refactor: 精简发给模型的上下文、隐藏接口原始报错，合并重复的删除逻辑

- 发给模型的对话只留最近 20 条，除最后一条 AI 回复外旧回复的代码块换成占位文字，
  对话越长越省 token，也不会超出上下文长度
- 学生看到的 AI 报错只显示超时 / 繁忙 / 出错（状态码），完整错误写日志；
  客户端自己断开导致的中止不再记错误
- 管理员查询对话必须带 task_id 或 user_id，不再一次查出全站所有对话
- “已评分或获奖不能删”和“删除一轮对话”合并到 services/cascade.ts，补上删除一轮对话的测试

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

server/src/services/cascade.ts

@@ -1,4 +1,5 @@
-import { all, run } from "../db"
+import { all, get, run } from "../db"
+import { getPrecedingUserMessage, type MessageRow } from "./prompt"
 
 // Django 的 on_delete=CASCADE 是在应用层做的，这里手动按依赖顺序删除。
 // 调用方负责包在事务里。
@@ -15,6 +16,22 @@ export function deleteSubmissions(where: string, params: Record<string, unknown>
   run(`DELETE FROM submission_submission WHERE ${where}`, params)
 }
 
+// 被评过分或获过奖的提交只有超管能删
+export const hasRatingsOrAwards = (submissionId: string) =>
+  !!get(
+    `SELECT 1 WHERE EXISTS (SELECT 1 FROM submission_rating WHERE submission_id = $id)
+       OR EXISTS (SELECT 1 FROM submission_submissionaward WHERE submission_id = $id)`,
+    { id: submissionId },
+  )
+
+// 删除一轮对话：AI 回复、它前面那条用户提问，以及回复关联的提交
+export function deleteRound(asst: MessageRow) {
+  const userMsg = getPrecedingUserMessage(asst)
+  if (userMsg) run(`DELETE FROM prompt_message WHERE id = $id`, { id: userMsg.id })
+  run(`DELETE FROM prompt_message WHERE id = $id`, { id: asst.id })
+  if (asst.submission_id) deleteSubmissions("id = $id", { id: asst.submission_id })
+}
+
 export function deleteTask(taskId: number) {
   const p = { task_id: taskId }
   deleteSubmissions("task_id = $task_id", p)

server/src/services/llm.ts

@@ -158,9 +158,36 @@ 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
+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")
+  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```") }
+      : m,
+  )
+}
+
+// 给学生看的错误：不把模型接口的原始返回（可能带接口细节）原样发出去，完整信息写日志
+export function userFacingError(e: unknown) {
+  console.error("LLM request failed:", e)
+  const message = e instanceof Error ? e.message : String(e)
+  if (message === "模型响应超时" || (e instanceof Error && e.name === "TimeoutError")) {
+    return "AI 响应超时，请重试"
+  }
+  const status = /^Error code: (\d+)/.exec(message)?.[1]
+  if (status === "429") return "AI 服务繁忙，请稍后重试"
+  return status ? `AI 服务出错（${status}），请稍后重试` : "AI 服务出错，请稍后重试"
+}
+
 export function streamChat(history: ChatMessage[], model?: string, signal?: AbortSignal) {
   return streamCompletion(
-    [{ role: "system", content: SYSTEM_PROMPT }, ...history],
+    [{ role: "system", content: SYSTEM_PROMPT }, ...trimHistory(history)],
     chatModel(model),
     signal,
   )