plan_draft_cache.py 7.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193
  1. """服务端的训练计划草稿缓存(对应需求清单 BUG-05)。
  2. **要解决的问题**
  3. "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
  4. (`most_recent_complete_training_plan(history)`)。但那份 history 已经被
  5. `context_budget._normalize_history` 做过两级截断:
  6. * 单条上限 `HISTORY_ITEM_MAX_CHARS = 4000`,超出会追加 `[历史消息已截断]`;
  7. * 总量上限 `HISTORY_TOTAL_MAX_CHARS = 12000`,超出会追加 `[历史上下文预算已用尽]`。
  8. 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,
  9. 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
  10. content_hash 去重,用户重存一次也修不回来。
  11. 同一条路径上还有两个副作用:
  12. * `HISTORY_CONTEXT_MAX_ITEMS = 8`,生成计划后再聊 8 轮,它就彻底不在
  13. history 里了;
  14. * 前端的 `sessionConversationHistory` 是内存变量,刷新页面或新开标签页
  15. 即归零,"保存刚才的计划"直接 400。
  16. **做法**
  17. 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
  18. 聊天历史。history 只作为进程重启后的兜底。
  19. 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"
  20. 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。
  21. """
  22. from __future__ import annotations
  23. import threading
  24. from dataclasses import dataclass, field
  25. from datetime import datetime, timedelta, timezone
  26. from typing import Callable
  27. from uuid import uuid4
  28. DEFAULT_MAX_ITEMS = 20
  29. DEFAULT_TTL_SECONDS = 6 * 60 * 60
  30. @dataclass(frozen=True)
  31. class PlanDraft:
  32. id: str
  33. content: str
  34. subject: str
  35. title: str
  36. suggested_date: str
  37. created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
  38. class PlanDraftCache:
  39. """进程内、有上限、带 TTL 的计划草稿缓存。"""
  40. def __init__(
  41. self,
  42. *,
  43. max_items: int = DEFAULT_MAX_ITEMS,
  44. ttl_seconds: int = DEFAULT_TTL_SECONDS,
  45. ) -> None:
  46. self._max_items = max_items
  47. self._ttl = timedelta(seconds=ttl_seconds)
  48. self._drafts: list[PlanDraft] = []
  49. self._lock = threading.Lock()
  50. # ── 内部 ──────────────────────────────────────────────────────────
  51. def _prune(self, now: datetime) -> None:
  52. cutoff = now - self._ttl
  53. self._drafts = [draft for draft in self._drafts if draft.created_at > cutoff]
  54. if len(self._drafts) > self._max_items:
  55. self._drafts = self._drafts[-self._max_items:]
  56. # ── 对外 ──────────────────────────────────────────────────────────
  57. def remember(
  58. self,
  59. *,
  60. content: str,
  61. subject: str = "",
  62. title: str = "",
  63. suggested_date: str = "",
  64. now: datetime | None = None,
  65. ) -> PlanDraft:
  66. """记住一份刚生成的完整计划,返回带 id 的草稿。"""
  67. now = now or datetime.now(timezone.utc)
  68. draft = PlanDraft(
  69. id=uuid4().hex,
  70. content=content,
  71. subject=subject,
  72. title=title,
  73. suggested_date=suggested_date,
  74. created_at=now,
  75. )
  76. with self._lock:
  77. self._drafts.append(draft)
  78. self._prune(now)
  79. return draft
  80. def get(self, draft_id: str | None, *, now: datetime | None = None) -> PlanDraft | None:
  81. if not draft_id:
  82. return None
  83. now = now or datetime.now(timezone.utc)
  84. with self._lock:
  85. self._prune(now)
  86. for draft in reversed(self._drafts):
  87. if draft.id == draft_id:
  88. return draft
  89. return None
  90. def latest(self, *, now: datetime | None = None) -> PlanDraft | None:
  91. """最近一次生成且未过期的计划。"""
  92. now = now or datetime.now(timezone.utc)
  93. with self._lock:
  94. self._prune(now)
  95. return self._drafts[-1] if self._drafts else None
  96. def find(
  97. self,
  98. *,
  99. subject: str = "",
  100. suggested_date: str = "",
  101. now: datetime | None = None,
  102. ) -> PlanDraft | None:
  103. """按科目/日期挑一份草稿,取不到再回落到最近一份。
  104. 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
  105. `latest()` 会在用户说"保存那份腿的计划"时静默存错。这里按线索
  106. 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。
  107. """
  108. now = now or datetime.now(timezone.utc)
  109. wanted_subject = (subject or "").strip()
  110. wanted_date = (suggested_date or "").strip()
  111. with self._lock:
  112. self._prune(now)
  113. candidates = list(reversed(self._drafts))
  114. if not candidates:
  115. return None
  116. if wanted_subject and wanted_date:
  117. for draft in candidates:
  118. if draft.subject == wanted_subject and draft.suggested_date == wanted_date:
  119. return draft
  120. if wanted_subject:
  121. for draft in candidates:
  122. if draft.subject == wanted_subject:
  123. return draft
  124. if wanted_date:
  125. for draft in candidates:
  126. if draft.suggested_date == wanted_date:
  127. return draft
  128. return candidates[0]
  129. def clear(self) -> None:
  130. with self._lock:
  131. self._drafts.clear()
  132. def __len__(self) -> int:
  133. with self._lock:
  134. return len(self._drafts)
  135. TRUNCATED_CONTENT_ERROR = (
  136. "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。"
  137. "请让 Agent 重新生成完整计划后再保存。"
  138. )
  139. def resolve_plan_save_content(
  140. payload: dict,
  141. cache: PlanDraftCache,
  142. *,
  143. is_truncated: Callable[[str], bool],
  144. ) -> tuple[str, str | None]:
  145. """决定保存计划时用哪份正文,返回 `(content, error)`。
  146. 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`
  147. 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。
  148. 截断判断由调用方注入(而不是在这里 import context_budget),这样按文件
  149. 路径单独加载本模块时也不需要包上下文(对照 ARCH-02)。
  150. 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
  151. 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
  152. 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。
  153. 只有回落到前端正文时才需要截断拦截——`plan_store` 按 content_hash 去重,
  154. 残缺版本一旦落盘,用户重存一次也覆盖不掉。
  155. """
  156. draft = cache.get(str(payload.get("draft_id") or ""))
  157. if draft is not None and draft.content:
  158. return draft.content, None
  159. content = str(payload.get("content") or "")
  160. if is_truncated(content):
  161. return "", TRUNCATED_CONTENT_ERROR
  162. return content, None