diff --git a/PROGRESS.md b/PROGRESS.md index 7c28246..e5ed16f 100644 --- a/PROGRESS.md +++ b/PROGRESS.md @@ -8,7 +8,7 @@ --- -## 进行中：AI 对话（聊天记录）持久化 · develop 上新功能 +## ✅ 已交付：AI 对话（聊天记录）持久化（AC-1..AC-4 全绿）· develop 上新功能 > **背景**：[2026-07-09] 用户要求把作者↔AI 每轮往复（refine/rewrite/clarify/continue/工具箱生成器）**落库为聊天记录**，每章一个「AI 对话」抽屉展示。多 agent（4 学科起草→4 视角交叉评审→合稿）产出实施计划 `docs/design/ai-chat-history-plan.md`。**决策全 A**：clarify buffer-until-concluded / v1 只记 5 类交互 / append-only 不删。 > **契约与不变量纪律**：@db 建表(契约 C2 扩)→@backend 端点(C3 扩)→@frontend gen:api+接线→@qa E2E，逐级等上游契约「稳定」。守 #1/#3/#8：聊天记录是**旁路留痕**（同 `usage_ledger` 物种），前端**显式 append**、生成端点**保持只读不变**、**绝不**喂 `assemble`/prompt（计划 §8 有围栏 + code-review 守卫）。 @@ -18,7 +18,7 @@ | AC-1 `ai_messages` 表 + 迁移 | ✅ @db 2026-07-09 | @db | — | 计划 §2 canonical DDL；迁移 `ceabd3c2c00d`(down=`f6a7b8c9d0e1`)；C2 扩已「稳定」，AC-2 可动工 | | AC-2 repo + append/list 端点 + schema + 契约 | ✅ @backend 2026-07-09 | @backend | AC-1 | `c675a74`；`SqlAiMessageRepo`(批 append seq/thread_id + newest-first union/分页) + `POST/GET /projects/{id}/ai-messages` + schemas；C3 扩已「稳定」→ AC-3 可动工（须 `pnpm gen:api`） | | AC-3 gen:api + AiConversationDrawer + useAiConversation + 接线 5 处交互 | ✅ @frontend 2026-07-09 | @frontend | AC-2 | `dce882c`/`ddf7fa0`/`59f2f60`/`33a9b76`；gen:api + 纯视图模型 + hook + 抽屉 + 五类交互留痕（clarify buffer-until-concluded）；vitest 706 passed / cov 95% / build OK | -| AC-4 E2E + repo/端点/hook 测试 | ⬜ | @qa | AC-3 | 计划 §6；剩 E2E 6 流程（含 clarify-abandon 不落库、refine 漂移不盲替换、项目级现于本章抽屉） | +| AC-4 E2E + #1/#6 围栏守卫 | ✅ @qa 2026-07-09 | @qa | AC-3 | 计划 §6/§8；`tests/test_ai_chat_history_e2e.py`（6 用例，真 pg 零 LLM）+ `tests/test_ai_messages_not_in_generation_path.py`（AST 围栏，能变红）；pytest **1002 passed**（+10）。**AI 对话（聊天记录）AC-1..AC-4 全交付** | --- @@ -208,6 +208,7 @@ T0.1 monorepo 骨架 ✅ @devops · T0.2 16 MVP 表迁移（无漂移，users st > 格式：`- [YYYY-MM-DD] @skill 完成/进展 Txx — 一句话结果 + 影响的契约/文件` +- [2026-07-09] @qa — **✅ AC-4 E2E + #1/#6 真源围栏守卫｜AI 对话（聊天记录）AC-1..AC-4 全交付**（计划 §6/§8）。**E2E**（`tests/test_ai_chat_history_e2e.py`，真 pg、**零 LLM 无网关**——纯 CRUD 侧记录，镜像 `test_t6_toolbox_e2e.py` 的 `e2e_sm` 生命周期）覆盖 §6 全 6 用例：① 缓冲多轮线程一批落库（作者含糊→AI 反问〔选项进 meta〕→作者答→AI 定稿，同 `thread_id`/`seq` 0..3）→ GET newest-first（`seq`/`role`/`thread_id`/`meta{options,version_no,segment}` 往复保真、全文 content 未截断）；② 作用域分区（章级 ∪ 项目级 `chapter_no IS NULL`——`?chapter_no=N` 两者都返回、缺省仅项目级、项目级现于任一章抽屉）；③ 跨多批 newest-first + `limit/offset` 分页 + `kind` 过滤；④ **append 只写 `ai_messages`**（UsageLedger/Chapter/ChapterReview/WorldEntity/Job 逐表零变化、`usage_ledger` 恒 0 证无网关）、GET 只读不建行；⑤ 未知项目 POST/GET 皆 404、非法 role/kind·空 messages·超上限 content/meta 五类 422 且失败零落库；⑥ clarify-abandon 代理（服务端无部分写路径，只持久化被显式 POST 的已定线程）。**围栏**（`tests/test_ai_messages_not_in_generation_path.py`，AST 静态守卫，**能变红**）：memory/orchestrator 两目录 23 文件无一 import/引用 `AiMessage*`/`ai_message_repo`（Name/Attribute/Import 全扫）+ `MemoryRepos` 捆绑与 `assemble` 形参都不含它 + 植入违规自证探测器非空——把 ai_message repo 接进 `assemble()` 会直接失败（已用真 assemble.py 源注入验证变红）。守 #1/#6：聊天记录是 append-only 旁路侧记录（同 `usage_ledger` 物种），绝不喂 prompt/记忆注入。**门禁全绿**：ruff/format 干净 · mypy **238 文件** · **`alembic check` 无漂移** · pytest **1002 passed**（992 基线 + 10：E2E 6 + 围栏 4）。gotcha 已登记（ai_messages 真源围栏 = code-review BLOCKER）。仅动 `tests/` + `memory/gotchas.md` + `PROGRESS.md`，**未碰任何产品代码**。 - [2026-07-09] @frontend — **✅ AC-3 gen:api + AI 对话 hook/抽屉 + 五类交互留痕**（计划 §4，`dce882c`/`ddf7fa0`/`59f2f60`/`33a9b76`）。**gen:api**（`dce882c`）：`POST/GET /projects/{id}/ai-messages` + `AiMessage*` 4 schema 入 `schema.d.ts`（openapi.json 为 gitignore 构建产物），`lib/api/types.ts` re-export 码生成名。**核心**（`ddf7fa0`）：`lib/workbench/aiConversation.ts` 纯视图模型（`kindLabel/kindBadge`、`bubbleSide`、`groupByThread` 按 `(created_at,seq)` 组内+线程定序〔**绝不单靠 created_at**——同批共享 created_at 靠 repo 的 seq 消歧〕、`partitionByScope`〔本章 ∪ 项目级〕、`acceptActionFor`〔rewrite→替换整章 / continue·generator→插入正文 / refine→回填选段 / clarify→无〕）+ 11 单测；`lib/workbench/useAiConversation.ts` Workbench 级单实例（懒加载去重 + `mergeById` 防并发覆盖 + 乐观插入/正位对账/失败回滚，**fail-soft 绝不 throw 进生成/HITL**〔守 #3 意图〕；`silent` 直投供无抽屉的工具箱页）+ 9 renderHook 测试；`components/workbench/AiConversationDrawer.tsx` 克隆 ContextDrawer 的 a11y（`role=dialog/log`、`aria-live` additions、Esc、focus trap、还原焦点、prefers-reduced-motion），本章对话 + 项目级生成两段气泡、clarify 只读选项芯片、ai 气泡挂 DRAFT-only 再接受。**hook 返回产物**（`59f2f60`）：四生成 hook 加返回值（`RefineVersion|null`/`string|null`/`RewriteVersion|null`/`GeneratorArtifact|null`，向后兼容，补成功用例断言）。**五处接线**（`33a9b76`）：RefinePanel/ChapterRewritePanel 结束后 append `[author(意见/答复), ai(产出)]`，**clarify buffer-until-concluded**——澄清 Q&A（原含糊意见→AI 反问〔选项进 meta〕→作者答复）缓冲，随最终产出批一起落库同 `thread_id`，**纯放弃不记**；refine ai 行 meta 存 `version_no`+原选段（每版重锚）；ContinuePanel ai-only（续写无作者输入）meta `candidate_index`；GeneratorRunner append `[author(字段摘要,meta{tool_key,input_fields}), ai(渲染预览,meta{tool_key,output_kind})]`（空产物不记，`lib/toolbox/aiSummary.ts` 纯助手+5 单测），InlineToolbox 挂本章号、ToolboxPage silent 项目级(`chapter_no=null`)；Workbench 挂 hook 单实例 + 抽屉 + 底栏「AI 对话」入口（`AI_CONVERSATION_ENABLED` 灰度）+ 三个再接受回调（成功关抽屉+toast、refine 漂移不盲替换）。**留痕永不改「是否/何时写正文」**——正文仍只在验收事务落库（守 #3）。**门禁全绿**：lint/typecheck 干净 · vitest **706 passed**（681 基线 + 25：aiConversation 11 + useAiConversation 9 + aiSummary 5）· `test:coverage` 通过（`lib/**` 95.44% ≥ 80，新纯模块/hook 全测）· `build` OK。→ 剩 **AC-4 @qa** E2E（计划 §6）+ P4 #1/#6 围栏。 - [2026-07-09] @backend — **✅ AC-2 repo + append/list 端点 + schema + 契约**（计划 §3，`c675a74`）。repo `packages/core/ww_core/domain/ai_message_repo.py`（`AiMessageRepo`/`SqlAiMessageRepo`/`AiMessageView`/`AiTurnRow`，经 `ww_core.domain` 导出）：`append`(一批多条 bubble → 批内 `seq`=enumerate 0 基位置、同 `thread_id`、meta 新 dict 不原地 mutate、`add_all`+一次 flush+refresh，**只 flush 不 commit**) + `list_for_chapter`(`created_at DESC, seq DESC` newest-first；`chapter_no` 非 None → 本章 ∪ 项目级 `IS NULL`，None → 仅项目级；可选 `kind`；`limit/offset`)。schemas `apps/api/ww_api/schemas/ai_messages.py`：批级 append 信封 `AiMessageAppendRequest{thread_id,chapter_no(ge=1)|null,kind,tool_key,messages}` + `AiMessageInput{role,content(min1,max200_000),meta(序列化≤50_000)}` + `AiMessageView` + `AiMessageListResponse{messages}`；`Literal` 为 kind/role 枚举权威（边界 422）。端点 `apps/api/ww_api/routers/ai_messages.py`（prefix `/projects`，tag `ai-messages`，已注册 main）：**`POST /projects/{id}/ai-messages`**(201，项目 404，校验 422，append 只 flush + 端点 `commit()`) + **`GET /projects/{id}/ai-messages?chapter_no=&kind=`+`PageDep`**(200，newest-first，只读)。**这是显式记录写入**——五个生成端点逐字节不变、只读（守 #3）；`ai_messages` 旁路侧记录，`assemble()`/agent 永不读（守 #1/#6）。注入缝 `get_ai_message_repo`。**TDD**：repo 真 pg 单测 9 passed（seq 0..n-1 / 同 thread_id / `created_at DESC,seq DESC` / union 含 NULL / 分页 / flush-only 跨 session 不可见）；端点集成 12 passed（round-trip / 本章∪项目级 / kind 过滤 / 404×2 / 422×5〔非法 role/kind、空 messages、空 content、超上限 content、超上限 meta〕 / commit 边界 / 响应 key 集合不泄漏）。**门禁全绿**：ruff/format 干净 · mypy **238 文件** · **`alembic check` 无漂移**（本任务不加模型）· `export_openapi` 成功（新端点+4 schema 已入表面）· pytest **992 passed**（971 基线 + 21）· 新模块 cov router 100%/schemas 100%/repo 97%（余 Protocol `...` 桩），全局 89%。契约 **C3 扩（AC-2）已登记「稳定」** → **@frontend AC-3 须 `cd apps/web && pnpm gen:api`** 后消费码生成名。 - [2026-07-09] @db — **✅ AC-1 `ai_messages` 表 + 迁移**（计划 §2 canonical DDL）。`packages/db/ww_db/models.py` 加 `AiMessage(UuidPk, CreatedAt, Base)`（补 `Uuid` 导入）：`project_id`(FK projects CASCADE, NOT NULL, indexed)/`chapter_no`(Integer nullable=项目级)/`thread_id`(Uuid NOT NULL)/`seq`(Integer NOT NULL, repo 批内 0 基定序)/`kind`(Text 无 CHECK)/`tool_key`(Text nullable)/`role`(Text + `ck_ai_messages_role` CHECK author|ai)/`content`(Text NOT NULL 全文)/`meta`(JSONB NOT NULL server_default `'{}'`)/`id`+`created_at`(mixin)；append-only 无 updated_at。复合索引 `ix_ai_messages_project_chapter=(project_id,chapter_no,created_at DESC)` 手写 `sa.text` DESC + FK 单列索引。迁移 `ceabd3c2c00d`（down=`f6a7b8c9d0e1`，**autogenerate 仅捕获 ai_messages 无杂散漂移**，DESC 表达式归一为 `sa.text`）。**T0.2 MVP schema 后第一张新业务表**——对话历史是 18 表都无的能力、正交于 review/job/手稿，属真新增（非重复建表）；是**旁路侧记录非手稿真源**，`assemble()`/agent 永不读（守 #1/#3/#6，围栏留 AC-4/P4 收口）。**门禁全绿**：ruff/format 干净 · mypy **233 文件** · `alembic upgrade head` + **`alembic check` 无漂移** + **downgrade→upgrade 往复干净** · pytest **971 passed**（基线不变，无代码用新表）。契约 **C2 扩（AC-1）已登记「稳定」** → AC-2 @backend 可动工。 diff --git a/memory/gotchas.md b/memory/gotchas.md index 821b5a7..c718ac9 100644 --- a/memory/gotchas.md +++ b/memory/gotchas.md @@ -90,3 +90,4 @@ - [2026-06-24] @llm/@devops Prompt 外置：`prompts/*.md` 文件名按 **spec.name（连字符）** 非 Python 变量名——`style` agent 的 name 是 `"style"`→`style.md`（非 `style_drift`），另有 `character-gen.md`/`golden-finger.md`/`book-title.md`/`fine-outline.md`/`de-ai.md`。**改 prompt 内容后必须重生成金标准** `uv run python packages/agents/tests/_gen_golden.py`（覆盖 `tests/fixtures/prompt_hashes.json`），否则 `test_prompt_loader.py` 字节回归红；旧常量含反斜杠折行——外迁/比对一律用**运行时值**别用源码文本。`.gitattributes` 锁 `prompts/*.md text eol=lf`（路径含斜杠是**根锚定**，必须写完整嵌套路径 `packages/agents/ww_agents/prompts/*.md` 才匹配，裸 `prompts/*.md` 不生效）。 - [2026-06-24] @devops wheel 默认不打非-`.py` 数据文件——`prompts/*.md` 必须在 `packages/agents/pyproject.toml` 显式纳入（hatchling 用 `artifacts`，**别用 `force-include`**：hatchling 已含包目录下全部文件，force-include 会 duplicate-path 构建失败）；源码树 pytest **测不出**漏带（fail-fast 仅裸装时触发），靠 CI `agents-wheel-smoke`（build→裸装→`import ww_agents; assert ww_agents.SPECS`）守。`ww_agents` import 链拉起 `ww_llm_gateway`（`types.Tier`）→需 `structlog`，故 `packages/{llm_gateway,core}` 必须自声明 `structlog`（原靠 apps/api 传递，裸装 import 会 ModuleNotFoundError）。 - [2026-06-24] @backend `SpecResolver`：内置 name 走 `SPECS` 纯内存、**零 DB**；保留命名空间守卫在 **SkillRegistry 入库校验期**（非 resolver 读路径），用户 skill 与内置同名（`REVIEW_RESERVED_NAMES ∪ set(SPECS)`）→ `AppError(VALIDATION)`。name **精确字符串相等**：拼错近似名（`character_gen` vs `character-gen`）`output_schema_for` 返 None/不命中，是预期行为非 bug。 +- [2026-07-09] @qa **`ai_messages` 是 append-only 旁路侧记录（与 `usage_ledger` 同物种），绝不喂 `assemble()`/prompt——守不变量 #1/#6（计划 §8）。** 它是「作者↔AI 说了什么」的真源，**不是手稿/审稿真源**：无记忆注入(#6)、无裁决(#3/#4)权威。生成链（`packages/core/ww_core/{memory,orchestrator}`）任一模块 import/引用 `AiMessage`/`AiMessageRepo`/`ai_message_repo` = 制造第二个非确定性真源，**是 code-review BLOCKER**。围栏已工具化：`tests/test_ai_messages_not_in_generation_path.py`（AST 扫两目录 import/Name/Attribute + 断言 `MemoryRepos` 捆绑与 `assemble` 形参都不含它；植入违规自证非空守卫）——把 ai_message repo 接进 `assemble()`/prompt 会直接变红。日后加新 kind 或改 repo 都别越过这道墙。 diff --git a/tests/test_ai_chat_history_e2e.py b/tests/test_ai_chat_history_e2e.py new file mode 100644 index 0000000..5b2ccca --- /dev/null +++ b/tests/test_ai_chat_history_e2e.py @@ -0,0 +1,531 @@ +"""AC-4 端到端：AI 对话（聊天记录）持久化（真实 pg，**零 LLM / 无网关**）。 + +本功能是纯 CRUD 旁路留痕（`docs/design/ai-chat-history-plan.md`）——五个生成端点逐字节 +不变、只读；作者↔AI 每轮往复经 **`POST /projects/{id}/ai-messages`** 显式 append，供每章 +「AI 对话」抽屉重放。故这里**不** override 任何网关（无 `_FakeAdapter`）：只打两端点、断 +DB 真源。镜像 `tests/test_t6_toolbox_e2e.py` 的 `e2e_sm` engine/session 生命周期；无 pg → skip。 + +覆盖计划 §6 的 API 级用例： +1. 缓冲多轮线程：一批落一次「已定的 refine+clarify 交换」（作者含糊→AI 反问→作者答→AI 定稿）， + 同 `thread_id`、`seq` 0..n；GET newest-first、`seq`/`role`/`thread_id`/`meta` 往复保真。 +2. 作用域分区：章级线程(chapter_no=N) ∪ 项目级(chapter_no=null) 生成——GET `?chapter_no=N` + 两者都返回；GET 缺省（项目级）只返回项目级。 +3. newest-first + 分页（跨多批已提交）+ `kind` 过滤。 +4. append-only / read-only：append 只写 `ai_messages`（其它业务表零变化、`usage_ledger` 恒 0 + 证明无网关）；GET 只读不建行。 +5. 未知项目 → 404；非法 role/kind、空 messages、超上限 content/meta → 422（失败零落库）。 +6. clarify-abandon 的 API 级代理：服务端无部分写路径——缓冲在客户端，端点只持久化被显式 + POST 的已定线程；从未 POST 的放弃线程在库里不留痕。 +""" + +from __future__ import annotations + +import asyncio +import uuid +from collections.abc import AsyncIterator + +import httpx +import pytest +from asgi_lifespan import LifespanManager +from sqlalchemy import delete, func, select +from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker +from ww_db import get_sessionmaker +from ww_db.models import ( + AiMessage, + Chapter, + ChapterReview, + Job, + Project, + UsageLedger, + WorldEntity, +) + +# content 超上限探针 = 手稿上限 + 1（AI_MESSAGE_CONTENT_MAX = 200_000）。 +_CONTENT_CAP = 200_000 +# meta 序列化上界（_META_MAX_SERIALIZED = 50_000）——用超此长度的字符串塞 meta 触发 422。 +_META_CAP = 50_000 + + +@pytest.fixture +async def e2e_sm() -> AsyncIterator[async_sessionmaker[AsyncSession]]: + """真实 DB session 工厂；无 pg 时跳过（每测试清缓存重建 engine、结束 dispose）。""" + get_sessionmaker.cache_clear() + maker = get_sessionmaker() + try: + async with maker() as probe: + await probe.execute(select(1)) + except Exception: + pytest.skip("postgres not reachable") + yield maker + await maker.kw["bind"].dispose() + get_sessionmaker.cache_clear() + + +async def _count(sm: async_sessionmaker[AsyncSession], model: type, project_uuid: uuid.UUID) -> int: + async with sm() as s: + return int( + ( + await s.execute( + select(func.count()).select_from(model).where(model.project_id == project_uuid) # type: ignore[attr-defined] + ) + ).scalar_one() + ) + + +async def _cleanup(sm: async_sessionmaker[AsyncSession], project_uuid: uuid.UUID) -> None: + """按 FK 顺序清理（ai_messages 虽 CASCADE，显式删更清晰）。""" + async with sm() as cleanup: + await cleanup.execute(delete(AiMessage).where(AiMessage.project_id == project_uuid)) + await cleanup.execute(delete(Project).where(Project.id == project_uuid)) + await cleanup.commit() + + +async def _create_project(client: httpx.AsyncClient, title: str) -> uuid.UUID: + resp = await client.post("/projects", json={"title": title, "genre": "玄幻"}) + assert resp.status_code == 201, resp.text + return uuid.UUID(resp.json()["id"]) + + +async def test_buffered_multi_turn_thread_persists_and_lists_newest_first( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 1/4：一批落「已定的 refine+clarify 交换」→ GET newest-first，seq/role/meta 保真。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + thread_id = uuid.uuid4() + original_segment = "他站在雨里，一动不动。" + options = [ + {"label": "保留第一人称", "value": "keep_first"}, + {"label": "改第三人称", "value": "third"}, + ] + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + project_uuid = await _create_project(client, "AC-4 缓冲线程作品") + + # 一批落库：作者(含糊)→AI(反问)→作者(答)→AI(定稿)；选项/version_no/segment 进 meta。 + resp = await client.post( + f"/projects/{project_uuid}/ai-messages", + json={ + "thread_id": str(thread_id), + "chapter_no": 5, + "kind": "refine", + "messages": [ + {"role": "author", "content": "帮我把这段润色得更有张力", "meta": {}}, + { + "role": "ai", + "content": "你希望保留原文的第一人称视角吗？", + "meta": { + "parent_kind": "refine", + "options": options, + "allow_free_text": True, + "verification": "视角一致性", + }, + }, + {"role": "author", "content": "保留第一人称", "meta": {}}, + { + "role": "ai", + "content": "雨丝斜织，他钉在原地，任凭寒意浸透脊背。", + "meta": {"version_no": 1, "segment": original_segment}, + }, + ], + }, + ) + assert resp.status_code == 201, resp.text + appended = resp.json()["messages"] + # 201 回显：4 条、seq 0..3（repo 赋值）、同 thread_id、同 chapter_no。 + assert [m["seq"] for m in appended] == [0, 1, 2, 3] + assert {m["thread_id"] for m in appended} == {str(thread_id)} + assert {m["chapter_no"] for m in appended} == {5} + assert all(m["id"] for m in appended) # id/created_at 已回填供乐观对账。 + + # GET newest-first：同批共享 created_at → 靠 seq DESC 定序 → [3,2,1,0]。 + listed = await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=5") + assert listed.status_code == 200 + rows = listed.json()["messages"] + assert [r["seq"] for r in rows] == [3, 2, 1, 0] + assert [r["role"] for r in rows] == ["ai", "author", "ai", "author"] + assert {r["thread_id"] for r in rows} == {str(thread_id)} + # meta 往复保真：AI 反问的选项列表 + AI 定稿的 version_no/segment。 + ai_question = next(r for r in rows if r["seq"] == 1) + assert ai_question["meta"]["options"] == options + assert ai_question["meta"]["allow_free_text"] is True + ai_final = next(r for r in rows if r["seq"] == 3) + assert ai_final["meta"]["version_no"] == 1 + # segment 存于 AI 定稿行 → 抽屉「回填选段」可重锚（计划 §3.4/§4.5）。 + assert ai_final["meta"]["segment"] == original_segment + # 抽屉 re-accept 直读全文 content（未截断）。 + assert ai_final["content"] == "雨丝斜织，他钉在原地，任凭寒意浸透脊背。" + + # DB 真源：恰 4 行。 + assert project_uuid is not None + assert await _count(e2e_sm, AiMessage, project_uuid) == 4 + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) + + +async def test_scope_partition_chapter_union_and_project_only( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 2/6：章级 ∪ 项目级(NULL)——GET `?chapter_no=N` 两者都返回；缺省只返回项目级。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + chapter_thread = uuid.uuid4() + project_thread = uuid.uuid4() + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + project_uuid = await _create_project(client, "AC-4 作用域分区作品") + + # A) 章级线程（rewrite，chapter_no=3）。 + a = await client.post( + f"/projects/{project_uuid}/ai-messages", + json={ + "thread_id": str(chapter_thread), + "chapter_no": 3, + "kind": "rewrite", + "messages": [ + { + "role": "author", + "content": "整章节奏太平", + "meta": {"version_no": 1}, + }, + { + "role": "ai", + "content": "<重写后的整章正文>", + "meta": {"version_no": 1}, + }, + ], + }, + ) + assert a.status_code == 201, a.text + + # B) 项目级工具箱生成（generator，chapter_no 缺省 = NULL）。 + b = await client.post( + f"/projects/{project_uuid}/ai-messages", + json={ + "thread_id": str(project_thread), + "kind": "generator", + "tool_key": "brainstorm", + "messages": [ + { + "role": "author", + "content": "想要三个突破常规的脑洞", + "meta": { + "tool_key": "brainstorm", + "input_fields": {"brief": "..."}, + }, + }, + { + "role": "ai", + "content": "1. 天降系统…\n2. 转生灵宠…\n3. 末世种田…", + "meta": {"tool_key": "brainstorm", "output_kind": "IdeaListResult"}, + }, + ], + }, + ) + assert b.status_code == 201, b.text + assert all(m["chapter_no"] is None for m in b.json()["messages"]) + + # GET ?chapter_no=3 → 本章 ∪ 项目级 = A + B（4 行，两线程都在）。 + union = ( + await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=3") + ).json()["messages"] + assert len(union) == 4 + assert {r["thread_id"] for r in union} == { + str(chapter_thread), + str(project_thread), + } + assert {r["chapter_no"] for r in union} == {3, None} + + # GET 缺省（无 chapter_no）→ 仅项目级 B（2 行，chapter_no 全 NULL）。 + project_only = (await client.get(f"/projects/{project_uuid}/ai-messages")).json()[ + "messages" + ] + assert len(project_only) == 2 + assert {r["thread_id"] for r in project_only} == {str(project_thread)} + assert all(r["chapter_no"] is None for r in project_only) + + # GET 另一无对话的章 → union 仍带出项目级 B（项目级现于每章抽屉，计划 §6-(6)）。 + other_chapter = ( + await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=99") + ).json()["messages"] + assert {r["thread_id"] for r in other_chapter} == {str(project_thread)} + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) + + +async def test_newest_first_pagination_and_kind_filter_across_batches( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 3：跨多批已提交 → newest-first + 分页 + `kind` 过滤。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + continue_thread = uuid.uuid4() + refine_thread = uuid.uuid4() + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + project_uuid = await _create_project(client, "AC-4 分页过滤作品") + + async def _post(body: dict) -> None: + resp = await client.post(f"/projects/{project_uuid}/ai-messages", json=body) + assert resp.status_code == 201, resp.text + # 分批提交须 created_at 单调递增 → 睡极短保跨批定序确定（同批仍靠 seq）。 + await asyncio.sleep(0.02) + + # 批1：续写候选0（ai-only，续写无作者输入）。 + await _post( + { + "thread_id": str(continue_thread), + "chapter_no": 7, + "kind": "continue", + "tool_key": "continue", + "messages": [ + {"role": "ai", "content": "候选一…", "meta": {"candidate_index": 0}} + ], + } + ) + # 批2：同线程「再来一个」候选1。 + await _post( + { + "thread_id": str(continue_thread), + "chapter_no": 7, + "kind": "continue", + "tool_key": "continue", + "messages": [ + {"role": "ai", "content": "候选二…", "meta": {"candidate_index": 1}} + ], + } + ) + # 批3：refine 两条（最新批）。 + await _post( + { + "thread_id": str(refine_thread), + "chapter_no": 7, + "kind": "refine", + "messages": [ + {"role": "author", "content": "再紧凑些", "meta": {}}, + { + "role": "ai", + "content": "定稿…", + "meta": {"version_no": 1, "segment": "原段"}, + }, + ], + } + ) + + # newest-first：批3(created_at 最大) → 批2 → 批1；批内 seq DESC。 + rows = ( + await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=7") + ).json()["messages"] + assert [r["kind"] for r in rows] == [ + "refine", # 批3 seq1 (ai 定稿) + "refine", # 批3 seq0 (author) + "continue", # 批2 + "continue", # 批1 + ] + assert [r["role"] for r in rows] == ["ai", "author", "ai", "ai"] + + # 分页：limit=1 取最新一行；offset 前移。 + page0 = ( + await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=7&limit=1") + ).json()["messages"] + assert len(page0) == 1 and page0[0]["content"] == "定稿…" + page1 = ( + await client.get( + f"/projects/{project_uuid}/ai-messages?chapter_no=7&limit=1&offset=1" + ) + ).json()["messages"] + assert len(page1) == 1 and page1[0]["role"] == "author" + + # kind 过滤：只取 continue（批2、批1），candidate_index newest-first [1,0]。 + only_continue = ( + await client.get( + f"/projects/{project_uuid}/ai-messages?chapter_no=7&kind=continue" + ) + ).json()["messages"] + assert [r["kind"] for r in only_continue] == ["continue", "continue"] + assert [r["meta"]["candidate_index"] for r in only_continue] == [1, 0] + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) + + +async def test_append_writes_only_ai_messages_and_get_is_read_only( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 4：append 只写 ai_messages（其它业务表零变、usage_ledger 恒 0）；GET 只读不建行。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + project_uuid = await _create_project(client, "AC-4 旁路留痕作品") + + # 关键旁路表快照（append 前）。 + side_models = [UsageLedger, Chapter, ChapterReview, WorldEntity, Job] + before = {m.__name__: await _count(e2e_sm, m, project_uuid) for m in side_models} + + resp = await client.post( + f"/projects/{project_uuid}/ai-messages", + json={ + "thread_id": str(uuid.uuid4()), + "kind": "generator", + "tool_key": "book-title", + "messages": [ + {"role": "author", "content": "起个书名", "meta": {}}, + {"role": "ai", "content": "《九霄剑主》", "meta": {}}, + ], + }, + ) + assert resp.status_code == 201, resp.text + + # 只写 ai_messages：其它业务表逐一零变化。 + after = {m.__name__: await _count(e2e_sm, m, project_uuid) for m in side_models} + assert after == before, f"append touched a non-ai_messages table: {before}->{after}" + # 无网关/无 LLM：usage_ledger 恒 0（纯 CRUD 证据）。 + assert await _count(e2e_sm, UsageLedger, project_uuid) == 0 + assert await _count(e2e_sm, AiMessage, project_uuid) == 2 + + # GET 只读：读后 ai_messages 计数不变（不建行）。 + assert ( + await client.get(f"/projects/{project_uuid}/ai-messages") + ).status_code == 200 + assert await _count(e2e_sm, AiMessage, project_uuid) == 2 + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) + + +async def test_unknown_project_404_and_validation_422( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 5：未知项目 → 404；各类非法请求（role/kind/空/超上限）→ 422，且零落库。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + # 未知项目：POST + GET 皆 404 NOT_FOUND 信封。 + ghost = uuid.uuid4() + miss_post = await client.post( + f"/projects/{ghost}/ai-messages", + json={ + "thread_id": str(uuid.uuid4()), + "kind": "refine", + "messages": [{"role": "author", "content": "x", "meta": {}}], + }, + ) + assert miss_post.status_code == 404 + assert miss_post.json()["error"]["code"] == "NOT_FOUND" + miss_get = await client.get(f"/projects/{ghost}/ai-messages") + assert miss_get.status_code == 404 + assert miss_get.json()["error"]["code"] == "NOT_FOUND" + + project_uuid = await _create_project(client, "AC-4 校验作品") + thread = str(uuid.uuid4()) + base = {"thread_id": thread, "kind": "refine"} + + # 各 422 非法体：非法 role / 非法 kind / 空 messages / 超上限 content / meta。 + invalid_bodies = [ + {**base, "messages": [{"role": "system", "content": "x"}]}, # 非法 role + { + **base, + "kind": "nonsense", # 非法 kind + "messages": [{"role": "author", "content": "x"}], + }, + {**base, "messages": []}, # 空 messages + { + **base, + "messages": [{"role": "author", "content": "x" * (_CONTENT_CAP + 1)}], + }, # content 超上限 + { + **base, + "messages": [ + {"role": "author", "content": "x", "meta": {"blob": "y" * _META_CAP}} + ], + }, # meta 序列化超上限 + ] + for body in invalid_bodies: + bad = await client.post(f"/projects/{project_uuid}/ai-messages", json=body) + assert bad.status_code == 422, ( + f"expected 422 for {body!r}, got {bad.status_code}" + ) + + # 失败零落库：所有 422 都没有写入任何行。 + assert project_uuid is not None + assert await _count(e2e_sm, AiMessage, project_uuid) == 0 + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) + + +async def test_clarify_abandon_persists_only_explicit_posts( + e2e_sm: async_sessionmaker[AsyncSession], +) -> None: + """用例 6：clarify-abandon 代理——服务端无部分写路径，只持久化被显式 POST 的已定线程。""" + from ww_api.main import create_app + + app = create_app() + transport = httpx.ASGITransport(app=app) + concluded_thread = uuid.uuid4() + project_uuid: uuid.UUID | None = None + try: + async with LifespanManager(app): + async with httpx.AsyncClient(transport=transport, base_url="http://test") as client: + project_uuid = await _create_project(client, "AC-4 澄清放弃作品") + + # 从未 POST 的（放弃的）线程 → 库里无痕：GET 空、计数 0。 + # 缓冲是客户端的；纯放弃 = 不 POST，服务端没有任何部分写入路径。 + empty = await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=8") + assert empty.status_code == 200 + assert empty.json()["messages"] == [] + assert await _count(e2e_sm, AiMessage, project_uuid) == 0 + + # 只有「已定」线程被显式 POST 才落库。 + await client.post( + f"/projects/{project_uuid}/ai-messages", + json={ + "thread_id": str(concluded_thread), + "chapter_no": 8, + "kind": "refine", + "messages": [ + {"role": "author", "content": "润色这段", "meta": {}}, + { + "role": "ai", + "content": "润色结果", + "meta": {"version_no": 1, "segment": "原段"}, + }, + ], + }, + ) + + # GET 恰返回被显式 POST 的 2 行——无任何「缓冲/放弃」幻影行。 + after = ( + await client.get(f"/projects/{project_uuid}/ai-messages?chapter_no=8") + ).json()["messages"] + assert len(after) == 2 + assert {r["thread_id"] for r in after} == {str(concluded_thread)} + assert project_uuid is not None + assert await _count(e2e_sm, AiMessage, project_uuid) == 2 + finally: + if project_uuid is not None: + await _cleanup(e2e_sm, project_uuid) diff --git a/tests/test_ai_messages_not_in_generation_path.py b/tests/test_ai_messages_not_in_generation_path.py new file mode 100644 index 0000000..091987e --- /dev/null +++ b/tests/test_ai_messages_not_in_generation_path.py @@ -0,0 +1,144 @@ +"""AC-4 / 计划 §8(#1/#6) 真源围栏守卫：`ai_messages` 绝不进生成链。 + +`ai_messages` 是 append-only 旁路侧记录（与 `usage_ledger` 同物种）——记录「作者↔AI 说了 +什么」，**不是手稿真源**。守不变量 #1（DB 单一真源指手稿/审稿真源）/ #6（记忆注入是确定性 +选择，非把聊天记录喂进 prompt）：`assemble()` 与任何 memory/orchestrator 节点**永不读它**。 + +这是一道**能变红的**真守卫（非空断言）： +- 若日后有人把 `AiMessageRepo`/`AiMessage` 接入 `packages/core/ww_core/{memory,orchestrator}` + 下任一模块（import 或引用），`test_generation_path_never_references_ai_messages` 直接失败。 +- 若有人把 ai_message repo 塞进 `assemble()` 消费的 `MemoryRepos` 捆绑， + `test_assemble_repo_bundle_excludes_ai_messages` 直接失败。 +- `test_detector_flags_a_planted_violation` 用植入的违规源证明探测器确会命中——守卫非空。 + +纯静态分析（AST + 反射），无 DB / 无网络 / 无 LLM。 +""" + +from __future__ import annotations + +import ast +import dataclasses +import inspect +from pathlib import Path + +import ww_core.memory.assemble as assemble_mod +from ww_core.domain.repositories import MemoryRepos +from ww_core.memory import assemble + +# 生成链禁止出现的标识符（repo/ORM 模型/视图/行）。 +_FORBIDDEN_NAMES = frozenset( + { + "AiMessage", + "AiMessageRepo", + "SqlAiMessageRepo", + "AiMessageView", + "AiTurnRow", + } +) +# 禁止 import 的模块子串（catch `from ww_core.domain.ai_message_repo import ...`）。 +_FORBIDDEN_MODULE_SUBSTR = "ai_message" + +# 受守卫的生成链目录（记忆装配 + 编排器；相对仓库根）。 +_GUARDED_DIRS = ( + "packages/core/ww_core/memory", + "packages/core/ww_core/orchestrator", +) + +_REPO_ROOT = Path(__file__).resolve().parents[1] + + +def _ai_message_refs(source: str, filename: str) -> list[str]: + """返回 `source` 里对 ai_messages 生成链禁忌标识符的所有引用（AST 精确匹配）。 + + 命中面（真实的「接入生成链」向量）： + - `import ...ai_message...` / `from ...ai_message... import ...`（模块子串）； + - `from X import AiMessageRepo`（导入名在禁忌集）； + - 代码里任意 `AiMessageRepo` 名引用 / `x.AiMessageView` 属性访问（Name/Attribute）。 + docstring/注释是 `ast.Constant`，不计入——避免「文档提到它」的假阳性。 + """ + hits: list[str] = [] + tree = ast.parse(source, filename=filename) + for node in ast.walk(tree): + if isinstance(node, ast.ImportFrom): + if node.module and _FORBIDDEN_MODULE_SUBSTR in node.module.lower(): + hits.append(f"from {node.module} import ...") + for alias in node.names: + if alias.name in _FORBIDDEN_NAMES: + hits.append(f"from {node.module} import {alias.name}") + elif isinstance(node, ast.Import): + for alias in node.names: + if _FORBIDDEN_MODULE_SUBSTR in alias.name.lower(): + hits.append(f"import {alias.name}") + elif isinstance(node, ast.Name) and node.id in _FORBIDDEN_NAMES: + hits.append(f"name {node.id}") + elif isinstance(node, ast.Attribute) and node.attr in _FORBIDDEN_NAMES: + hits.append(f"attr .{node.attr}") + return hits + + +def _guarded_py_files() -> list[Path]: + files: list[Path] = [] + for rel in _GUARDED_DIRS: + d = _REPO_ROOT / rel + assert d.is_dir(), f"guarded dir missing (moved?): {d}" + files.extend(sorted(d.rglob("*.py"))) + return files + + +def test_generation_path_never_references_ai_messages() -> None: + """memory/orchestrator 下任一模块都不得 import/引用 ai_messages（否则真源蔓延）。""" + files = _guarded_py_files() + # 非空守卫：确有文件被扫（否则断言形同虚设）。 + assert len(files) >= 20, f"expected the generation path to have modules; scanned {len(files)}" + + violations: dict[str, list[str]] = {} + for path in files: + refs = _ai_message_refs(path.read_text(encoding="utf-8"), str(path)) + if refs: + violations[str(path.relative_to(_REPO_ROOT))] = refs + + assert not violations, ( + "ai_messages leaked into the generation path (violates invariant #1/#6 — it is an " + f"append-only side-record, never a truth-source): {violations}" + ) + + +def test_assemble_repo_bundle_excludes_ai_messages() -> None: + """`assemble()` 消费的 `MemoryRepos` 捆绑不得含任何 ai_message repo 字段。""" + for field in dataclasses.fields(MemoryRepos): + assert "ai_message" not in field.name.lower(), ( + f"MemoryRepos.{field.name} wires ai_messages into assemble() — forbidden (#1/#6)" + ) + annotation = str(field.type) + assert "AiMessage" not in annotation, ( + f"MemoryRepos.{field.name} type {annotation!r} references AiMessage — forbidden (#1/#6)" + ) + + +def test_assemble_signature_only_takes_memory_repos_bundle() -> None: + """`assemble` 的 repo 入参恒为 `MemoryRepos`——阻止旁路加一个 ai_message repo 形参。""" + # eval_str 求值 PEP 563 字符串化注解（assemble.py 有 `from __future__ import annotations`）。 + sig = inspect.signature(assemble, eval_str=True) + repos_param = sig.parameters["repos"] + assert repos_param.annotation is MemoryRepos + # assemble 模块本身也不得引用禁忌标识符（双保险，直接扫其源）。 + source = inspect.getsource(assemble_mod) + assert _ai_message_refs(source, "assemble.py") == [] + + +def test_detector_flags_a_planted_violation() -> None: + """证明守卫非空：植入的违规源确被探测器命中（否则上面的绿全是假绿）。""" + planted_import = "from ww_core.domain.ai_message_repo import AiMessageRepo\n" + assert _ai_message_refs(planted_import, "") + + planted_usage = ( + "from ww_core.domain import AiMessageRepo\n" + "async def assemble(repos, ai_log: AiMessageRepo):\n" + " rows = await ai_log.list_for_chapter(...)\n" + " return rows\n" + ) + assert _ai_message_refs(planted_usage, "") + + # 干净源不命中（避免误报——探测器有区分度）。 + clean = "from ww_core.domain.repositories import MemoryRepos\nx = 1\n" + assert _ai_message_refs(clean, "") == []