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

@@ -19,7 +19,7 @@ The backend was rewritten from Django in 2026-10 to cut memory use (~25 MB vs. 1
   - 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 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)
@@ -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/`.