| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193 |
- """服务端的训练计划草稿缓存(对应需求清单 BUG-05)。
- **要解决的问题**
- "保存刚才的训练计划"以前是从聊天历史里把计划正文捞回来的
- (`most_recent_complete_training_plan(history)`)。但那份 history 已经被
- `context_budget._normalize_history` 做过两级截断:
- * 单条上限 `HISTORY_ITEM_MAX_CHARS = 4000`,超出会追加 `[历史消息已截断]`;
- * 总量上限 `HISTORY_TOTAL_MAX_CHARS = 12000`,超出会追加 `[历史上下文预算已用尽]`。
- 一份正常的训练计划轻松超过 4000 字符,而完整性判断只要求 ≥300 字符,
- 于是**被截断的残缺计划照样通过检查并原样落盘**;`plan_store` 又按
- content_hash 去重,用户重存一次也修不回来。
- 同一条路径上还有两个副作用:
- * `HISTORY_CONTEXT_MAX_ITEMS = 8`,生成计划后再聊 8 轮,它就彻底不在
- history 里了;
- * 前端的 `sessionConversationHistory` 是内存变量,刷新页面或新开标签页
- 即归零,"保存刚才的计划"直接 400。
- **做法**
- 生成计划时就把**完整正文**在服务端存一份,保存时按 id 取原文,不再回捞
- 聊天历史。history 只作为进程重启后的兜底。
- 这是一个进程内的短期缓存,不做持久化:它服务的是"刚生成完、马上保存"
- 这个窗口,重启后回退到原有行为即可。单用户本地应用,因此不区分会话。
- """
- from __future__ import annotations
- import threading
- from dataclasses import dataclass, field
- from datetime import datetime, timedelta, timezone
- from typing import Callable
- from uuid import uuid4
- DEFAULT_MAX_ITEMS = 20
- DEFAULT_TTL_SECONDS = 6 * 60 * 60
- @dataclass(frozen=True)
- class PlanDraft:
- id: str
- content: str
- subject: str
- title: str
- suggested_date: str
- created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
- class PlanDraftCache:
- """进程内、有上限、带 TTL 的计划草稿缓存。"""
- def __init__(
- self,
- *,
- max_items: int = DEFAULT_MAX_ITEMS,
- ttl_seconds: int = DEFAULT_TTL_SECONDS,
- ) -> None:
- self._max_items = max_items
- self._ttl = timedelta(seconds=ttl_seconds)
- self._drafts: list[PlanDraft] = []
- self._lock = threading.Lock()
- # ── 内部 ──────────────────────────────────────────────────────────
- def _prune(self, now: datetime) -> None:
- cutoff = now - self._ttl
- self._drafts = [draft for draft in self._drafts if draft.created_at > cutoff]
- if len(self._drafts) > self._max_items:
- self._drafts = self._drafts[-self._max_items:]
- # ── 对外 ──────────────────────────────────────────────────────────
- def remember(
- self,
- *,
- content: str,
- subject: str = "",
- title: str = "",
- suggested_date: str = "",
- now: datetime | None = None,
- ) -> PlanDraft:
- """记住一份刚生成的完整计划,返回带 id 的草稿。"""
- now = now or datetime.now(timezone.utc)
- draft = PlanDraft(
- id=uuid4().hex,
- content=content,
- subject=subject,
- title=title,
- suggested_date=suggested_date,
- created_at=now,
- )
- with self._lock:
- self._drafts.append(draft)
- self._prune(now)
- return draft
- def get(self, draft_id: str | None, *, now: datetime | None = None) -> PlanDraft | None:
- if not draft_id:
- return None
- now = now or datetime.now(timezone.utc)
- with self._lock:
- self._prune(now)
- for draft in reversed(self._drafts):
- if draft.id == draft_id:
- return draft
- return None
- def latest(self, *, now: datetime | None = None) -> PlanDraft | None:
- """最近一次生成且未过期的计划。"""
- now = now or datetime.now(timezone.utc)
- with self._lock:
- self._prune(now)
- return self._drafts[-1] if self._drafts else None
- def find(
- self,
- *,
- subject: str = "",
- suggested_date: str = "",
- now: datetime | None = None,
- ) -> PlanDraft | None:
- """按科目/日期挑一份草稿,取不到再回落到最近一份。
- 一个会话里可能生成过多份计划(先练背、又改练腿),无条件用
- `latest()` 会在用户说"保存那份腿的计划"时静默存错。这里按线索
- 从新到旧找:科目+日期都对上最优先,其次任一对上,最后才是最近一份。
- """
- now = now or datetime.now(timezone.utc)
- wanted_subject = (subject or "").strip()
- wanted_date = (suggested_date or "").strip()
- with self._lock:
- self._prune(now)
- candidates = list(reversed(self._drafts))
- if not candidates:
- return None
- if wanted_subject and wanted_date:
- for draft in candidates:
- if draft.subject == wanted_subject and draft.suggested_date == wanted_date:
- return draft
- if wanted_subject:
- for draft in candidates:
- if draft.subject == wanted_subject:
- return draft
- if wanted_date:
- for draft in candidates:
- if draft.suggested_date == wanted_date:
- return draft
- return candidates[0]
- def clear(self) -> None:
- with self._lock:
- self._drafts.clear()
- def __len__(self) -> int:
- with self._lock:
- return len(self._drafts)
- TRUNCATED_CONTENT_ERROR = (
- "这份计划正文是被上下文长度限制截断过的残缺内容,已拒绝保存。"
- "请让 Agent 重新生成完整计划后再保存。"
- )
- def resolve_plan_save_content(
- payload: dict,
- cache: PlanDraftCache,
- *,
- is_truncated: Callable[[str], bool],
- ) -> tuple[str, str | None]:
- """决定保存计划时用哪份正文,返回 `(content, error)`。
- 抽成纯函数是为了让这段判断能被真正调用着测——原先它内联在 `/plans`
- 路由里,测试只能对 main.py 做字符串断言,改个变量名测试就绿着失效。
- 截断判断由调用方注入(而不是在这里 import context_budget),这样按文件
- 路径单独加载本模块时也不需要包上下文(对照 ARCH-02)。
- 顺序很关键(BUG-05):**先取服务端草稿**。draft_id 的设计目的就是绕过
- 任何中途加工,若先校验前端 content,一个带合法 draft_id 的请求会因为
- 前端那份恰好带了截断标记而先被拒,服务端的完整原文根本没机会生效。
- 只有回落到前端正文时才需要截断拦截——`plan_store` 按 content_hash 去重,
- 残缺版本一旦落盘,用户重存一次也覆盖不掉。
- """
- draft = cache.get(str(payload.get("draft_id") or ""))
- if draft is not None and draft.content:
- return draft.content, None
- content = str(payload.get("content") or "")
- if is_truncated(content):
- return "", TRUNCATED_CONTENT_ERROR
- return content, None
|