLlm(content: string): Promise<number | null> {
         { role: "system", content: CLASSIFY_SYSTEM_PROMPT },
         { role: "user", content },
       ],
-      // 同一句提示词要得到同一个层级
+      // 只回一个数字，不用思考；关了思考 temperature 才生效，同一句提示词得到同一个层级
-      { maxTokens: 10, temperature: 0, timeoutMs: 30_000 },
+      { thinking: false, maxTokens: 10, temperature: 0, timeoutMs: 30_000 },
     )
     const match = /[1-6]/.exec(text)
     if (!match) {

server/src/services/llm.ts

@@ -59,9 +59,12 @@ export interface ChatMessage {
   content: string
 }
 
+// DeepSeek（deepseek-flash）默认先思考再回答，思考的内容在 reasoning_content 里，也算输出 token；
+// 思考模式下 temperature 不起作用。thinking 只发给 DeepSeek，豆包不传
 interface CompletionOptions {
   model?: string
   stream: boolean
+  thinking?: boolean
   maxTokens?: number
   temperature?: number
   signal?: AbortSignal
@@ -82,6 +85,7 @@ function buildRequest(messages: ChatMessage[], opts: CompletionOptions) {
     messages,
     stream: opts.stream,
   }
+  if (!ark && opts.thinking !== undefined) body.thinking = { type: opts.thinking ? "enabled" : "disabled" }
   if (opts.maxTokens) body.max_tokens = opts.maxTokens
   if (opts.temperature !== undefined) body.temperature = opts.temperature
   const apiKey = ark ? config.arkApiKey : config.llmApiKey
@@ -109,8 +113,7 @@ async function ensureOk(res: Response) {
 // 读取 OpenAI 兼容接口的 SSE 流，逐段产出文本
 async function* streamCompletion(
   messages: ChatMessage[],
-  model?: string,
+  { model, thinking, signal }: { model?: string; thinking?: boolean; signal?: AbortSignal },
-  signal?: AbortSignal,
 ): AsyncGenerator<string> {
   // 带思考的模型生成整页代码可能超过两分钟，只要一直有数据就不该中断
   const idle = new AbortController()
@@ -125,6 +128,7 @@ async function* streamCompletion(
     const res = await buildRequest(messages, {
       model,
       stream: true,
+      thinking,
       signal: AbortSignal.any(signals),
       timeoutMs: STREAM_MAX_MS,
     })
@@ -141,6 +145,7 @@ async function* streamCompletion(
         if (data === "[DONE]") return true
         const json = JSON.parse(data)
         if (json.error) throw new Error(json.error.message ?? JSON.stringify(json.error))
+        // 思考阶段只有 delta.reasoning_content，不发给学生（它也会重置空闲计时）
         const content = json.choices?.[0]?.delta?.content
         if (content) yield content
       }
@@ -187,29 +192,31 @@ export function userFacingError(e: unknown) {
   return status ? `AI 服务出错（${status}），请稍后重试` : "AI 服务出错，请稍后重试"
 }
 
+// 生成整页代码：开着思考，质量优先
 export function streamChat(history: ChatMessage[], model?: string, signal?: AbortSignal) {
-  return streamCompletion(
+  return streamCompletion([{ role: "system", content: SYSTEM_PROMPT }, ...trimHistory(history)], {
-    [{ role: "system", content: SYSTEM_PROMPT }, ...trimHistory(history)],
+    model: chatModel(model),
-    chatModel(model),
+    thinking: true,
     signal,
-  )
+  })
 }
 
+// 引导只是简短追问，不用思考，回得快
 export function streamGuidance(history: ChatMessage[], signal?: AbortSignal) {
-  return streamCompletion(
+  return streamCompletion([{ role: "system", content: GUIDANCE_SYSTEM_PROMPT }, ...history], {
-    [{ role: "system", content: GUIDANCE_SYSTEM_PROMPT }, ...history],
+    thinking: false,
-    undefined,
     signal,
-  )
+  })
 }
 
 export async function complete(
   messages: ChatMessage[],
-  opts: { model?: string; maxTokens?: number; temperature?: number; timeoutMs?: number } = {},
+  opts: { model?: string; thinking?: boolean; maxTokens?: number; temperature?: number; timeoutMs?: number } = {},
 ): Promise<string> {
   const res = await buildRequest(messages, {
     model: opts.model ?? config.llmModel,
     stream: false,
+    thinking: opts.thinking,
     maxTokens: opts.maxTokens,
     temperature: opts.temperature,
     timeoutMs: opts.timeoutMs,

server/tests/unit.test.ts

@@ -5,6 +5,7 @@ import {
   extractCode,
   parseGuidanceResponse,
   streamChat,
+  streamGuidance,
   trimHistory,
   userFacingError,
 } from "../src/services/llm"
@@ -77,12 +78,13 @@ test("CSV 转义", () => {
 })
 
 describe("streamChat", () => {
-  test("不认识的模型换成默认模型，最后一行没有换行也能读到", async () => {
+  test("不认识的模型换成默认模型，最后一行没有换行也能读到；生成代码开思考，引导不开", async () => {
     let model = ""
+    let thinking: unknown
     const server = Bun.serve({
       port: 0,
       async fetch(req) {
-        model = ((await req.json()) as any).model
+        ;({ model, thinking } = (await req.json()) as any)
         const line = (s: string) => `data: ${JSON.stringify({ choices: [{ delta: { content: s } }] })}`
         // 一行拆在两个分块之间，最后一行不带换行
         const body = `${line("你好")}\n\n${line("世")}\n${line("界")}`
@@ -108,6 +110,9 @@ describe("streamChat", () => {
       for await (const s of streamChat([{ role: "user", content: "hi" }], "deepseek-v4-flash-thinking")) out += s
       expect(out).toBe("你好世界")
       expect(model).toBe(config.llmModel)
+      expect(thinking).toEqual({ type: "enabled" })
+      for await (const _ of streamGuidance([{ role: "user", content: "hi" }]));
+      expect(thinking).toEqual({ type: "disabled" })
     } finally {
       Object.assign(config, saved)
       server.stop(true)