From 43c18cdb88fe46aaace5077832fdcc6b1f01e26d Mon Sep 17 00:00:00 2001
From: yuetsh <517252939@qq.com>
Date: Tue, 27 Oct 2026 21:36:06 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20AI=20=E5=9B=9E=E5=A4=8D=E8=AF=B4?=
 =?UTF-8?q?=E6=98=8E=E6=9B=BF=E5=AD=A6=E7=94=9F=E5=AE=9A=E4=BA=86=E4=BB=80?=
 =?UTF-8?q?=E4=B9=88=E5=B9=B6=E9=97=AE=E4=B8=80=E4=B8=AA=E9=97=AE=E9=A2=98?=
 =?UTF-8?q?=EF=BC=8C=E2=80=9C=E5=B8=AE=E6=88=91=E5=86=99=E6=B8=85=E6=A5=9A?=
 =?UTF-8?q?=E2=80=9D=E8=83=BD=E7=9C=8B=E5=88=B0=E9=A2=98=E7=9B=AE=E5=92=8C?=
 =?UTF-8?q?=E4=B9=8B=E5=89=8D=E7=9A=84=E5=AF=B9=E8=AF=9D?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 生成回复分“做了什么 / 我替你定的 / 想一想”三项；只问问题时直接回答、不给代码
- 没有代码的回复发 code: null，不再清空网页；上下文保留最近一份带代码的回复
- 教练带上题目要求、最近的提问和 AI 上一次的说明，照搬题目不算写清楚

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 CLAUDE.md                         |  2 +-
 server/src/services/classifier.ts | 11 +-----
 server/src/services/llm.ts        | 64 +++++++++++++++++++------------
 server/src/services/prompt.ts     | 43 +++++++++++++++++++++
 server/src/ws.ts                  | 11 +++++-
 server/tests/unit.test.ts         | 15 +++++++-
 server/tests/ws.test.ts           | 50 ++++++++++++++++++++++--
 7 files changed, 155 insertions(+), 41 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 749cc44..9d379ee 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -17,7 +17,7 @@ The backend was rewritten from Django in 2026-10 to cut memory use (~25 MB vs. 1
   - `src/auth.ts` — Django-compatible password hashing; JWT login, no cookies: `POST /account/login` returns a 15-min HS256 `access_token` (sent as `Authorization: Bearer`, signed with `JWT_SECRET`, random per process if unset) and a 14-day opaque `refresh_token` (only its sha256 is stored, in `auth_refresh_token`; `POST /account/refresh` trades it for a new access token, logout / password change / deactivation delete it). An invalid or expired Bearer token is a 401 even on public endpoints so the client knows to refresh. `requireUser/requireAdmin/requireSuper` (401 / 403)
   - `src/routes/` — `account`, `tasks` (tutorial / challenge / creation), `files` (task assets + image upload), `submission`, `showcase` (mounted under `/submission/showcase`), `review` (mounted under `/submission/review`: teacher class × task wall — no `classname` means all classes —, students' peer works with `scope=class|all`), `prompt`
   - Who can read a submission (`canView` in `routes/submission.ts`): staff, the owner, anyone if it's in the showcase, or any student who has submitted the same task (any class). Students' `GET /submission/` only lists their own. Peer works and the random-rating item are anonymous until the viewer rates them (`PUT /:id/score` returns the author). `flag` now only carries `present` (teacher's in-class presentation queue); the old four colors are no longer shown
-  - `src/services/` — `score.ts` (rating weights super×0.5, admin×0.3, normal×0.2, then Bayesian smoothing with C=3 and zones), `gradebook.ts`, `llm.ts` (OpenAI-compatible streaming via fetch; DeepSeek default, ARK for doubao models; chat context is the last 20 messages with older code blocks elided), `classifier.ts` (Bloom-level L1–L6 prompt tagging, runs in background; the model sees the task brief and the previous prompt, temperature 0, prompts under 4 chars get no level), `cascade.ts` (Django's CASCADE deletes done by hand, plus the shared delete rules: `hasRatingsOrAwards`, `deleteRound`), `pyround.ts` (`pyRound`/`pySum` reproduce Python's `round()` and `sum()` so scores match the old system bit-for-bit)
+  - `src/services/` — `score.ts` (rating weights super×0.5, admin×0.3, normal×0.2, then Bayesian smoothing with C=3 and zones), `gradebook.ts`, `llm.ts` (OpenAI-compatible streaming via fetch; DeepSeek default, ARK for doubao models; chat context is the last 20 messages with code kept only in the newest reply that has code. Chat replies are written as **做了什么 / 我替你定的 / 想一想** — the decisions the AI made for the student and one open question, to prompt the next round; a reply to a plain question has no code and the WebSocket sends `code: null`. The “帮我写清楚” coach gets the task brief, recent prompts and the last AI explanation from `guidanceContext` in `services/prompt.ts`, and judges by what the student added beyond the brief, like the classifier), `classifier.ts` (Bloom-level L1–L6 prompt tagging, runs in background; the model sees the task brief and the previous prompt, temperature 0, prompts under 4 chars get no level), `cascade.ts` (Django's CASCADE deletes done by hand, plus the shared delete rules: `hasRatingsOrAwards`, `deleteRound`), `pyround.ts` (`pyRound`/`pySum` reproduce Python's `round()` and `sum()` so scores match the old system bit-for-bit)
   - `src/ws.ts` — `/ws/prompt/<task_id>/?token=<access_token>` (code generation chat; the AI context and the page's chat include `conversation`, `edit` and `manual` messages) and `/ws/guidance/<task_id>/?token=...` (prompt coaching); one message per connection at a time (a message sent while the AI is still replying gets an error, payloads capped at 64 KB); on disconnect the pending user message is deleted and the LLM request aborted
   - `scripts/` — one-off CLIs: `create-superuser.ts`, `classify-prompts.ts`, `recalculate-scores.ts`
   - Roles: `super`, `admin`, `normal`. Task model is multi-table: `task_task` + one of `task_tutorial` / `task_challenge` / `task_creation` (`task_ptr_id`)
diff --git a/server/src/services/classifier.ts b/server/src/services/classifier.ts
index 14afb58..f6f3bfd 100644
--- a/server/src/services/classifier.ts
+++ b/server/src/services/classifier.ts
@@ -1,5 +1,6 @@
 import { all, get, run } from "../db"
 import { complete, llmConfigured } from "./llm"
+import { briefOf } from "./prompt"
 
 // 学生在这里是“指挥 AI 做网页”，看的是这句提问里学生自己想了多少，不是要 AI 做的东西有多大：
 // 把题目要求原样贴过去，功能列得再多也只是照着做（L3）
@@ -24,7 +25,6 @@ const CLASSIFY_SYSTEM_PROMPT = `你是一个教育评估专家。学生在网页
 只返回一个数字（1-6），不要解释。`
 
 const MIN_PROMPT_CHARS = 4
-const BRIEF_CHARS = 800
 const PROMPT_CHARS = 1500
 
 const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms))
@@ -52,15 +52,6 @@ async function callLlm(content: string): Promise<number | null> {
   }
 }
 
-// 题目要求去掉图片、压成一段，太长的截掉
-function briefOf(content: string) {
-  const text = content
-    .replace(/!\[[^\]]*\]\([^)]*\)/g, "")
-    .replace(/\s+/g, " ")
-    .trim()
-  return text.length > BRIEF_CHARS ? text.slice(0, BRIEF_CHARS) + "……" : text
-}
-
 function clip(text: string) {
   return text.length > PROMPT_CHARS ? text.slice(0, PROMPT_CHARS) + "……" : text
 }
diff --git a/server/src/services/llm.ts b/server/src/services/llm.ts
index 2179102..5c02605 100644
--- a/server/src/services/llm.ts
+++ b/server/src/services/llm.ts
@@ -1,18 +1,29 @@
 import { config } from "../config"
 
-export const SYSTEM_PROMPT = `你是一个网页生成助手。根据用户的需求描述，生成网页代码。
+// AI 回复里的说明是学生唯一会读的文字，用它引导学生想下一步：
+// “我替你定的”指出学生没说、AI 自己拿的主意，学生看了才会说“不要这样，因为……”；“想一想”只问不答
+export const SYSTEM_PROMPT = `你是一个网页生成助手，和学生一起做网页。根据学生的描述生成网页代码。
 
 规则：
-1. 使用一个 \`\`\`html 代码块返回所有代码
+1. 生成或修改网页时，使用一个 \`\`\`html 代码块返回所有代码
 2. HTML 代码只需要 body 内的内容，不需要完整的 HTML 文档结构
 3. CSS 样式写在 <style> 标签内，JavaScript 写在 <script> 标签内，都放在代码块里
-4. 用中文回复，先简要说明你做了什么，然后给出代码
-5. 在已有代码基础上修改时，返回完整的修改后代码，不要只返回片段
-6. 由于任何外部链接都被屏蔽，使用纯 HTML、CSS 和 JS 实现功能，不要依赖外部库
+4. 在已有代码基础上修改时，返回完整的修改后代码，不要只返回片段
+5. 由于任何外部链接都被屏蔽，使用纯 HTML、CSS 和 JS 实现功能，不要依赖外部库
+6. 用中文回复，对学生称“你”。先写说明，再给代码。说明分下面三项，标题照抄，每项一两句话，不要写长：
+   - **做了什么**：这次按学生的要求做了哪些改动
+   - **我替你定的**：学生没说、你自己拿主意的地方，比如配色、布局、规则、数值、文字内容；列 1-3 条，只写在网页上看得到、学生可以要求改的。学生这次已经说清楚、没有需要你拿主意的地方，就省略这一项
+   - **想一想**：问学生一个问题，帮他想清楚下一步想要什么。问他自己的想法或判断（如“……的时候你希望它怎样？”“这两种你觉得哪个更适合……，为什么？”），不要列一串功能让他挑，也不要替他回答
+7. 学生只是在问问题（为什么、是什么意思、怎么做到的），没有要求改网页时，用浅显的话直接回答，不给代码，也不用分三项
+8. 学生只说“不行”“有 bug”“没反应”，没说清哪里不对时，照样尽力修好，并在“想一想”里请他下次说清楚：做了什么操作、想看到什么、实际看到了什么
 
 输出格式示例（必须严格遵守，只用一个代码块）：
 
-好的，我为你创建了一个点击按钮变色的示例。
+**做了什么**：做了一个按钮，点一下会变成红色。
+
+**我替你定的**：按钮放在页面左上角；变色后不会再变回来。
+
+**想一想**：再点一次的时候，你希望它变回原来的颜色，还是换成别的颜色？
 
 \`\`\`html
 <style>
@@ -28,23 +39,27 @@ document.getElementById('btn').onclick = function() {
 </script>
 \`\`\``
 
-export const GUIDANCE_SYSTEM_PROMPT = `你是一个提示词写作教练，帮助学生写出清晰、具体的网页需求描述。
+// 判断标准和分类器一致：看学生自己想了什么，照搬题目写得再长也不算清楚
+export const GUIDANCE_SYSTEM_PROMPT = `你是一个提示词写作教练。学生在网页制作课上让 AI 生成或修改网页，你帮他把这一次要对 AI 说的话想清楚、说具体。后面会给你题目要求、他之前对 AI 说过的话和 AI 上一次的说明。
 
-判断标准——满足以下条件视为"够好"：
-- 有明确主题（例如：登录页、计时器、商品卡片列表）
-- 至少包含以下一项具体描述：颜色/布局/风格、交互行为（按钮点击、动画等）、页面内容（文字、数量、图标）
+判断标准——看学生这一次的话，满足下面一条就算"够好"：
+- 第一次描述作品：除了题目要求，有他自己的想法，至少有一处具体的视觉、交互或内容是他自己定的（如配色风格、玩法规则、页面上放什么）
+- 修改已有网页：说清楚要改哪里、改成什么样；或者说清楚出了什么问题——做了什么操作、想看到什么、实际看到了什么
 
 判断为"不够好"的典型情况：
-- 目标太泛，如"做一个好看的页面"
-- 只有主题，完全没有任何视觉、交互或内容描述
+- 只有作品名称，如"做一个象棋游戏"
+- 照搬或改写题目要求，没有自己的想法——哪怕写得很长、功能很多
+- 只说"不行""有 bug""再好看一点"，没说是哪里、要怎样
 
 规则：
-1. 如果提示词不够好，用 1-2 个启发性问题引导学生补充细节，不要直接给出答案
-2. 如果提示词已经够好，以 [READY] 开头回复，简短夸奖学生并说明可以生成了
-3. 用中文回复，语气鼓励，简洁明了
-4. 使用 Markdown 语法高亮关键词，优先突出 **主题**、**视觉**、**交互**、**内容**、**可以生成** 等重点
-5. 如果回复以 [READY] 开头，[READY] 不要加粗，必须保持原始文本
-6. 不要生成任何代码`
+1. 如果不够好，用 1-2 个启发性问题引导学生补充，问他自己的想法，不要直接给出答案，也不要列一串功能让他挑
+2. 照搬题目要求时，先肯定他读懂了题目，再问题目之外他想让作品有什么不一样
+3. AI 上一次的说明里如果有“我替你定的”，可以问他同不同意这些决定
+4. 如果已经够好，以 [READY] 开头回复，简短说明好在哪里，告诉他可以生成了；他还给了理由（为什么这样改）的，特别夸一下
+5. 用中文回复，语气鼓励，简洁明了
+6. 使用 Markdown 语法高亮关键词，优先突出 **视觉**、**交互**、**内容**、**你的想法**、**可以生成** 等重点
+7. 如果回复以 [READY] 开头，[READY] 不要加粗，必须保持原始文本
+8. 不要生成任何代码`
 
 // 走火山方舟（ARK）接口的模型
 const ARK_MODELS = new Set(["doubao-seed-2-1-lite-260915"])
@@ -165,14 +180,15 @@ async function* streamCompletion(
   }
 }
 
-// 发给模型的对话上下文：只留最近 20 条；除最后一条 AI 回复外，旧回复里的代码块换成占位文字。
-// 每次回复都是完整代码，模型只需要最新那份；否则对话越长 token 越多，最后会超出上下文长度
+// 发给模型的对话上下文：只留最近 20 条；除最后一条带代码的 AI 回复外，旧回复里的代码块换成占位文字。
+// 每次回复都是完整代码，模型只需要最新那份；否则对话越长 token 越多，最后会超出上下文长度。
+// 只回答问题的回复没有代码，不能拿它当“最后一条”，不然最新的代码也被省略了
 const MAX_HISTORY = 20
 const CODE_BLOCK = /```[\s\S]*?```/g
 
 export function trimHistory(history: ChatMessage[]): ChatMessage[] {
   const recent = history.slice(-MAX_HISTORY)
-  const last = recent.findLastIndex((m) => m.role === "assistant")
+  const last = recent.findLastIndex((m) => m.role === "assistant" && m.content.includes("```"))
   return recent.map((m, i) =>
     m.role === "assistant" && i !== last
       ? { ...m, content: m.content.replace(CODE_BLOCK, "```\n（旧版本代码已省略）\n```") }
@@ -201,9 +217,9 @@ export function streamChat(history: ChatMessage[], model?: string, signal?: Abor
   })
 }
 
-// 引导只是简短追问，不用思考，回得快
-export function streamGuidance(history: ChatMessage[], signal?: AbortSignal) {
-  return streamCompletion([{ role: "system", content: GUIDANCE_SYSTEM_PROMPT }, ...history], {
+// 引导只是简短追问，不用思考，回得快。context 是题目和学生之前的对话（见 guidanceContext）
+export function streamGuidance(history: ChatMessage[], context: string, signal?: AbortSignal) {
+  return streamCompletion([{ role: "system", content: `${GUIDANCE_SYSTEM_PROMPT}\n\n${context}` }, ...history], {
     thinking: false,
     signal,
   })
diff --git a/server/src/services/prompt.ts b/server/src/services/prompt.ts
index 29407b1..bd6dfea 100644
--- a/server/src/services/prompt.ts
+++ b/server/src/services/prompt.ts
@@ -99,6 +99,49 @@ export function codeAsMarkdown(code: {
   return "```html\n" + parts.join("\n\n") + "\n```"
 }
 
+// 题目要求去掉图片、压成一段，太长的截掉
+const BRIEF_CHARS = 800
+
+export function briefOf(content: string) {
+  const text = content
+    .replace(/!\[[^\]]*\]\([^)]*\)/g, "")
+    .replace(/\s+/g, " ")
+    .trim()
+  return text.length > BRIEF_CHARS ? text.slice(0, BRIEF_CHARS) + "……" : text
+}
+
+const clipTo = (text: string, n: number) => (text.length > n ? text.slice(0, n) + "……" : text)
+
+// “帮我写清楚”的教练要知道题目和学生已经做到哪了，才分得清照搬题目和自己的想法：
+// 题目要求、学生最近几次对 AI 说的话、AI 上一次的说明（代码去掉）
+export function guidanceContext(conversationId: string, taskId: number) {
+  const task = get<{ title: string; content: string }>(
+    `SELECT title, content FROM task_task WHERE id = $id`,
+    { id: taskId },
+  )
+  const asked = all<{ content: string }>(
+    `SELECT content FROM prompt_message
+     WHERE conversation_id = $id AND role = 'user' AND source IN ('conversation', 'manual')
+     ORDER BY created DESC, id DESC LIMIT 5`,
+    { id: conversationId },
+  ).reverse()
+  const reply = get<{ content: string }>(
+    `SELECT content FROM prompt_message
+     WHERE conversation_id = $id AND role = 'assistant' AND source = 'conversation'
+     ORDER BY created DESC, id DESC LIMIT 1`,
+    { id: conversationId },
+  )
+  const replyText = reply?.content.replace(/```[\s\S]*?(```|$)/g, "").trim()
+  return [
+    `题目：${task?.title ?? "（未知）"}`,
+    `题目要求：${task ? briefOf(task.content) || "（无）" : "（无）"}`,
+    `学生之前对 AI 说过的话：${
+      asked.length ? asked.map((m, i) => `\n${i + 1}. ${clipTo(m.content.trim(), 300)}`).join("") : "（还没有，这是第一次描述作品）"
+    }`,
+    `AI 上一次的说明：${replyText ? clipTo(replyText, 600) : "（无）"}`,
+  ].join("\n\n")
+}
+
 export function getPrecedingUserMessage(asst: Pick<MessageRow, "id" | "conversation_id" | "created">) {
   return get<MessageRow>(
     `SELECT * FROM prompt_message
diff --git a/server/src/ws.ts b/server/src/ws.ts
index bb573f4..feb1360 100644
--- a/server/src/ws.ts
+++ b/server/src/ws.ts
@@ -5,6 +5,7 @@ import {
   canAccessTask,
   codeAsMarkdown,
   getOrCreateActiveConversation,
+  guidanceContext,
   insertMessage,
 } from "./services/prompt"
 import {
@@ -151,7 +152,9 @@ async function handlePrompt(ws: WS, data: any) {
     // 有了回复，这条提问才留下来；在后台打 L1–L6 层级
     inBackground(classifyMessage(userMessageId))
     ws.data.pendingUserMessageId = null
-    send(ws, { type: "complete", code, message_id: messageId })
+    // 只回答问题、没给代码的回复不是新版本；发 null，免得工作台把网页清空
+    const hasCode = !!(code.html || code.css || code.js)
+    send(ws, { type: "complete", code: hasCode ? code : null, message_id: messageId })
   } finally {
     dropPendingUserMessage(ws)
   }
@@ -172,7 +175,11 @@ async function handleGuidance(ws: WS, data: any) {
   try {
     let full: string
     try {
-      full = await streamToClient(ws, streamGuidance(ws.data.guidanceHistory, ws.data.abort.signal))
+      const context = guidanceContext(ws.data.conversationId, ws.data.taskId)
+      full = await streamToClient(
+        ws,
+        streamGuidance(ws.data.guidanceHistory, context, ws.data.abort.signal),
+      )
     } catch (e) {
       // 这条提问没有回复，留在上下文里会出现连续两条用户消息
       ws.data.guidanceHistory.pop()
diff --git a/server/tests/unit.test.ts b/server/tests/unit.test.ts
index 860e9bc..208db90 100644
--- a/server/tests/unit.test.ts
+++ b/server/tests/unit.test.ts
@@ -111,7 +111,7 @@ describe("streamChat", () => {
       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" }]));
+      for await (const _ of streamGuidance([{ role: "user", content: "hi" }], "题目：x"));
       expect(thinking).toEqual({ type: "disabled" })
     } finally {
       Object.assign(config, saved)
@@ -140,6 +140,19 @@ test("trimHistory 只留最近 20 条，旧回复的代码换成占位", () => {
   expect(out.filter((m) => m.content.includes("<p>"))).toHaveLength(1)
 })
 
+test("trimHistory 最后一条回复没有代码时，保留之前最新的那份代码", () => {
+  const out = trimHistory([
+    { role: "user", content: "做个页面" },
+    { role: "assistant", content: "第1版\n```html\n<p>1</p>\n```" },
+    { role: "user", content: "加个按钮" },
+    { role: "assistant", content: "第2版\n```html\n<p>2</p>\n```" },
+    { role: "user", content: "为什么按钮是圆的？" },
+    { role: "assistant", content: "因为加了 border-radius。" },
+  ])
+  expect(out[3]!.content).toContain("<p>2</p>")
+  expect(out[1]!.content).not.toContain("<p>1</p>")
+})
+
 test("userFacingError 不透出接口原始返回", () => {
   const quiet = console.error
   console.error = () => {}
diff --git a/server/tests/ws.test.ts b/server/tests/ws.test.ts
index 6cc17db..f43b1e0 100644
--- a/server/tests/ws.test.ts
+++ b/server/tests/ws.test.ts
@@ -2,12 +2,13 @@ import { afterAll, beforeAll, beforeEach, describe, expect, test } from "bun:tes
 import type { Server } from "bun"
 import { app } from "../src/app"
 import { config } from "../src/config"
-import { all } from "../src/db"
+import { all, run } from "../src/db"
 import { upgrade, websocket, type WsData } from "../src/ws"
 import { client, freshDb, makeTask, makeUser } from "./helpers"
 
 // 假的模型接口：每段之间停一下，方便测试中途断开
-const PIECES = ["好的", "```html\n<p>hi</p>\n```"]
+const DEFAULT_PIECES = ["好的", "```html\n<p>hi</p>\n```"]
+let PIECES = DEFAULT_PIECES
 // 最近一次发给模型的请求；分类（非流式）单独记，固定回“3”
 let lastRequest: any = null
 let lastClassify: any = null
@@ -56,7 +57,10 @@ afterAll(() => {
   llm.stop(true)
 })
 
-beforeEach(freshDb)
+beforeEach(() => {
+  freshDb()
+  PIECES = DEFAULT_PIECES
+})
 
 interface Conn {
   ws: WebSocket
@@ -176,9 +180,49 @@ describe("WebSocket 代码生成", () => {
     c.ws.close()
   })
 
+  test("只回答问题、没给代码时 code 是 null", async () => {
+    const { task, token } = await setup()
+    PIECES = ["因为加了 ", "border-radius。"]
+    const c = (await connect(`/ws/prompt/${task}/?token=${token}`))!
+    await c.next("init")
+    c.ws.send(JSON.stringify({ type: "message", content: "为什么按钮是圆的？" }))
+    const done = await c.next("complete")
+    expect(done.code).toBeNull()
+    expect(done.message_id).toBeGreaterThan(0)
+    c.ws.close()
+  })
+
   test("token 无效或任务不存在时握手失败", async () => {
     const { task, token } = await setup()
     expect(await connect(`/ws/prompt/${task}/?token=bad`)).toBeNull()
     expect(await connect(`/ws/prompt/9999/?token=${token}`)).toBeNull()
   })
 })
+
+describe("WebSocket 帮我写清楚", () => {
+  test("教练能看到题目要求、之前的提问和 AI 上一次的说明", async () => {
+    const { task, token } = await setup()
+    run(`UPDATE task_task SET content = $c WHERE id = $id`, { c: "做一个计时器，![图](a.png) 有开始和暂停按钮", id: task })
+    const p = (await connect(`/ws/prompt/${task}/?token=${token}`))!
+    await p.next("init")
+    p.ws.send(JSON.stringify({ type: "message", content: "做个计时器" }))
+    await p.next("complete")
+    p.ws.close()
+
+    const g = (await connect(`/ws/guidance/${task}/?token=${token}`))!
+    await g.next("init")
+    g.ws.send(JSON.stringify({ type: "message", content: "改成粉色" }))
+    await g.next("complete")
+    const system = lastRequest.messages[0]
+    expect(system.role).toBe("system")
+    expect(system.content).toContain("题目要求：做一个计时器， 有开始和暂停按钮")
+    expect(system.content).toContain("1. 做个计时器")
+    expect(system.content).toContain("AI 上一次的说明：好的")
+    expect(system.content).not.toContain("<p>hi</p>")
+    expect(lastRequest.messages.at(-1)).toEqual({ role: "user", content: "改成粉色" })
+    // 等后台打完层级再结束，不然假接口关了它会报错
+    const level = () => all(`SELECT prompt_level FROM prompt_message WHERE source = 'guidance' AND role = 'user'`)[0]?.prompt_level
+    for (let i = 0; i < 50 && level() == null; i++) await Bun.sleep(20)
+    g.ws.close()
+  })
+})