feat: 加 prompt-stats 脚本，对比某天前后学生和 AI 对话的情况

轮数分布、层级分布、第 2 轮起的 L1、只粘贴外部代码的比例、“出问题了”“帮我写清楚”
和交作品反思的使用情况。只读。

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

CLAUDE.md

@@ -17,9 +17,9 @@ 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`. For challenges the AI also gets the task brief (`challengeBrief`) as a checklist only (`BRIEF_RULES`): it builds just what the student asked for and asks about one missing requirement in “想一想”; creations get no brief. 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`
+  - `scripts/` — one-off CLIs: `create-superuser.ts`, `classify-prompts.ts`, `recalculate-scores.ts`, `prompt-stats.ts` (read-only before/after comparison of AI-chat behavior: rounds, levels, external-code share, “出问题了” / reflection usage)
   - Roles: `super`, `admin`, `normal`. Task model is multi-table: `task_task` + one of `task_tutorial` / `task_challenge` / `task_creation` (`task_ptr_id`)
 
 - **`client/`** — Vue 3 + TypeScript SPA built with **Vite**; Bun is the package manager and runtime (no Node/npm)
@@ -30,7 +30,7 @@ The backend was rewritten from Django in 2026-10 to cut memory use (~25 MB vs. 1
   - `src/router.ts` — Vue Router; `meta.auth` needs login, `meta.role` (`staff` = admin+super, `super`) is checked in the guard after `store/session.loadUser()`, `meta.bare` pages render without the top bar
   - `src/theme.ts` — the only place colors are defined: one palette (light + dark values) becomes both `--wp-*` CSS variables and Naive UI `themeOverrides`; use `var(--wp-...)` in styles, never hard-coded colors (iframes stay white: they show students' pages). No `color-mix()` (Chrome 105). `store/theme.ts` holds the mode (auto / light / dark, toggled in the top bar) and sets `data-theme` on `<html>`; `styles/dark-content.css` is generated from the GitHub dark markdown and highlight.js styles (selectors prefixed with `:root[data-theme="dark"]`); CodeMirror uses One Dark highlighting in dark mode
   - `src/icons.ts` — `<Icon icon="lucide:name" />` renders bundled `lucide-vue-next` icons (works offline); register new names in its map. Fonts are bundled too (`@fontsource/jetbrains-mono`)
-  - `components/workbench/Workbench.vue` — shared challenge / creation page: left = brief + AI chat (`PromptPanel`), right = preview / code (`CodeTabs`) + `VersionBar`. Every assistant message with code is a version; hand edits and pasted external-AI code are saved as versions via `POST /prompt/versions` (`source` `edit` / `manual`, stored as a user+assistant pair). Challenges are submitted only when the student clicks “交这一版” (`POST /submission/` with `message_id`), never automatically. Editor code is per task (`composables/useTaskCode.ts`, localStorage `web-code:<user>:<task>`); the tutorial page edits the shared `store/editors.ts` refs but `Workspace.vue` saves them under the same per-task key and `TutorialContent` restores them (example code only the first time)
+  - `components/workbench/Workbench.vue` — shared challenge / creation page: left = brief + AI chat (`PromptPanel`), right = preview / code (`CodeTabs`) + `VersionBar`. Every assistant message with code is a version; hand edits and pasted external-AI code are saved as versions via `POST /prompt/versions` (`source` `edit` / `manual`, stored as a user+assistant pair). Challenges are submitted only when the student clicks “交这一版” (`POST /submission/` with `message_id`), never automatically; `SubmitModal` first asks an optional one-line reflection (最满意的改动和理由), stored in `submission_reflection` (`ON DELETE CASCADE`, returned by `GET /submission/:id` only to staff and the author, shown as “学生说” in `ReviewPanel`). In `PromptPanel`, “出问题了” is a three-field form (做了什么 / 想看到什么 / 实际看到了什么) joined into one prompt; a short vague complaint (“还是不行”) gets a one-time hint offering it. Pasted external-AI prompts need at least 8 characters Editor code is per task (`composables/useTaskCode.ts`, localStorage `web-code:<user>:<task>`); the tutorial page edits the shared `store/editors.ts` refs but `Workspace.vue` saves them under the same per-task key and `TutorialContent` restores them (example code only the first time)
   - `pages/ChallengeHome.vue` — challenge list page; `pages/TaskEditor.vue` (prop `kind`) is the single tutorial / challenge authoring page
   - `pages/Review.vue` (“批改”, default teaching page) + `components/review/ReviewPanel.vue` (also opened from the gradebook) — opens as a large preview (wall hidden, Esc returns; “小面板” switches to a side panel, remembered); keys: J/K previous/next student (by key position), 1–5 rate, P present; “全部班级” groups the wall by class; `components/review/PeersDrawer.vue` is the students' “同学作品” (本班 / 全部班级)
   - `components/layout/AppShell.vue` — top bar for every page; `pages/Dashboard.vue` is the teaching sidebar (“教学”) under `/dashboard`
@@ -62,6 +62,7 @@ bun run typecheck
 bun scripts/create-superuser.ts <username> <password>
 bun scripts/classify-prompts.ts [--task-id N] [--force] [--dry-run]
 bun scripts/recalculate-scores.ts [--dry-run]
+bun scripts/prompt-stats.ts <split-date> [--from date] [--to date]   # dates in Beijing time
 ```
 Config comes from `server/.env` (see `server/.env.example`); dev DB is `server/data/db.sqlite3`, uploads go to the repo-root `media/`.
 

client/src/api.ts

@@ -264,6 +264,7 @@ export const Submission = {
       prompt?: string
     },
     messageId?: number,
+    reflection?: string,
   ) {
     const { prompt, ...rest } = code
     const data = {
@@ -271,6 +272,7 @@ export const Submission = {
       ...rest,
       prompt: prompt || null,
       message_id: messageId ?? null,
+      reflection: reflection || null,
     }
     const res = await http.post("/submission/", data)
     return res.data

client/src/components/ai/PromptPanel.vue

@@ -77,63 +77,105 @@
     </div>
     <GuidancePanel @generate="onGuidanceGenerate" />
     <div class="input-area">
-      <n-flex
+      <!-- 只说“不行”“没反应”时提醒一下，可以直接发 -->
-        class="prompt-input-row"
+      <div v-if="vagueHint" class="vague-hint">
-        align="flex-start"
+        <span>说清楚哪里不对，AI 更容易一次修好</span>
-        :wrap="false"
+        <n-flex :size="4" :wrap="false">
-        :size="8"
+          <n-button size="tiny" type="primary" secondary @click="openProblem">说清楚</n-button>
-      >
+          <n-button size="tiny" quaternary @click="sendAnyway">直接发</n-button>
-        <n-input
+        </n-flex>
-          v-model:value="input"
+      </div>
-          type="textarea"
+      <div v-if="problemOpen" class="problem">
-          :autosize="{ minRows: 1, maxRows: 4 }"
+        <div class="problem-title">出问题了？说清楚这三件事</div>
-          placeholder="描述你想要的网页效果..."
+        <label>我做了什么</label>
-          :disabled="streaming"
+        <n-input v-model:value="problem.did" placeholder="比如：点了「开始」按钮" />
-          style="flex: 1; min-width: 0"
+        <label>想看到什么</label>
-          @keydown.enter.exact.prevent="send"
+        <n-input v-model:value="problem.want" placeholder="比如：倒计时开始走" />
-        />
+        <label>实际看到了什么</label>
-        <n-tooltip>
+        <n-input v-model:value="problem.saw" placeholder="比如：数字一直停在 60 不动" />
-          <template #trigger>
+        <n-flex justify="end" :size="8" style="margin-top: 10px">
-            <n-button
+          <n-button @click="problemOpen = false">取消</n-button>
-              secondary
+          <n-button
-              :disabled="!input.trim() || !connected || !currentTaskId || streaming"
+            type="primary"
-              style="flex-shrink: 0"
+            :disabled="!problem.did.trim() || !problem.saw.trim() || !connected || streaming"
-              @click="startGuidance"
+            @click="sendProblem"
-            >
+          >
-              <template #icon>
+            发给 AI
-                <Icon icon="lucide:lightbulb" :width="16" />
+          </n-button>
-              </template>
+        </n-flex>
-              帮我写清楚
+      </div>
-            </n-button>
+      <template v-else>
-          </template>
+        <n-flex
-          AI 教练会问你几个问题，帮你把想要的效果描述清楚
+          class="prompt-input-row"
-        </n-tooltip>
+          align="flex-start"
-      </n-flex>
+          :wrap="false"
-      <n-flex justify="flex-end" align="center" style="margin-top: 8px">
+          :size="8"
-        <n-select
-          v-model:value="selectedModel"
-          :options="modelOptions"
-          style="width: 150px"
-          :disabled="streaming"
-        />
-        <n-button v-if="streaming" type="error" @click="stopPrompt">
-          停止
-        </n-button>
-        <n-button
-          v-else
-          type="primary"
-          :disabled="!input.trim() || !connected"
-          @click="send"
         >
-          发送
+          <n-input
-        </n-button>
+            v-model:value="input"
-      </n-flex>
+            type="textarea"
+            :autosize="{ minRows: 1, maxRows: 4 }"
+            placeholder="用你自己的话说说想做成什么样…"
+            :disabled="streaming"
+            style="flex: 1; min-width: 0"
+            @keydown.enter.exact.prevent="send"
+          />
+          <n-tooltip>
+            <template #trigger>
+              <n-button
+                secondary
+                :disabled="!input.trim() || !connected || !currentTaskId || streaming"
+                style="flex-shrink: 0"
+                @click="startGuidance"
+              >
+                <template #icon>
+                  <Icon icon="lucide:lightbulb" :width="16" />
+                </template>
+                帮我写清楚
+              </n-button>
+            </template>
+            AI 教练会问你几个问题，帮你把想要的效果描述清楚
+          </n-tooltip>
+        </n-flex>
+        <n-flex justify="space-between" align="center" style="margin-top: 8px">
+          <n-tooltip>
+            <template #trigger>
+              <n-button quaternary :disabled="!hasVersion || streaming" @click="openProblem">
+                <template #icon>
+                  <Icon icon="lucide:bug" :width="16" />
+                </template>
+                出问题了
+              </n-button>
+            </template>
+            网页哪里不对？说清楚做了什么、想看到什么、实际看到了什么
+          </n-tooltip>
+          <n-flex align="center" :size="8">
+            <n-select
+              v-model:value="selectedModel"
+              :options="modelOptions"
+              style="width: 150px"
+              :disabled="streaming"
+            />
+            <n-button v-if="streaming" type="error" @click="stopPrompt">
+              停止
+            </n-button>
+            <n-button
+              v-else
+              type="primary"
+              :disabled="!input.trim() || !connected"
+              @click="send"
+            >
+              发送
+            </n-button>
+          </n-flex>
+        </n-flex>
+      </template>
     </div>
   </div>
 </template>
 
 <script setup lang="ts">
-import { ref, watch, nextTick, computed, onUnmounted } from "vue"
+import { ref, reactive, watch, nextTick, computed, onUnmounted } from "vue"
 import { useStorage } from "@vueuse/core"
 import { useMessage } from "naive-ui"
 import { Icon } from "../../icons"
@@ -233,6 +275,46 @@ async function deletePair(assistantMsgId: number) {
 async function send() {
   const text = input.value.trim()
   if (!text || streaming.value) return
+  // 只说“不行”“没反应”，AI 只能瞎猜；提醒一次，学生选“直接发”以后这次就不再拦
+  if (!skipVagueCheck && hasVersion.value && [...text].length <= 15 && VAGUE.test(text)) {
+    vagueHint.value = true
+    return
+  }
+  input.value = ""
+  await props.beforeSend?.()
+  sendPrompt(text, selectedModel.value)
+}
+
+// “出问题了”：做了什么、想看到什么、实际看到了什么，拼成一句发给 AI
+const VAGUE =
+  /不行|不对|没反应|没用|没效果|用不了|不能用|不能玩|动不了|进不去|打不开|没显示|不显示|显示不出|bug|出错|报错|错了|坏了|白屏|空白|有问题/i
+const vagueHint = ref(false)
+let skipVagueCheck = false
+const problemOpen = ref(false)
+const problem = reactive({ did: "", want: "", saw: "" })
+
+watch(input, () => (vagueHint.value = false))
+
+const hasVersion = computed(() => messages.value.some((m) => m.role === "assistant" && hasCode(m)))
+
+function openProblem() {
+  vagueHint.value = false
+  problemOpen.value = true
+}
+
+function sendAnyway() {
+  skipVagueCheck = true
+  vagueHint.value = false
+  send()
+}
+
+async function sendProblem() {
+  const part = (s: string) => s.trim().replace(/[。，,.！!；;]+$/, "")
+  const did = part(problem.did).replace(/^我/, "")
+  const want = part(problem.want)
+  const text = `我${did}，${want ? `想看到${want}，` : ""}但实际${part(problem.saw)}。`
+  Object.assign(problem, { did: "", want: "", saw: "" })
+  problemOpen.value = false
   input.value = ""
   await props.beforeSend?.()
   sendPrompt(text, selectedModel.value)
@@ -447,6 +529,31 @@ watch([() => messages.value.length, streamingContent], () => {
   animation: pulse 1.5s ease-in-out infinite;
 }
 
+.vague-hint {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+  margin-bottom: 8px;
+  padding: 6px 10px;
+  border-radius: 6px;
+  background: var(--wp-brand-soft);
+  font-size: 13px;
+}
+
+.problem-title {
+  margin-bottom: 4px;
+  font-size: 13px;
+  font-weight: 600;
+}
+
+.problem label {
+  display: block;
+  margin: 8px 0 4px;
+  font-size: 12px;
+  color: var(--wp-muted);
+}
+
 .input-area {
   flex-shrink: 0;
   padding: 12px;

client/src/components/review/ReviewPanel.vue

@@ -119,6 +119,11 @@
           </n-popconfirm>
         </section>
 
+        <section v-if="sub.reflection">
+          <h4>学生说</h4>
+          <div class="reflection">{{ sub.reflection }}</div>
+        </section>
+
         <section>
           <n-flex justify="space-between" align="center" class="chain-head">
             <h4>提示词 <span class="muted" v-if="chainSteps">· {{ chainSteps }} 轮</span></h4>
@@ -479,6 +484,17 @@ h4 {
   flex-wrap: wrap;
 }
 
+/* 交的时候写的反思 */
+.reflection {
+  padding: 6px 10px;
+  border-left: 3px solid var(--wp-brand);
+  background: var(--wp-brand-soft);
+  font-size: 13px;
+  line-height: 1.6;
+  white-space: pre-wrap;
+  word-break: break-word;
+}
+
 .chain {
   margin: 0;
   padding-left: 20px;

client/src/components/workbench/PasteCodeModal.vue

@@ -5,12 +5,16 @@
     title="粘贴外部 AI 生成的代码"
     style="width: 640px; max-width: 94vw"
   >
-    <n-form-item label="你在外部 AI 里用的提示词" :show-feedback="false">
+    <n-form-item
+      label="你在外部 AI 里说了什么"
+      :validation-status="promptTooShort ? 'warning' : undefined"
+      :feedback="promptTooShort ? `再写具体一点：你让 AI 做了什么、改了什么（至少 ${MIN_PROMPT} 个字）` : ''"
+    >
       <n-input
         v-model:value="prompt"
         type="textarea"
         :autosize="{ minRows: 2, maxRows: 6 }"
-        placeholder="把你发给豆包、DeepSeek 等 AI 的话粘贴到这里"
+        placeholder="把你发给豆包、DeepSeek 等 AI 的话粘贴到这里；说了好几轮的，每一轮都贴进来，一轮一行"
       />
     </n-form-item>
     <n-form-item label="完整的网页代码" :show-feedback="false" style="margin-top: 12px">
@@ -30,7 +34,7 @@
         <n-button @click="show = false">取消</n-button>
         <n-button
           type="primary"
-          :disabled="!prompt.trim() || !raw.trim()"
+          :disabled="[...prompt.trim()].length < MIN_PROMPT || !raw.trim()"
           :loading="saving"
           @click="save"
         >
@@ -41,7 +45,7 @@
   </n-modal>
 </template>
 <script lang="ts" setup>
-import { ref } from "vue"
+import { computed, ref } from "vue"
 import { splitHtml, type Code } from "../../utils/code"
 
 const show = defineModel<boolean>("show", { required: true })
@@ -50,6 +54,12 @@ const props = defineProps<{
 }>()
 
 const prompt = ref("")
+// 以前四分之一的外部提示词只有“1”“倒计时”这样几个字，看不出学生想了什么
+const MIN_PROMPT = 8
+const promptTooShort = computed(() => {
+  const n = [...prompt.value.trim()].length
+  return n > 0 && n < MIN_PROMPT
+})
 const raw = ref("")
 const saving = ref(false)
 

client/src/components/workbench/SubmitModal.vue

@@ -0,0 +1,67 @@
+<template>
+  <n-modal
+    v-model:show="show"
+    preset="card"
+    :title="title"
+    style="width: 520px; max-width: 94vw"
+    :mask-closable="!saving"
+  >
+    <div class="question">
+      {{ hasEarlier ? "比起第一版，这一版你最满意的改动是什么？为什么这样改？" : "这一版你最满意的地方是什么？为什么？" }}
+    </div>
+    <n-input
+      v-model:value="reflection"
+      type="textarea"
+      :autosize="{ minRows: 3, maxRows: 6 }"
+      :maxlength="500"
+      placeholder="比如：我把按钮从五个减到三个，因为按钮太多了看不出先点哪个"
+    />
+    <n-text depth="3" class="hint">可以不写，直接交。写了的话，老师批改时能看到。</n-text>
+    <template #footer>
+      <n-flex justify="end">
+        <n-button :disabled="saving" @click="show = false">取消</n-button>
+        <n-button type="primary" :loading="saving" @click="save">交</n-button>
+      </n-flex>
+    </template>
+  </n-modal>
+</template>
+<script lang="ts" setup>
+import { ref } from "vue"
+
+// 交作品前问一句反思：最满意的改动和理由。不拦着交，只是让学生回头看一眼自己改了什么
+const show = defineModel<boolean>("show", { required: true })
+const props = defineProps<{
+  title: string
+  // 之前还有别的版本，就问“比起第一版”
+  hasEarlier: boolean
+  onSubmit: (reflection: string) => Promise<void>
+}>()
+
+const reflection = ref("")
+const saving = ref(false)
+
+async function save() {
+  saving.value = true
+  try {
+    await props.onSubmit(reflection.value.trim())
+    reflection.value = ""
+    show.value = false
+  } finally {
+    saving.value = false
+  }
+}
+</script>
+<style scoped>
+.question {
+  margin-bottom: 10px;
+  font-size: 14px;
+  line-height: 1.6;
+  color: var(--wp-ink);
+}
+
+.hint {
+  display: block;
+  margin-top: 8px;
+  font-size: 12px;
+}
+</style>

client/src/components/workbench/Workbench.vue

@@ -103,6 +103,12 @@
     </template>
   </n-split>
   <PasteCodeModal v-model:show="showPaste" :on-save="savePasted" />
+  <SubmitModal
+    v-model:show="showSubmit"
+    :title="submitTitle"
+    :has-earlier="submitHasEarlier"
+    :on-submit="submitChallenge"
+  />
   <n-drawer v-model:show="showHistory" :width="420" placement="right">
     <n-drawer-content title="全部历史" :native-scrollbar="false" body-content-style="padding: 0">
       <PromptHistoryPanel
@@ -128,6 +134,7 @@ import Preview from "../editor/Preview.vue"
 import CodeTabs from "./CodeTabs.vue"
 import VersionBar, { type VersionItem } from "./VersionBar.vue"
 import PasteCodeModal from "./PasteCodeModal.vue"
+import SubmitModal from "./SubmitModal.vue"
 import { Prompt, Submission } from "../../api"
 import { useTaskCode } from "../../composables/useTaskCode"
 import {
@@ -174,6 +181,7 @@ const selectedId = ref<number | null>(null)
 const busy = ref(false)
 const briefOpen = ref(true)
 const showPaste = ref(false)
+const showSubmit = ref(false)
 const showHistory = ref(false)
 const historyKey = ref(0)
 
@@ -308,21 +316,46 @@ function openHistory() {
 }
 
 async function submit() {
+  if (props.kind === "challenge") {
+    showSubmit.value = true
+    return
+  }
+  busy.value = true
+  try {
+    await props.saveWork?.({ ...code })
+    message.success("作品已保存")
+  } catch (err: any) {
+    message.error(err.response?.data?.detail ?? "保存失败，请重试")
+  } finally {
+    busy.value = false
+  }
+}
+
+const submitTitle = computed(() => {
+  if (modified.value || !selectedId.value) return "交当前修改"
+  const n = versions.value.find((v) => v.id === selectedId.value)?.n
+  return n ? `交 v${n}` : "交这一版"
+})
+
+// 要交的这一版之前还有版本（当前修改会存成新的一版）
+const submitHasEarlier = computed(() => {
+  if (modified.value || !selectedId.value) return versions.value.length > 0
+  return (versions.value.find((v) => v.id === selectedId.value)?.n ?? 1) > 1
+})
+
+// 挑战：先在弹框里写一句反思（可以不写）再交；出错时抛出去，弹框不关
+async function submitChallenge(reflection: string) {
   busy.value = true
   try {
-    if (props.kind === "creation") {
-      await props.saveWork?.({ ...code })
-      message.success("作品已保存")
-      return
-    }
     const id =
       modified.value || !selectedId.value ? await saveAsVersion() : selectedId.value
-    await Submission.create(props.taskId, { ...code }, id)
+    await Submission.create(props.taskId, { ...code }, id, reflection)
     markMessageSubmitted(id)
     const n = versions.value.find((v) => v.id === id)?.n
     message.success(n ? `已提交 v${n}` : "已提交")
   } catch (err: any) {
     message.error(err.response?.data?.detail ?? "提交失败，请重试")
+    throw err
   } finally {
     busy.value = false
   }

client/src/icons.ts

@@ -4,6 +4,7 @@ import {
   ArrowRight,
   Award,
   BookOpen,
+  Bug,
   ChartNoAxesColumn,
   Check,
   ChevronDown,
@@ -49,6 +50,7 @@ const registry: Record<string, Component> = {
   "arrow-right": ArrowRight,
   award: Award,
   "book-open": BookOpen,
+  bug: Bug,
   "bar-chart-2": ChartNoAxesColumn,
   check: Check,
   "chevron-down": ChevronDown,

client/src/utils/type.ts

@@ -167,6 +167,8 @@ export interface SubmissionAll {
   view_count: number
   created: Date
   modified: Date
+  // 交的时候写的反思，只有老师和作者本人拿得到
+  reflection?: string | null
 }
 
 export interface UserTag {

server/scripts/prompt-stats.ts

@@ -0,0 +1,109 @@
+/**
+ * 对比某个日期前后，学生和 AI 对话的情况（比如改了 AI 的回复方式以后有没有变化）。
+ * 日期按北京时间，分界日期那天算“之后”。只读，不改数据。
+ *
+ *   bun scripts/prompt-stats.ts <分界日期> [--from 日期] [--to 日期]
+ *   bun scripts/prompt-stats.ts 2026-10-28 --from 2026-09-01
+ *
+ * 层级是打标签时的标准；分类标准改过以后，先跑 classify-prompts.ts --force 重打，前后才好比。
+ */
+import { all, toDbTime } from "../src/db"
+
+const args = process.argv.slice(2)
+const opt = (name: string) => {
+  const i = args.indexOf(name)
+  return i >= 0 ? args[i + 1] : undefined
+}
+const split = args[0]
+if (!split || !/^\d{4}-\d{2}-\d{2}$/.test(split)) {
+  console.log("用法：bun scripts/prompt-stats.ts <分界日期> [--from 日期] [--to 日期]")
+  process.exit(1)
+}
+const day = (d: string) => toDbTime(new Date(`${d}T00:00:00+08:00`))
+const nextDay = (d: string) => toDbTime(new Date(new Date(`${d}T00:00:00+08:00`).getTime() + 86_400_000))
+const periods = [
+  { name: `${opt("--from") ?? "最早"} ~ ${split} 之前`, from: opt("--from") ? day(opt("--from")!) : "", to: day(split) },
+  { name: `${split} ~ ${opt("--to") ?? "现在"}`, from: day(split), to: opt("--to") ? nextDay(opt("--to")!) : "9999" },
+]
+
+// 和 PromptPanel 里一样：只说“不行”“没反应”的短抱怨；“出问题了”拼出来的句子
+const VAGUE =
+  /不行|不对|没反应|没用|没效果|用不了|不能用|不能玩|动不了|进不去|打不开|没显示|不显示|显示不出|bug|出错|报错|错了|坏了|白屏|空白|有问题/i
+const PROBLEM = /^我.+，(想看到.+，)?但实际.+。$/s
+
+type Msg = {
+  conversation_id: string
+  source: string
+  content: string
+  prompt_level: number | null
+  created: string
+  // 在这个对话的 AI 对话里是第几轮（只数 conversation）
+  round: number | null
+}
+
+// 批处理脚本，整表扫一遍没关系
+const msgs = all<Msg>(
+  `SELECT conversation_id, source, content, prompt_level, created,
+     CASE WHEN source = 'conversation'
+       THEN ROW_NUMBER() OVER (PARTITION BY conversation_id, source = 'conversation' ORDER BY created, id) END AS round
+   FROM prompt_message WHERE role = 'user'`,
+)
+const subs = all<{ created: string; reflected: number }>(
+  `SELECT s.created, r.submission_id IS NOT NULL AS reflected
+   FROM submission_submission s JOIN task_task t ON t.id = s.task_id
+   LEFT JOIN submission_reflection r ON r.submission_id = s.id
+   WHERE t.task_type = 'challenge'`,
+)
+
+const pct = (n: number, d: number) => (d ? `${((n / d) * 100).toFixed(1)}%` : "—")
+const chars = (s: string) => [...s.trim()].length
+
+function stats(from: string, to: string) {
+  const inRange = (t: string) => t >= from && t < to
+  const ms = msgs.filter((m) => inRange(m.created))
+  const byConv = new Map<string, Msg[]>()
+  for (const m of ms) byConv.set(m.conversation_id, [...(byConv.get(m.conversation_id) ?? []), m])
+
+  const convs = [...byConv.values()]
+  const ai = convs.filter((c) => c.some((m) => m.source === "conversation"))
+  const manualOnly = convs.filter((c) => c.some((m) => m.source === "manual") && !c.some((m) => m.source === "conversation"))
+  const rounds = ai.map((c) => c.filter((m) => m.source === "conversation").length)
+  const bucket = (lo: number, hi: number) => rounds.filter((n) => n >= lo && n <= hi).length
+
+  const asked = ms.filter((m) => m.source === "conversation")
+  const levels = [1, 2, 3, 4, 5, 6].map((l) => asked.filter((m) => m.prompt_level === l).length)
+  const leveled = levels.reduce((a, b) => a + b, 0)
+  const later = asked.filter((m) => (m.round ?? 1) >= 2 && m.prompt_level)
+  const vague = asked.filter((m) => chars(m.content) <= 15 && VAGUE.test(m.content))
+  const problem = asked.filter((m) => PROBLEM.test(m.content.trim()))
+  const manual = ms.filter((m) => m.source === "manual")
+  const s = subs.filter((x) => inRange(x.created))
+
+  return [
+    ["有对话或版本的作品", `${convs.length}`],
+    ["只粘贴外部代码", `${manualOnly.length}（${pct(manualOnly.length, convs.length)}）`],
+    ["和平台 AI 对话的作品", `${ai.length}`],
+    ["  平均轮数", ai.length ? (rounds.reduce((a, b) => a + b, 0) / ai.length).toFixed(2) : "—"],
+    ["  1 轮", pct(bucket(1, 1), ai.length)],
+    ["  2–3 轮", pct(bucket(2, 3), ai.length)],
+    ["  4–6 轮", pct(bucket(4, 6), ai.length)],
+    ["  7 轮以上", pct(bucket(7, Infinity), ai.length)],
+    ["提问（AI 对话）", `${asked.length}，已分级 ${leveled}`],
+    ...levels.map((n, i) => [`  L${i + 1}`, pct(n, leveled)]),
+    ["  L4–L6", pct(levels[3]! + levels[4]! + levels[5]!, leveled)],
+    ["  第 2 轮起的 L1", pct(later.filter((m) => m.prompt_level === 1).length, later.length)],
+    ["  只说“不行”的短抱怨", `${vague.length}（${pct(vague.length, asked.length)}）`],
+    ["  用“出问题了”", `${problem.length}（${pct(problem.length, asked.length)}）`],
+    ["用了“帮我写清楚”的作品", `${convs.filter((c) => c.some((m) => m.source === "guidance")).length}（占和 AI 对话的 ${pct(ai.filter((c) => c.some((m) => m.source === "guidance")).length, ai.length)}）`],
+    ["外部 AI 提示词", `${manual.length}，12 字以内 ${pct(manual.filter((m) => chars(m.content) <= 12).length, manual.length)}`],
+    ["挑战提交", `${s.length}，写了反思 ${pct(s.filter((x) => x.reflected).length, s.length)}`],
+  ]
+}
+
+const [a, b] = periods.map((p) => stats(p.from, p.to))
+// 终端里中文占两格
+const cols = (k: string) => [...k].reduce((n, ch) => n + (ch.codePointAt(0)! >= 0x2e80 ? 2 : 1), 0)
+const width = Math.max(...a!.map(([k]) => cols(k!))) + 2
+const pad = (k: string) => k + " ".repeat(width - cols(k))
+console.log(`${pad("")}${periods[0]!.name}  →  ${periods[1]!.name}`)
+for (const [i, [k, v]] of a!.entries()) console.log(`${pad(k!)}${v}  →  ${b![i]![1]}`)

server/src/db.ts

@@ -51,6 +51,9 @@ CREATE INDEX IF NOT EXISTS "prompt_message_prompt_level_e461b596" ON "prompt_mes
 CREATE INDEX IF NOT EXISTS "prompt_message_chain" ON "prompt_message" ("conversation_id", "created", "id", "role", "source", "prompt_level");
 CREATE INDEX IF NOT EXISTS "prompt_message_submission" ON "prompt_message" ("submission_id", "conversation_id", "created", "source") WHERE "submission_id" IS NOT NULL;
 
+-- 学生交作品时写的一句反思（最满意的改动和理由）；可以不写。删提交时跟着删
+CREATE TABLE IF NOT EXISTS "submission_reflection" ("submission_id" char(32) NOT NULL PRIMARY KEY REFERENCES "submission_submission" ("id") ON DELETE CASCADE, "content" text NOT NULL, "modified" datetime NOT NULL);
+
 -- 登录用 JWT：access token 不落库，refresh token 只存 sha256
 DROP TABLE IF EXISTS "auth_session";
 CREATE TABLE IF NOT EXISTS "auth_refresh_token" ("token_hash" text NOT NULL PRIMARY KEY, "user_id" bigint NOT NULL REFERENCES "account_user" ("id") ON DELETE CASCADE, "expire_at" datetime NOT NULL);

server/src/routes/submission.ts

@@ -120,6 +120,7 @@ router.post("/", async (c) => {
       js: z.string().nullable().optional(),
       prompt: z.string().nullable().optional(),
       message_id: z.number().int().nullable().optional(), // 关联的 assistant message
+      reflection: z.string().max(1000).nullable().optional(), // 交的时候写的一句反思，可以不写
     }),
   )
   const task = orNotFound(
@@ -131,6 +132,8 @@ router.post("/", async (c) => {
 
   const code = { html: p.html ?? null, css: p.css ?? null, js: p.js ?? null }
 
+  const reflection = p.reflection?.trim() || null
+
   const { submissionId, newUserMsgId } = tx(() => {
     let manualAsstMsgId: number | null = null
     let linkedMsg: MessageRow | null = null
@@ -203,6 +206,15 @@ router.post("/", async (c) => {
     return { submissionId: id, newUserMsgId }
   })
 
+  // 同一版再交时没写就留着上次写的
+  if (reflection) {
+    run(
+      `INSERT INTO submission_reflection (submission_id, content, modified) VALUES ($id, $content, $ts)
+       ON CONFLICT (submission_id) DO UPDATE SET content = excluded.content, modified = excluded.modified`,
+      { id: submissionId, content: reflection, ts: now() },
+    )
+  }
+
   // 只对新增的提示词做层级分类
   if (
     newUserMsgId !== null &&
@@ -540,7 +552,13 @@ router.get("/:id", (c) => {
   const id = uuidParam(c, "id")
   requireView(user, id)
   const s = orNotFound(get(`${SUBMISSION_SELECT} ${SUBMISSION_FROM} WHERE s.id = $id`, { me: user.id, id }))
-  const out = submissionOut(s)
+  // 反思只给老师和作者本人看
+  const reflection =
+    isStaff(user) || s.user_id === user.id
+      ? (get<{ content: string }>(`SELECT content FROM submission_reflection WHERE submission_id = $id`, { id })
+          ?.content ?? null)
+      : null
+  const out = { ...submissionOut(s), reflection }
   // 和同学作品一样：学生给同学打分之前不告诉作者和得分（作品展里的本来就署名）
   if (!isStaff(user) && s.user_id !== user.id && !out.my_score && !inShowcase(id)) {
     return c.json({ ...out, userid: null, username: null, score: null })

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
 }

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
+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. 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. 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 +39,36 @@ 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。
+// 挑战题目要求只当对照清单：学生没说的不替他做，挑一项放进“想一想”问他打算怎么做，
+// 要做成什么样还是学生自己说；整段贴了题目的照做（不拦，拦了只会去用外部 AI）
+export const BRIEF_RULES = `下面是这道题的题目要求。它是对照清单，不是让你照着做——这是课堂练习，题目要求要学生自己一步步提出来，这正是这节课要练的：
+- 只做学生在对话里说了的内容，做出来要能打开、能用。学生说的话里包含了题目要求（比如整段贴了题目），就照着做
+- 题目要求里学生还没提到的功能、规则和细节（如计时方式、对错的颜色、计分、重来按钮），哪怕很常见、顺手就能做，也先不做；也不要照搬题目里的类名、变量名和数据
+- 从学生没提到的要求里挑一项最重要的放进“想一想”，问他打算怎么做（如“题目还要求标出打错的字，你希望打错的时候是什么样子？”）
+- 学生要的和题目要求冲突时，按学生说的做，在“想一想”里提醒他题目是怎么要求的
+- 题目要求都做到了，“想一想”就问他自己的想法，不再提题目`
 
-判断标准——满足以下条件视为"够好"：
+// 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
-- 有明确主题（例如：登录页、计时器、商品卡片列表）
+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。
-- 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"
+- 只有作品名称，如"做一个象棋游戏"
-- 只有主题，完全没有任何视觉、交互或内容描述
+- 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
+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. 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
+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. 用中文回复，语气鼓励，简洁明了
+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. 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
+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. 如果回复以 [READY] 开头，[READY] 不要加粗，必须保持原始文本
+5. 用中文回复，语气鼓励，简洁明了
-6. 不要生成任何代码`
+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 +189,15 @@ 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。
+// 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
+// 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```") }
@@ -192,18 +217,19 @@ 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
+// 生成整页代码：开着思考，质量优先。brief 是挑战的题目要求（见 challengeBrief），自由创作没有
-export function streamChat(history: ChatMessage[], model?: string, signal?: AbortSignal) {
+export function streamChat(history: ChatMessage[], model?: string, signal?: AbortSignal, brief?: string | null) {
-  return streamCompletion([{ role: "system", content: SYSTEM_PROMPT }, ...trimHistory(history)], {
+  const system = brief ? `${SYSTEM_PROMPT}\n\n${BRIEF_RULES}\n\n${brief}` : SYSTEM_PROMPT
+  return streamCompletion([{ role: "system", content: system }, ...trimHistory(history)], {
     model: chatModel(model),
     thinking: true,
     signal,
   })
 }
 
-// 引导只是简短追问，不用思考，回得快
+// 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[], signal?: AbortSignal) {
+export function streamGuidance(history: ChatMessage[], context: string, signal?: AbortSignal) {
-  return streamCompletion([{ role: "system", content: GUIDANCE_SYSTEM_PROMPT }, ...history], {
+  return streamCompletion([{ role: "system", content: `${GUIDANCE_SYSTEM_PROMPT}\n\n${context}` }, ...history], {
     thinking: false,
     signal,
   })

server/src/services/prompt.ts

@@ -99,6 +99,62 @@ 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, max = BRIEF_CHARS) {
+  const text = content
+    .replace(/!\[[^\]]*\]\([^)]*\)/g, "")
+    .replace(/\s+/g, " ")
+    .trim()
+  return clipTo(text, max)
+}
+
+function clipTo(text: string, n: number) {
+  return text.length > n ? text.slice(0, n) + "……" : text
+}
+
+// 挑战的题目要求，给生成网页的 AI 当对照清单（见 llm.ts 的 BRIEF_RULES）；要逐条对照，不能截得太短。
+// 自由创作的题目是学生自己写的介绍，不用对照，返回 null
+export function challengeBrief(taskId: number) {
+  const task = get<{ title: string; content: string }>(
+    `SELECT title, content FROM task_task WHERE id = $id AND task_type = 'challenge'`,
+    { id: taskId },
+  )
+  const brief = task && briefOf(task.content, 4000)
+  return brief ? `题目：${task.title}\n\n题目要求：${brief}` : null
+}
+
+// “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

server/src/ws.ts

@@ -3,8 +3,10 @@ import { all, get, run, iso, uuidOut } from "./db"
 import { userFromAccessToken, type User } from "./auth"
 import {
   canAccessTask,
+  challengeBrief,
   codeAsMarkdown,
   getOrCreateActiveConversation,
+  guidanceContext,
   insertMessage,
 } from "./services/prompt"
 import {
@@ -15,6 +17,7 @@ import {
   userFacingError,
   type ChatMessage,
 } from "./services/llm"
+import { classifyMessage, inBackground } from "./services/classifier"
 
 type Kind = "prompt" | "guidance"
 
@@ -116,11 +119,12 @@ async function handlePrompt(ws: WS, data: any) {
   const model = String(data.model ?? "")
   if (!prompt) return
 
-  ws.data.pendingUserMessageId = insertMessage({
+  const userMessageId = insertMessage({
     conversation_id: ws.data.conversationId,
     role: "user",
     content: prompt,
   })
+  ws.data.pendingUserMessageId = userMessageId
   try {
     const history: ChatMessage[] = conversationHistory(ws.data.conversationId).map((m) => ({
       role: m.role as ChatMessage["role"],
@@ -129,7 +133,8 @@ async function handlePrompt(ws: WS, data: any) {
     }))
     let full: string
     try {
-      full = await streamToClient(ws, streamChat(history, model, ws.data.abort.signal))
+      const brief = challengeBrief(ws.data.taskId)
+      full = await streamToClient(ws, streamChat(history, model, ws.data.abort.signal, brief))
     } catch (e) {
       // 客户端自己断开导致的中止不算出错
       if (!ws.data.closed) send(ws, { type: "error", content: userFacingError(e) })
@@ -146,8 +151,12 @@ async function handlePrompt(ws: WS, data: any) {
       code_css: code.css,
       code_js: code.js,
     })
+    // 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)
   }
@@ -157,17 +166,22 @@ async function handleGuidance(ws: WS, data: any) {
   const prompt = String(data.content ?? "").trim()
   if (!prompt) return
 
-  ws.data.pendingUserMessageId = insertMessage({
+  const userMessageId = insertMessage({
     conversation_id: ws.data.conversationId,
     role: "user",
     content: prompt,
     source: "guidance",
   })
+  ws.data.pendingUserMessageId = userMessageId
   ws.data.guidanceHistory.push({ role: "user", content: prompt })
   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()
@@ -185,6 +199,7 @@ async function handleGuidance(ws: WS, data: any) {
       content: clean,
       source: "guidance",
     })
+    inBackground(classifyMessage(userMessageId))
     ws.data.pendingUserMessageId = null
     send(ws, { type: "complete", message_id: messageId, is_ready: isReady })
   } finally {

server/tests/api.test.ts

@@ -276,6 +276,34 @@ describe("提交与删除", () => {
     expect(all(`SELECT COUNT(*) AS n FROM submission_submission WHERE id != $id`, { id: otherSub })).toEqual([{ n: 1 }])
   })
 
+  test("交的时候写的反思：再交没写就保留，只有老师和作者能看，删提交时一起删", async () => {
+    makeUser("stu")
+    makeUser("mate")
+    makeUser("teacher", "admin")
+    const task = makeTask("T")
+    const stu = await client("stu")
+    const edit = await stu("POST", "/api/prompt/versions", { task_id: task, html: "<p>1", css: "", js: "" })
+    const sub = await stu("POST", "/api/submission/", {
+      task_id: task,
+      html: "<p>1",
+      message_id: edit.json.message_id,
+      reflection: "  按钮改成三个，因为五个太乱  ",
+    })
+    const id = sub.json.id
+    await stu("POST", "/api/submission/", { task_id: task, html: "<p>1!", message_id: edit.json.message_id, reflection: "" })
+    expect((await stu("GET", `/api/submission/${id}`)).json.reflection).toBe("按钮改成三个，因为五个太乱")
+    expect((await (await client("teacher"))("GET", `/api/submission/${id}`)).json.reflection).toBe("按钮改成三个，因为五个太乱")
+
+    const mate = await client("mate")
+    await mate("POST", "/api/submission/", { task_id: task, html: "<p>m" })
+    const peer = await mate("GET", `/api/submission/${id}`)
+    expect(peer.status).toBe(200)
+    expect(peer.json.reflection).toBeNull()
+
+    expect((await stu("DELETE", `/api/submission/${id}`)).status).toBe(200)
+    expect(all(`SELECT * FROM submission_reflection`)).toEqual([])
+  })
+
   test("被评分的提交只有超管能删", async () => {
     const author = makeUser("author")
     const other = makeUser("other")

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 = () => {}

server/tests/ws.test.ts

@@ -2,18 +2,25 @@ 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```"]
-// 最近一次发给模型的请求
+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
 const llm = Bun.serve({
   port: 0,
   async fetch(req) {
-    lastRequest = await req.json()
+    const body: any = await req.json()
+    if (!body.stream) {
+      lastClassify = body
+      return Response.json({ choices: [{ message: { content: "3" } }] })
+    }
+    lastRequest = body
     const enc = new TextEncoder()
     const stream = new ReadableStream({
       async start(ctrl) {
@@ -50,7 +57,10 @@ afterAll(() => {
   llm.stop(true)
 })
 
-beforeEach(freshDb)
+beforeEach(() => {
+  freshDb()
+  PIECES = DEFAULT_PIECES
+})
 
 interface Conn {
   ws: WebSocket
@@ -106,6 +116,19 @@ describe("WebSocket 代码生成", () => {
     c.ws.close()
   })
 
+  test("回复完后在后台给提问打层级", async () => {
+    const { task, token } = await setup()
+    const c = (await connect(`/ws/prompt/${task}/?token=${token}`))!
+    await c.next("init")
+    c.ws.send(JSON.stringify({ type: "message", content: "做个页面" }))
+    await c.next("complete")
+    const level = () => all(`SELECT prompt_level FROM prompt_message WHERE role = 'user'`)[0]?.prompt_level
+    for (let i = 0; i < 50 && level() == null; i++) await Bun.sleep(20)
+    expect(level()).toBe(3)
+    expect(lastClassify.thinking).toEqual({ type: "disabled" })
+    c.ws.close()
+  })
+
   test("AI 回复中再发消息会被拒绝", async () => {
     const { task, token } = await setup()
     const c = (await connect(`/ws/prompt/${task}/?token=${token}`))!
@@ -157,9 +180,70 @@ 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("挑战的题目要求作为对照清单发给 AI，自由创作不发", async () => {
+    const { task, token } = await setup()
+    run(`UPDATE task_task SET content = $c WHERE id = $id`, { c: "做一个计时器，有开始和暂停按钮", id: task })
+    const c = (await connect(`/ws/prompt/${task}/?token=${token}`))!
+    await c.next("init")
+    c.ws.send(JSON.stringify({ type: "message", content: "做个页面" }))
+    await c.next("complete")
+    expect(lastRequest.messages[0].content).toContain("它是对照清单")
+    expect(lastRequest.messages[0].content).toContain("题目要求：做一个计时器，有开始和暂停按钮")
+    c.ws.close()
+
+    const creation = makeTask("我的作品", "creation", 2)
+    run(`UPDATE task_task SET content = '我想做一个相册' WHERE id = $id`, { id: creation })
+    const d = (await connect(`/ws/prompt/${creation}/?token=${token}`))!
+    await d.next("init")
+    d.ws.send(JSON.stringify({ type: "message", content: "做个页面" }))
+    await d.next("complete")
+    expect(lastRequest.messages[0].content).not.toContain("题目要求")
+    d.ws.close()
+  })
+
   test("token 无效或任务不存在时握手失败", async () => {
     const { task, token } = await setup()
     expect(await connect(`/ws/prompt/${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()
+  })
+})