test(qa): craft 成稿质量前后对照评估回路（灵感 F2）

固定 3 题材金标准输入（仙侠/都市/悬疑，各含最小 spec+world+outline 及开篇章），
经真实 assemble()+build_write_request() 组装 craft-on / craft-off 两版写章请求：

- dry 模式（默认，零成本、免真实 LLM，可进 CI）：断言 craft 教条块、genre 片段、
  开篇黄金三章 marker 确实按 fixture 注入正确内容，且 craft-off 基线不含这三者；
  并把两版将发送给模型的 prompt 并排 dump 供人眼看。这些断言即 tests/test_craft_eval.py，
  是 craft/审查注入路径的回归门禁。
- --live 模式（真实付费 LLM，绝不进 CI）：经 gateway 各生成一章 + analyst 五维裁判，
  凭据读 DB、路由走 tier_routing，与生产写章同一网关路径。

入口 `uv run python -m tests.craft_eval [--live]`；用法/成本/门禁说明见 README.md。

PROGRESS.md

@@ -161,6 +161,11 @@ 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`
 
+- [2026-07-06] @backend/@llm/@frontend — **✅ 灵感⑦ 群像按定位配比·缩减版 + §0 归属纠偏（develop 直提）**。角色群像生成支持「按定位配比」下发。**契约变更** `CharacterGenerateRequest` 加 `role_mix:dict[str,int]|None`——`model_validator(mode="after")` 定义与既有 `count` 的关系：在场时校验（非空、每项≥1、总和≤`MAX_GENERATE_COUNT`=12）并把 `count` 归一为各值之和（构造期归一化，忽略客户端 count）；缺省退回单一 `role`+`count`。**§0 归属纠偏**：`build_character_gen_context`+`run_character_gen`（`ww_core/orchestrator/generation_node.py`）确为 **@llm 域**——加末位默认 `role_mix=None` 关键字参，在场时注入「角色定位配比」块（保插入序、确定性、无时间戳，守 #6/#7 assemble 纯函数；覆盖单一 role 行）；`routers/generation.py` 穿 `role_mix=body.role_mix`。`character-gen.md` 加「按配比严格分配 role、各定位之和=总数」纪律 → **regen `prompt_hashes.json`**（仅 character-gen 哈希变，**不进 SPECS→不 bump 计数**）。**前端** `cards.ts` 加 `sanitizeRoleMix`（去空/去重/clamp≥1/按 12 截断）+`roleMixTotal`+`buildCharacterGenerateRequest` 第 4 参 roleMix；`useCharacterGen` `CharacterGenInput.roleMix` 穿参；新 `RoleMixEditor.tsx` + `CharacterGenerator` SegmentedControl 单一/配比双模。**TDD**：schema 校验 7 例（总和≤12/空 dict/非正值拒绝）+ context 注入 2 例 + run 穿参 1 例 + 前端 cards 构造/clamp 5 例 + hook renderHook 1 例。**门禁全绿**：后端 ruff/format 干净 · mypy **222** · pytest **858**（+11）· 无迁移（未动模型，未跑 alembic）；前端 gen:api/lint/typecheck 干净 · vitest **543**（+6）· coverage 95%（cards.ts 100%）· build OK。记 contracts C3。守不变量 #2/#6/#7/#9。**无 DB 迁移**。
+
+- [2026-07-06] @llm/@db/@backend/@frontend — **✅ 灵感③ 人物塑造 advisory 审（develop 直提）**。写→审链新增第五审「人物塑造」：**advisory 单维——仅建议、不阻断验收**（不进 conflicts、不碰 `assert_conflicts_resolved`、**不入 `REVIEW_RESERVED_NAMES`**，D2）；只做人物塑造、不做氛围（氛围噪声大折进 style/pace）。**新 SPEC** `characterization_spec`（`tier="analyst"` 不写 model，`reads=("characters",)` 以人物卡 motive/appearance 为客观锚点，`writes=()` 只读，`prompts/characterization.md` **显式忽略文本长度与辞藻华丽度**、引用逐字片段+置信度对抗 LLM-judge 冗长/辞藻偏见）+ schema `CharacterizationReview{issues:[…character/aspect/where/quote/diagnosis/suggestion/confidence]}`（全字段默认值守解析韧性）+ 注册 `SCHEMA_CATALOG["characterization"]`；**计数三处 22→23**（`specs.py` assert / `test_prompt_loader.py` len / `schema_catalog.py` docstring）+ **regen `prompt_hashes.json`**（仅加一条，无其余漂移）。**DB** `chapter_reviews` 加 1 列 `characterization`（**JSONB nullable**，迁移 `f6a7b8c9d0e1`，nullable 免 backfill）。**契约变更** `ReviewHistoryItem`+`ReviewView` 加 `characterization`；接线链 `graph.py REVIEW_SPECS` 末位追加 + `chain/nodes.py ReviewRecordRepo` + `collect.py {CHARACTERIZATION/extract_characterization/ReviewRecorder/collect_reviews}` + `review_repo.py {Protocol/SqlReviewRepo/_to_view}` + `sse.py {EVENT_CHARACTERIZATION/characterization_event/_section_result_events}`（normalize 按 sorted 键自动纳入，字典序排最前不打乱既有四审断言）。**前端** `lib/review/{sse.ts,history.ts,useReviewStream.ts}` + 新 `CharacterizationPanel.tsx`（advisory 只展示、**无裁决 UI**、按置信度降序 + 低置信度折叠控注意力预算）+ `ReviewReport.tsx` 挂面板 + 双 seed。**门禁全绿**：后端 ruff/format 干净 · mypy **221** · pytest **848**（+核心 char 审 12 + agents spec 11 + api accept 2 等）· alembic 无漂移；前端 gen:api/lint/typecheck 干净 · vitest **536**（+10 含 `test_characterization_does_not_block_accept` 对应前端 reducer/normalize/parse）· coverage 95% · build OK。记 contracts C3+C4。守不变量 #2/#3/#9。依赖 ⑧（motive/appearance 已入 CharacterCard 契约，作客观锚点）。
+
+- [2026-07-06] @backend/@llm/@frontend — **✅ 灵感⑤ AI 立项方案生成+预填（develop 直提）**。向导阶段（project 未建）一键推演结构化书级蓝图预填立项向导。**新 SPEC** `project_plan_spec`（`tier="analyst"` 不写 model，`reads=()`/`writes=()` 纯预览，`prompts/project-plan.md`）+ schema `ProjectPlanResult`（书名候选 list + 时空背景/叙事结构/故事核心/结局设计/基调，**全字段默认值守解析韧性**）+ 注册 `SCHEMA_CATALOG["project-plan"]`；**计数三处 21→22**（`specs.py` assert / `test_prompt_loader.py` len / `schema_catalog.py` docstring）+ **regen `prompt_hashes.json`**（仅加一条）。**新端点** `POST /skills/project-plan/generate`（**不带 project 前缀**绕开 toolbox 404）← `ProjectPlanGenerateRequest`（全字段 `max_length` CR-H9，`brief` 用 `str|None` 令 codegen 可选）→ `ProjectPlanView`；**种子门控**缺 genre/logline→422，无凭据→503。复用 `build_brief_context`+`run_generator`（`run_generator`/`_build_request` 的 `project_id` 放宽 `uuid.UUID|None`——向导无 project，ledger.project_id 本 nullable）；新缝 `get_project_plan_gateway`（analyst）。**零迁移**。**前端** `wizard.ts` 加 `mapPlanResultToWizardForm`（书名/结构/基调→对应字段；时空背景/结局设计→折进 premise **不静默丢**；结局取向/视角预设枚举不臆造）+ `applyPlanPatch`（**仅回填空字段、保护已编辑**）+ `canGeneratePlan`/`toPlanSeedRequest` + `useProjectPlan` hook（renderHook 测试）+ `ProjectWizard.tsx` 第 2 步 `PlanAssistant`（生成→预览→按字段回填）。**门禁全绿**：后端 ruff/format 干净 · mypy **218** · pytest **814**（+新端点 6 + agents spec 8 测）· alembic 无漂移（零 ORM 变更）；前端 gen:api/lint/typecheck 干净 · vitest **526**（+28）· coverage 95%（wizard/generation 100%）· build OK。记 contracts C3 新端点。守不变量 #2/#3/#9。依赖 ④（已交付）。
 - [2026-06-26] @orchestrator — **全项目 Code Review（七维度）完成 → `QA-CR` 整改批立项**。6 个专科 reviewer 并行（后端 FastAPI 架构 / Python 规范 / LangGraph 最佳实践 / 前端 React 架构 / 前端 TS 类型安全 / 应用安全）+ 依赖时效自查；212 py + 168 tsx 全覆盖。绿门禁下仍揪出 **2 CRITICAL + 13 HIGH + 一批 MEDIUM/LOW**：**C1** Next.js 15.1.3 RCE（CVE-2025-66478「React2Shell」，App Router+RSC 受影响，升 15.1.9/15.5.7/16.x + 轮换密钥）；**C2** 多章链缺 `recursion_limit`（默认 25 supersteps、每章 4 节点→超 ~6 章即崩，批量产卷不可用）。HIGH 含 Kimi OAuth 占 DB 会话 / httpx 泄漏 / Anthropic 计费记 0 / request_id 未贯通 / `.env.example` 提交可用 Fernet key / 文本输入无 `max_length` / 前端 errorCode 重复 / Drawer 焦点 / RefineView 未取消。**确认无虞**：SSRF（base_url 仅硬编码）/SQL 注入（全 ORM）/XSS（无 dangerouslySetInnerHTML）/错误信封不泄密/Kimi device-flow 无 CSRF。已登记 `QA-CR` 任务表（CR-C1/C2 + CR-H1..H12 + CR-M1..M5 + CR-L1 + CR-D1）按 owner 分派 @frontend/@backend/@llm/@devops，落地顺序见当前阶段 background。**本批为评审立项，尚无代码改动。**
 - [2026-06-23] @orchestrator — **🎉 多章工作流链（Chain Workflow）交付 + 多 agent 交叉评审 + CRITICAL 修复 + 合并 develop（merge `513bf71`）**。对标星月「一键多章」+ 差异化「每章过四审、遇冲突才停人」。设计契约 `docs/design/chain-workflow.md`（commit `5ee9799`）。**多 agent 逐学科建造**（契约先行、目录单写者）：**C1 @llm**(`7f3eaab`) LangGraph **cyclic 链图** `orchestrator/chain/{state,nodes,graph}.py`（`build_chain_graph`：write→review→decide→accept 循环 + **interrupt-on-conflict** + resume；删死代码 `build_write_graph`，`write_node` 本体留用；7 单测 mock 网关+fake session+MemorySaver）·**C2 @backend**(`29349dc`) 服务/3 端点/schema/checkpointer 接线（见下条）·**C3 @db**(`548b7ab`) langgraph 检查点 4 表迁移 `d3e4f5a6b7c8`（`autocommit_block` 避开 `CREATE INDEX CONCURRENTLY` 与 alembic 事务死锁；`env.py` `include_object` 豁免这 4 表防漂移误报）·**C4 @qa**(`061792d`) E2E `tests/test_chain_workflow_e2e.py` 真 pg + mock 网关零 token（无冲突两章全自动 + 冲突→interrupt→awaiting→resume→accept）。**3 路对抗交叉评审**（python/fastapi/不变量）在**绿门禁下仍揪出 1 CRITICAL + 7 HIGH**——证伪"测试过=没问题"：CRITICAL=链把**绑请求 session 的 gateway** 传进 BackgroundTask，后台跑时 session 已关→write/review 的 `usage_ledger.record()` 写进死 session **静默丢失**（链计费全丢，违 #1；E2E 清了 ledger 却没断言其存在故漏掉）。**修复**(`d1ea83b`)：网关改 `chain_gateway_builder` 按节点 session 重建（仿 `digest_gateway_builder`/`style.py`）；日志脱敏（`_classify_job_error`+`exc_type`，去 `str(exc)`，§5）；resume 原子 claim `JobRepo.claim_awaiting_to_running`（防并发双跑同 thread_id，非 awaiting→409）+ 所有权校验（`JobView.project_id`，跨项目→404）；`ChainResumeAccepted`（去 start/count=0 哨兵）；删死导入 `extract_conflicts`；**补回归断言**（链 `usage_ledger` 必有行 + `chapter_reviews` 每章一行）。**出口（DoD）✅**：后端门禁绿 ruff/format 干净 · mypy **194 Success** · alembic **无漂移**（业务表零迁移，复用 `jobs.status=awaiting_input` + `result.awaiting_chapter`；唯一新 DDL=langgraph 检查点表）· pytest **608 passed**（+6 回归测）。首次真正启用 LangGraph checkpointer/interrupt。守不变量 #1/#3/#4/#5/#9。**剩余下一步（未做）**：① 前端链 UI（`pnpm gen:api` 纳入 run/resume + 发起页/进度轮询/awaiting 裁决续跑面板，复用 `ConflictAdjudication`）；② §12 文档回写 `ARCHITECTURE.md §5.2`（补 cyclic 链图 + 更正现状：写章直连流式/单章 accept 为事务/链图首启 checkpointer）；③ 未 push 远端。可选：续写/扩写节点入链；UGC 链市场需多租户（deferred）。
 - [2026-06-23] @backend — **✅ Chain Workflow C2（多章链 服务+端点+schema+checkpointer 接线，分支 `feat/chain-workflow`）**。承 C1(@llm)链图（`build_chain_graph` §3.3），落地服务/端点壳：① **3 端点**`routers/chain.py`（`POST .../chains/{key}/run`→202 `ChainRunAccepted`；`POST .../chains/runs/{job_id}/resume`→202；`GET /jobs/{id}` 复用）。校验：count 1..50→422、未知 chain_key→404、项目不存在→404、resume job 不存在→404、resume 非 awaiting→**409 新码 `CONFLICT`**、无凭据→503。② **schema**`schemas/chain.py`（`ChainRunRequest/ChainRunAccepted/ChainResumeRequest`，`ConflictDecision` 复用）。③ **服务**`services/chain_runner.py`（`run_chain_job` 仿 `run_job` 壳自建独立 session 驱动链图：set_running→ainvoke→据 `__interrupt__` 置 `awaiting_input` 或 `done`/`failed`；`build_accept_op` 在 apps/api 装配验收事务闭包注入图节点，守 #3/#4；token 不入 result/日志）+ `services/chain_deps.py`（`get_checkpointer_factory`：运行时 `AsyncPostgresSaver` 上下文 / 测试 MemorySaver）。④ **零迁移**（§7）：复用 `jobs`，新增 `status="awaiting_input"` + `JobRepo.set_awaiting`，awaiting 章经 `result.awaiting_chapter`；新错误码 `ErrorCode.CONFLICT`(409)。⑤ **新网关缝**`project_deps.build_chain_gateway`/`get_chain_gateway`（按请求 tier writer/analyst/light 分派——单档网关恒返该档会错路由 review/digest，见 gotcha）+ `get_digest_gateway_builder`。**单测**`apps/api/tests/test_chain.py` 12 用例（mock 网关 + MemorySaver + fake session/accept_op，无 DB/无网络/无真 LLM）：run→202、resume→202、未知 key 404、count 越界(>50/0) 422、resume 非 awaiting 409、项目/ job 不存在 404、run_chain_job 无冲突→done+written 满、冲突→awaiting→resume→done、错误脱敏、accept_op 冲突缺判→CONFLICT_UNRESOLVED。**全仓门禁绿**：ruff/format 干净、mypy **193 Success**、alembic **无漂移**（无 ORM 变更）、pytest **600 passed**。守不变量 #1/#3/#4/#5/#9。记 contracts C-Chain 稳定 + 2 decision + 3 gotcha。**唯一新增 DDL（langgraph 检查点表）= C3 迁移；前端链 UI = 契约稳定后 @frontend follow-up（本期非目标）。**
@@ -239,3 +244,4 @@ T0.1 monorepo 骨架 ✅ @devops · T0.2 16 MVP 表迁移（无漂移，users st
 - [2026-06-20] @frontend — **Tier4 完成**（T4-a 全局 AI 工具条 + T4-b 本章指令直通生成，分支 `feat/ux-tier4-ai-toolbar`，未并 develop）。T4-a：新 `lib/nav/ai-tools.ts`(aiToolItems) + `components/AiToolbar.tsx` 服务端组件，AppShell 顶栏下仅项目页渲染（写本章/审稿/大纲/设定库/工具箱常驻一排，写本章=朱砂主动作），用 `--chrome` CSS 变量统一 chrome 高度（项目页 7rem=header4+toolbar3，余 4rem），6 个固定高度页改用 `calc(100vh-var(--chrome,4rem))` 适配。T4-b（全链）：后端 `assemble/_build_volatile` 加 `directive` 形参（「本章指令」段领衔 volatile，**绝不入 stable_core**，守缓存前缀不变量 #9），`stream_draft` 加可选 body `DraftStreamRequest{directive}` 透传、`directive_len` 入 log 不记原文、不持久化无迁移，重生成 TS 客户端；前端新 `lib/workbench/directive.ts`(STYLE_PRESETS + composeDirective 纯函数)、`useDraftStream.start(p,no,directive?)` 非空时发 JSON body、Workbench 加可折叠「本章指令」textarea + 风格预设 chips。TDD：ai-tools 4 测 + directive 4 测 + assemble directive 测 + draft 端点 directive/向后兼容 2 测。门禁全绿：后端 ruff/format/mypy(163 文件)/**pytest 479 passed**；前端 lint/tsc/**vitest 213**/build。契约记 contracts C3 扩（T4-b）。
 - [2026-06-28] @backend/@frontend — **UI Phase G 后端依赖补齐**：`ProjectResponse` 增加 `updated_at` / `pending_review_count`，`chapters` 加 `updated_at` 迁移 `8c1d2e3f4a5b`；`SqlProjectRepo` 聚合项目自身、章节、审稿时间作为最近编辑，并按草稿是否晚于最近 accepted/review 统计待审稿章数。前端 `pnpm gen:api` 后项目库默认按最近编辑排序，新增“待审稿”筛选，作品卡显示最近编辑与待审稿徽标。契约记 `memory/contracts.md` C3 扩，UI 计划记 `docs/design/ui-improvement-plan.md` 0.16。
 - [2026-06-29] @frontend/@qa — **UI review follow-up 修复完成**：补 `lib/review/chapterNav.ts` 使新增章导航测试可运行；`ReviewSectionPanel` 改为本地 open 状态，避免用户折叠后重渲染强制弹开；`ReviewReport.hasReport` 纳入 clean review；新增 `tests/test_project_metadata_e2e.py` 用真实 DB 覆盖项目最近编辑/待审稿统计口径。验证：前端 lint/typecheck/vitest **490 passed**/build；后端 ruff/format/mypy、pytest **763 passed**；浏览器抽样 review/settings 无 console error/横向溢出。
+- [2026-07-06] @frontend — **✅ 灵感⑥ 关系图谱（缩减版）+ D4 完成**（纯前端，无后端契约/迁移/gen:api）。新增 `lib/characters/graphLayout.ts`（纯逻辑：`roleStyle` role→分组配色关键词映射、`buildGraphNodes`/`buildGraphEdges`(kind→标签，`DEFAULT_RELATION_LABEL` 命名常量)、`countDanglingRelations`、`layoutGraph`(@dagrejs/dagre 一次性确定性初始坐标、中心→左上、不可变)、`buildRelationshipGraph` 端到端）——**仅用现有 `CharacterCardView.relations={name,kind,note}`**：边按 name-join、kind 作中文标签、节点按 role 上色，无 closeness/阵营（数据不存在，不硬造）；悬空边（引用未入库角色）跳过并计数。新增 `components/characters/RelationshipGraph.tsx`（`'use client'`，React Flow @xyflow/react；CodexPage 经 `dynamic(ssr:false)` 挂载于人物 tab，显式容器高度 h-[28rem] + `import '@xyflow/react/dist/style.css'` + prefers-reduced-motion 关 fitView 动画 + 纸感 CSS 变量配色 + 悬空计数提示）。依赖：`@xyflow/react@^12.11.2`+`@dagrejs/dagre@^3.0.0`（见 contracts.md 2026-07-06）。TDD：graphLayout **+22 测**（role 映射/节点去空重名/kind 标签+默认/悬空跳过/自指跳过/无向去重/边 id 唯一/悬空计数/布局确定性+不可变/端到端）；组件走人工/E2E。前端门禁全绿：lint 干净 · typecheck 干净 · **vitest 565 passed** · **coverage lib 95.4%（graphLayout.ts 97.6%）** · **build OK**（codex 路由 7.58 kB，React Flow 懒加载独立 chunk）。注：`pnpm build` 门禁与运行中的 `next dev`(3001) 共享 `.next` → 触发 dev 服务临时 500（stale chunk），非代码缺陷、dev watcher 下次重编自愈，未重启服务。

apps/api/tests/fakes_projects.py

@@ -55,6 +55,9 @@ class FakeProjectRepo:
             theme=data.theme,
             selling_points=list(data.selling_points),
             structure=data.structure,
+            tone=data.tone,
+            ending_type=data.ending_type,
+            narrative_pov=data.narrative_pov,
             updated_at=self.now,
         )
         self.rows[pid] = (owner_id, view)
@@ -148,6 +151,7 @@ class FakeReviewRepo:
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> ReviewView:
         view = ReviewView(
@@ -159,6 +163,7 @@ class FakeReviewRepo:
             foreshadow_sug=[dict(s) for s in (foreshadow_sug or [])],
             style=style,
             pace=pace,
+            characterization=characterization,
             health_score=health_score,
         )
         # 新→旧：插到列表头部。

apps/api/tests/test_chain.py

@@ -115,6 +115,7 @@ class _FakeReviewRepo:
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> Any:
         rec = _FakeReviewRecord(chapter_no, list(conflicts))

apps/api/tests/test_generation.py

@@ -75,6 +75,8 @@ class _FakeCharacterWriteRepo:
         name: str,
         role: str,
         traits: list[str],
+        appearance: str,
+        motive: str,
         backstory: str,
         arc: str,
         speech_tics: list[str],
@@ -91,6 +93,8 @@ class _FakeCharacterWriteRepo:
                 {
                     "role": role,
                     "traits": list(traits),
+                    "appearance": appearance,
+                    "motive": motive,
                     "arc": arc,
                     "speech_tics": list(speech_tics),
                     "tags": list(tags),
@@ -106,6 +110,8 @@ class _FakeCharacterWriteRepo:
                 "name": name,
                 "role": role,
                 "traits": list(traits),
+                "appearance": appearance,
+                "motive": motive,
                 "arc": arc,
                 "speech_tics": list(speech_tics),
                 "tags": list(tags),
@@ -463,6 +469,62 @@ async def test_ingest_relations_persist_on_write() -> None:
     assert char_repo.rows[0]["relations"] == relations
 
 
+@pytest.mark.asyncio
+async def test_ingest_persists_appearance_and_motive() -> None:
+    # ⑧ 穿线：入库端点须把 appearance/motive 透传给写侧 repo（否则新列永远 NULL）。
+    repo = FakeProjectRepo()
+    pid = await _seed_project(repo)
+    gateway = _SchemaRoutingGateway({ContinuityReview: _no_conflicts()})
+    char_repo = _FakeCharacterWriteRepo()
+    app, _ = _make_app(project_repo=repo, gateway=gateway, char_repo=char_repo)
+
+    async with _client(app) as client:
+        resp = await client.post(
+            f"/projects/{pid}/characters",
+            json={
+                "cards": [
+                    {
+                        "name": "苏璃",
+                        "role": "CP",
+                        "traits": ["清冷毒舌"],
+                        "appearance": "银发赤瞳，玄色劲装",
+                        "motive": "替亡兄查清族灭真相",
+                        "backstory": "玄霜阁遗孤",
+                        "arc": "由敌转盟",
+                        "speech_tics": [],
+                        "tags": [],
+                        "relations": [],
+                    }
+                ]
+            },
+        )
+
+    assert resp.status_code == 201
+    assert char_repo.rows[0]["appearance"] == "银发赤瞳，玄色劲装"
+    assert char_repo.rows[0]["motive"] == "替亡兄查清族灭真相"
+
+
+def test_card_to_view_roundtrips_new_fields() -> None:
+    # ⑧ 穿线：schema↔view 双向形变须带上 appearance/motive（否则前后端静默丢字段）。
+    from ww_api.routers.generation import _card_to_view, _view_to_card
+
+    card = CharacterCard(
+        name="苏璃",
+        role="CP",
+        traits=["清冷毒舌"],
+        appearance="银发赤瞳，玄色劲装",
+        motive="替亡兄查清族灭真相",
+        backstory="玄霜阁遗孤",
+        arc="由敌转盟",
+    )
+    view = _card_to_view(card)
+    assert view.appearance == "银发赤瞳，玄色劲装"
+    assert view.motive == "替亡兄查清族灭真相"
+    back = _view_to_card(view)
+    assert back.appearance == card.appearance
+    assert back.motive == card.motive
+
+
 def test_relations_from_jsonb_parses_and_skips_dirty() -> None:
     # Arrange：混入脏条目（非 dict / 缺 name / 缺 kind）。
     from ww_api.routers.generation import _relations_from_jsonb
@@ -572,6 +634,8 @@ def _codex_memory() -> Any:
                     name="叶寒",
                     role="主角",
                     traits={"items": ["腹黑", "护短"]},
+                    appearance="剑眉星目，一袭青衫",
+                    motive="为母复仇、登临帝位",
                     backstory="孤儿出身",
                     arc={"text": "从孤儿到帝王"},
                     speech_tics={"items": ["哼"]},
@@ -668,6 +732,9 @@ async def test_list_characters_unpacks_jsonb_to_api_shape() -> None:
     assert card["speech_tics"] == ["哼"]
     assert card["arc"] == "从孤儿到帝王"
     assert card["tags"] == ["天才"]
+    # ⑧ 穿线：读端点暴露 appearance/motive（Text 列直落，不形变）。
+    assert card["appearance"] == "剑眉星目，一袭青衫"
+    assert card["motive"] == "为母复仇、登临帝位"
 
 
 @pytest.mark.asyncio

apps/api/tests/test_generation_schema.py

@@ -0,0 +1,51 @@
+"""CharacterGenerateRequest.role_mix 校验单测（⑦ 群像按定位配比）。
+
+纯 pydantic 校验，无 DB/网关。契约：role_mix 在场时 count 取各值之和，
+每项数量须 ≥ 1、总和须 ≤ MAX_GENERATE_COUNT。
+"""
+
+from __future__ import annotations
+
+import pytest
+from pydantic import ValidationError
+from ww_api.schemas.generation import MAX_GENERATE_COUNT, CharacterGenerateRequest
+
+
+def test_role_mix_absent_keeps_explicit_count() -> None:
+    req = CharacterGenerateRequest(brief="剑修", count=3)
+    assert req.role_mix is None
+    assert req.count == 3
+
+
+def test_role_mix_sets_count_to_sum() -> None:
+    req = CharacterGenerateRequest(brief="群像", role_mix={"主角": 1, "对手": 2, "配角": 3})
+    assert req.count == 6
+
+
+def test_role_mix_overrides_provided_count() -> None:
+    # role_mix 在场时 count 取各值之和，忽略客户端显式传入的 count
+    req = CharacterGenerateRequest(brief="群像", count=1, role_mix={"主角": 2, "对手": 2})
+    assert req.count == 4
+
+
+def test_role_mix_total_at_max_allowed() -> None:
+    req = CharacterGenerateRequest(brief="群像", role_mix={"主角": MAX_GENERATE_COUNT})
+    assert req.count == MAX_GENERATE_COUNT
+
+
+def test_role_mix_sum_over_max_rejected() -> None:
+    with pytest.raises(ValidationError):
+        # 6 + 7 = 13 > 12
+        CharacterGenerateRequest(brief="群像", role_mix={"主角": 6, "对手": 7})
+
+
+def test_role_mix_empty_dict_rejected() -> None:
+    with pytest.raises(ValidationError):
+        CharacterGenerateRequest(brief="群像", role_mix={})
+
+
+def test_role_mix_nonpositive_value_rejected() -> None:
+    with pytest.raises(ValidationError):
+        CharacterGenerateRequest(brief="群像", role_mix={"主角": 0})
+    with pytest.raises(ValidationError):
+        CharacterGenerateRequest(brief="群像", role_mix={"主角": -1})

apps/api/tests/test_project_plan_endpoint.py

@@ -0,0 +1,197 @@
+"""AI 立项方案生成端点测试（灵感⑤；内存替身，无 DB/无网络）。
+
+覆盖 `POST /skills/project-plan/generate`（不带 project 前缀）：
+- 无既有 project 也能生成（向导阶段 project 未建）→ 200 结构化预览 + 落 ledger。
+- 种子门控：缺 genre / logline → 422，不触达网关、不 commit。
+- 无凭据 → 503。
+- 向导草稿序列化进网关 context（种子 + 已填字段透传）。
+LLM 一律 mock，绝不联网。
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any
+
+import httpx
+import pytest
+from cryptography.fernet import Fernet
+from fakes_projects import FakeSession
+from ww_agents import ProjectPlanResult
+from ww_llm_gateway.types import LlmRequest, LlmResponse, ServedBy, Usage
+from ww_shared import AppError, ErrorCode
+
+
+class _CaptureGateway:
+    """记录最近一次 req.input（验证向导草稿进了 context），返固定 ProjectPlanResult。"""
+
+    def __init__(self, parsed: ProjectPlanResult) -> None:
+        self._parsed = parsed
+        self.last_input: str | None = None
+        self.calls = 0
+
+    async def run(self, req: LlmRequest) -> LlmResponse:
+        self.calls += 1
+        self.last_input = req.input if isinstance(req.input, str) else str(req.input)
+        return LlmResponse(
+            text=self._parsed.model_dump_json(),
+            parsed=self._parsed,
+            usage=Usage(
+                provider="fake",
+                model="fake",
+                input_tokens=1,
+                output_tokens=1,
+                cost_minor=0,
+                currency="USD",
+            ),
+            served_by=ServedBy(provider="fake", model="fake"),
+        )
+
+
+def _plan() -> ProjectPlanResult:
+    return ProjectPlanResult(
+        title_candidates=["逐光而行", "剑试九霄"],
+        setting="上古仙门世界",
+        narrative_structure="黄金三章立钩 → 升级",
+        story_core="废柴逆袭封神的代价",
+        ending_design="归隐收束（正剧）",
+        tone="热血",
+    )
+
+
+def _make_app(
+    *,
+    gateway: Any,
+    session: FakeSession | None = None,
+    no_creds: bool = False,
+) -> tuple[Any, FakeSession]:
+    os.environ.setdefault("CREDENTIAL_ENC_KEY", Fernet.generate_key().decode())
+    from ww_api.main import create_app
+    from ww_api.services.project_deps import get_project_plan_gateway
+    from ww_db import get_session
+
+    session = session or FakeSession()
+
+    async def _ok() -> Any:
+        return gateway
+
+    async def _no_creds() -> Any:
+        raise AppError(ErrorCode.LLM_UNAVAILABLE, "未配置凭据", {"provider": "deepseek"})
+
+    app = create_app()
+    app.dependency_overrides[get_project_plan_gateway] = _no_creds if no_creds else _ok
+    app.dependency_overrides[get_session] = lambda: session
+    return app, session
+
+
+def _client(app: Any) -> httpx.AsyncClient:
+    transport = httpx.ASGITransport(app=app, raise_app_exceptions=False)
+    return httpx.AsyncClient(transport=transport, base_url="http://test")
+
+
+@pytest.mark.asyncio
+async def test_generate_without_existing_project() -> None:
+    # 向导阶段 project 未建：无 project 前缀端点直接生成，返结构化方案 + 落 ledger。
+    gateway = _CaptureGateway(_plan())
+    app, session = _make_app(gateway=gateway)
+
+    async with _client(app) as client:
+        resp = await client.post(
+            "/skills/project-plan/generate",
+            json={"genre": "仙侠", "logline": "废柴少年觉醒禁忌血脉，逆势封神。"},
+        )
+
+    assert resp.status_code == 200
+    body = resp.json()
+    assert body["title_candidates"] == ["逐光而行", "剑试九霄"]
+    assert body["story_core"] == "废柴逆袭封神的代价"
+    assert body["ending_design"].endswith("（正剧）")
+    assert body["tone"] == "热血"
+    assert session.commits == 1  # 预览不写业务表，但落 ledger
+
+
+@pytest.mark.asyncio
+async def test_generate_serializes_wizard_draft_into_context() -> None:
+    # 种子 + 已填字段序列化进网关 context（有则贴合）。
+    gateway = _CaptureGateway(_plan())
+    app, _ = _make_app(gateway=gateway)
+
+    async with _client(app) as client:
+        resp = await client.post(
+            "/skills/project-plan/generate",
+            json={
+                "genre": "都市",
+                "logline": "外卖骑手觉醒读心术。",
+                "title": "听风者",
+                "tone": "轻松",
+                "selling_points": ["逆袭", "系统流"],
+                "brief": "偏悬疑向",
+            },
+        )
+
+    assert resp.status_code == 200
+    assert gateway.last_input is not None
+    assert "题材：都市" in gateway.last_input
+    assert "一句话故事：外卖骑手觉醒读心术。" in gateway.last_input
+    assert "暂定书名：听风者" in gateway.last_input
+    assert "核心卖点：逆袭、系统流" in gateway.last_input
+    assert "偏悬疑向" in gateway.last_input  # brief 进 context
+
+
+@pytest.mark.asyncio
+async def test_requires_seed_missing_logline_422() -> None:
+    # 种子门控：缺 logline → 422，不触达网关、不 commit。
+    gateway = _CaptureGateway(_plan())
+    app, session = _make_app(gateway=gateway)
+
+    async with _client(app) as client:
+        resp = await client.post("/skills/project-plan/generate", json={"genre": "仙侠"})
+
+    assert resp.status_code == 422
+    assert resp.json()["error"]["code"] == ErrorCode.VALIDATION
+    assert gateway.calls == 0
+    assert session.commits == 0
+
+
+@pytest.mark.asyncio
+async def test_requires_seed_missing_genre_422() -> None:
+    gateway = _CaptureGateway(_plan())
+    app, session = _make_app(gateway=gateway)
+
+    async with _client(app) as client:
+        resp = await client.post("/skills/project-plan/generate", json={"logline": "一句话故事。"})
+
+    assert resp.status_code == 422
+    assert gateway.calls == 0
+    assert session.commits == 0
+
+
+@pytest.mark.asyncio
+async def test_requires_seed_blank_strings_422() -> None:
+    # 纯空白种子也拒（strip 后为空）。
+    gateway = _CaptureGateway(_plan())
+    app, session = _make_app(gateway=gateway)
+
+    async with _client(app) as client:
+        resp = await client.post(
+            "/skills/project-plan/generate", json={"genre": "   ", "logline": "  "}
+        )
+
+    assert resp.status_code == 422
+    assert gateway.calls == 0
+    assert session.commits == 0
+
+
+@pytest.mark.asyncio
+async def test_generate_no_credentials_503() -> None:
+    app, session = _make_app(gateway=object(), no_creds=True)
+
+    async with _client(app) as client:
+        resp = await client.post(
+            "/skills/project-plan/generate",
+            json={"genre": "仙侠", "logline": "一句话故事。"},
+        )
+
+    assert resp.status_code == 503
+    assert resp.json()["error"]["code"] == ErrorCode.LLM_UNAVAILABLE
+    assert session.commits == 0

apps/api/tests/test_projects.py

@@ -150,6 +150,54 @@ async def test_create_project_returns_201() -> None:
     assert uuid.UUID(body["id"])
 
 
+@pytest.mark.asyncio
+async def test_create_project_persists_tone_ending_pov() -> None:
+    # ④ 立项新增基调/结局/视角：请求 → repo 落库 → 响应回显（snake_case 契约）。
+    client, repo, _, _ = _make_client()
+    async with client:
+        resp = await client.post(
+            "/projects",
+            json={
+                "title": "基调作品",
+                "tone": "热血",
+                "ending_type": "HE",
+                "narrative_pov": "第一人称",
+            },
+        )
+    assert resp.status_code == 201
+    body = resp.json()
+    assert body["tone"] == "热血"
+    assert body["ending_type"] == "HE"
+    assert body["narrative_pov"] == "第一人称"
+    (_owner, view) = next(iter(repo.rows.values()))
+    assert view.tone == "热血"
+    assert view.ending_type == "HE"
+    assert view.narrative_pov == "第一人称"
+
+
+@pytest.mark.asyncio
+async def test_get_project_exposes_tone_ending_pov() -> None:
+    client, _, _, _ = _make_client()
+    async with client:
+        created = (
+            await client.post(
+                "/projects",
+                json={
+                    "title": "书",
+                    "tone": "沉重",
+                    "ending_type": "BE",
+                    "narrative_pov": "多视角",
+                },
+            )
+        ).json()
+        resp = await client.get(f"/projects/{created['id']}")
+    assert resp.status_code == 200
+    body = resp.json()
+    assert body["tone"] == "沉重"
+    assert body["ending_type"] == "BE"
+    assert body["narrative_pov"] == "多视角"
+
+
 @pytest.mark.asyncio
 async def test_list_projects() -> None:
     client, repo, _, _ = _make_client()

apps/api/tests/test_review_accept.py

@@ -209,6 +209,70 @@ async def test_accept_promotes_appends_digest_records_decisions() -> None:
     assert session.commits == 1
 
 
+@pytest.mark.asyncio
+async def test_characterization_does_not_block_accept() -> None:
+    """灵感③ advisory：人物塑造问题（零 continuity 冲突）不阻断验收（D2）。
+
+    留痕带 characterization issues 但 conflicts=[] → 验收 gate 只读 continuity conflicts，
+    直通晋升；characterization 从不进 conflicts、从不产 decisions。
+    """
+    from ww_api.services.digest_extraction import ChapterDigestFacts
+
+    review_repo = FakeReviewRepo()
+    pid = uuid.uuid4()
+    # 零 continuity 冲突，但有 advisory 人物塑造问题。
+    await review_repo.record(
+        pid,
+        1,
+        conflicts=[],
+        characterization={
+            "issues": [
+                {
+                    "character": "沈砚",
+                    "aspect": "动机一致性",
+                    "where": "第3段",
+                    "quote": "他忽然转身就走。",
+                    "diagnosis": "与既定护妹动机断裂",
+                    "suggestion": "退缩扣回护妹动机",
+                    "confidence": 0.82,
+                }
+            ]
+        },
+    )
+    digest_gw = FakeReviewGateway(parsed=ChapterDigestFacts(summary="s"))
+    client, chapter_repo, _, digest_repo, session = _make_client(
+        review_repo=review_repo, digest_gateway=digest_gw
+    )
+    async with client:
+        resp = await client.post(
+            f"/projects/{pid}/chapters/1/accept",
+            # 无需对人物塑造做任何裁决（advisory 无 conflict 通道）。
+            json={"final_text": "作者终稿正文。", "decisions": []},
+        )
+    assert resp.status_code == 200
+    assert resp.json()["accepted_version"] == 1
+    assert session.commits == 1
+    # characterization 留痕不产 decisions（advisory）；晋升照常。
+    accepted = await chapter_repo.latest_accepted(pid, 1)
+    assert accepted is not None
+
+
+@pytest.mark.asyncio
+async def test_characterization_surfaces_in_review_history() -> None:
+    """列 characterization 经 GET .../reviews 透传到历史项（前端面板 seed 用）。"""
+    review_repo = FakeReviewRepo()
+    pid = uuid.uuid4()
+    await review_repo.record(
+        pid, 1, conflicts=[], characterization={"issues": [{"character": "沈砚"}]}
+    )
+    client, _, _, _, _ = _make_client(review_repo=review_repo)
+    async with client:
+        resp = await client.get(f"/projects/{pid}/chapters/1/reviews")
+    assert resp.status_code == 200
+    row = resp.json()["reviews"][0]
+    assert row["characterization"]["issues"][0]["character"] == "沈砚"
+
+
 @pytest.mark.asyncio
 async def test_accept_with_no_review_succeeds() -> None:
     """无审稿留痕（零冲突）→ gate 直通，仍晋升 + 提炼 digest。"""

apps/api/ww_api/routers/generation.py

@@ -26,8 +26,10 @@ from sqlalchemy.ext.asyncio import AsyncSession
 from ww_agents import (
     CharacterCard,
     CharacterRelation,
+    ProjectPlanResult,
     character_gen_spec,
     continuity_spec,
+    project_plan_spec,
     worldbuilder_spec,
 )
 from ww_core.domain import (
@@ -37,8 +39,10 @@ from ww_core.domain import (
 )
 from ww_core.domain.repositories import MemoryRepos
 from ww_core.orchestrator import (
+    build_brief_context,
     precheck_generated_cards,
     run_character_gen,
+    run_generator,
     run_worldbuilder,
 )
 from ww_db import get_session
@@ -64,6 +68,7 @@ from ww_api.schemas.generation import (
     WorldGenerateRequest,
     WorldGenPreviewResponse,
 )
+from ww_api.schemas.projects import ProjectPlanGenerateRequest, ProjectPlanView
 from ww_api.schemas.rules import RuleView
 from ww_api.services.credentials import STUB_OWNER_ID
 from ww_api.services.project_deps import (
@@ -71,6 +76,7 @@ from ww_api.services.project_deps import (
     get_character_write_repo,
     get_memory_repos,
     get_precheck_gateway,
+    get_project_plan_gateway,
     get_project_repo,
     get_rule_write_repo,
     get_skill_registry,
@@ -90,6 +96,7 @@ SkillRegistryDep = Annotated[SkillRegistry, Depends(get_skill_registry)]
 WorldGatewayDep = Annotated[Gateway, Depends(get_worldbuilder_gateway)]
 CharacterGatewayDep = Annotated[Gateway, Depends(get_character_gen_gateway)]
 PrecheckGatewayDep = Annotated[Gateway, Depends(get_precheck_gateway)]
+ProjectPlanGatewayDep = Annotated[Gateway, Depends(get_project_plan_gateway)]
 SessionDep = Annotated[AsyncSession, Depends(get_session)]
 
 
@@ -105,6 +112,46 @@ def _project_context(title: str, genre: str | None, premise: str | None, theme:
     return "\n".join(lines)
 
 
+def _plan_seed_context(body: ProjectPlanGenerateRequest) -> str:
+    """把立项向导草稿序列化为作品种子（喂 project-plan；仿 `_project_context`）。
+
+    确定性、无时间戳/UUID；只拼非空字段（有则贴合、无则不臆造）。
+    genre + logline 由端点种子门控保证非空。
+    """
+    lines: list[str] = []
+    label_fields: list[tuple[str, str | None]] = [
+        ("暂定书名", body.title),
+        ("题材", body.genre),
+        ("一句话故事", body.logline),
+        ("立意", body.premise),
+        ("主题", body.theme),
+        ("故事结构", body.structure),
+        ("基调", body.tone),
+        ("结局取向", body.ending_type),
+        ("叙事视角", body.narrative_pov),
+    ]
+    for label, value in label_fields:
+        text = (value or "").strip()
+        if text:
+            lines.append(f"{label}：{text}")
+    points = [p.strip() for p in body.selling_points if p.strip()]
+    if points:
+        lines.append(f"核心卖点：{'、'.join(points)}")
+    return "\n".join(lines)
+
+
+def _plan_to_view(plan: ProjectPlanResult) -> ProjectPlanView:
+    """ww_agents.ProjectPlanResult → API ProjectPlanView（字段同名直传）。"""
+    return ProjectPlanView(
+        title_candidates=list(plan.title_candidates),
+        setting=plan.setting,
+        narrative_structure=plan.narrative_structure,
+        story_core=plan.story_core,
+        ending_design=plan.ending_design,
+        tone=plan.tone,
+    )
+
+
 def _render_characters_context(cards: list[CharacterCard]) -> str:
     """已有角色简表（喂 character-gen 防雷同 + precheck 真相源）。确定性、保序。"""
     if not cards:
@@ -148,6 +195,8 @@ async def _existing_characters(memory: MemoryRepos, project_id: uuid.UUID) -> li
                 name=v.name,
                 role=v.role or "",
                 traits=traits,
+                appearance=v.appearance or "",
+                motive=v.motive or "",
                 backstory=v.backstory or "",
                 arc=arc or "",
                 speech_tics=tics,
@@ -244,6 +293,7 @@ async def generate_characters(
         gateway=gateway,
         user_id=STUB_OWNER_ID,
         project_id=project_id,
+        role_mix=body.role_mix,
     )
     await session.commit()
     log.info(
@@ -329,6 +379,8 @@ async def ingest_characters(
             name=card.name,
             role=card.role,
             traits=list(card.traits),
+            appearance=card.appearance,
+            motive=card.motive,
             backstory=card.backstory,
             arc=card.arc,
             speech_tics=list(card.speech_tics),
@@ -426,6 +478,53 @@ async def list_skills(registry: SkillRegistryDep) -> SkillListResponse:
     return SkillListResponse(skills=skills)
 
 
+# ---- AI 立项方案生成（灵感⑤；不带 project 前缀，立项前 project 未建）----
+
+
+@skills_router.post("/project-plan/generate")
+async def generate_project_plan(
+    body: ProjectPlanGenerateRequest,
+    request: Request,
+    gateway: ProjectPlanGatewayDep,
+    session: SessionDep,
+) -> ProjectPlanView:
+    """生成结构化立项方案预览（向导阶段，project 未建，不入库）。
+
+    **不带 project 前缀**——绕开 toolbox 的 `project_repo.get→404`（立项前无 project）。
+    **种子门控**：缺 genre 或 logline → 422（空向导不出方案，防 slop）。无凭据 → 503。
+    """
+    request_id = getattr(request.state, "request_id", None)
+    genre = (body.genre or "").strip()
+    logline = (body.logline or "").strip()
+    if not genre or not logline:
+        raise AppError(
+            ErrorCode.VALIDATION,
+            "生成立项方案至少需要「题材」与「一句话故事（logline）」两个种子",
+            {"has_genre": bool(genre), "has_logline": bool(logline)},
+        )
+
+    project_context = _plan_seed_context(body)
+    context = build_brief_context(brief=body.brief or "", project_context=project_context)
+    parsed = await run_generator(
+        project_plan_spec,
+        context=context,
+        gateway=gateway,
+        user_id=STUB_OWNER_ID,
+        project_id=None,  # 向导阶段 project 未建；usage_ledger.project_id 可空
+    )
+    # 提交边界：预览不写业务表，但网关 ledger 只 flush → 末尾 commit 落 usage（否则丢失）。
+    await session.commit()
+    if not isinstance(parsed, ProjectPlanResult):
+        # run_generator 已按 output_schema 校验；防御性收窄（不应触达）。
+        raise AppError(ErrorCode.INTERNAL, "立项方案生成返回类型异常")
+    log.info(
+        "project_plan_generate_done",
+        request_id=request_id,
+        title_count=len(parsed.title_candidates),
+    )
+    return _plan_to_view(parsed)
+
+
 # ---- schema <-> ww_agents 卡转换 ----
 
 
@@ -434,6 +533,8 @@ def _card_to_view(card: CharacterCard) -> CharacterCardView:
         name=card.name,
         role=card.role,
         traits=list(card.traits),
+        appearance=card.appearance,
+        motive=card.motive,
         backstory=card.backstory,
         arc=card.arc,
         speech_tics=list(card.speech_tics),
@@ -449,6 +550,8 @@ def _view_to_card(view: CharacterCardView) -> CharacterCard:
         name=view.name,
         role=view.role,
         traits=list(view.traits),
+        appearance=view.appearance,
+        motive=view.motive,
         backstory=view.backstory,
         arc=view.arc,
         speech_tics=list(view.speech_tics),

apps/api/ww_api/routers/projects.py

@@ -127,6 +127,9 @@ async def create_project(body: ProjectCreateRequest, repo: ProjectRepoDep) -> Pr
             theme=body.theme,
             selling_points=body.selling_points,
             structure=body.structure,
+            tone=body.tone,
+            ending_type=body.ending_type,
+            narrative_pov=body.narrative_pov,
         ),
     )
     log.info("project_created", project_id=str(view.id), title_len=len(view.title))
@@ -485,6 +488,7 @@ async def list_reviews(
             foreshadow_sug=v.foreshadow_sug,
             style=v.style,
             pace=v.pace,
+            characterization=v.characterization,
             health_score=v.health_score,
             decisions=v.decisions,
         )

apps/api/ww_api/schemas/generation.py

@@ -12,7 +12,7 @@ snake_case；前端经 OpenAPI 生成 TS 类型消费。改字段 → 前端必
 
 from __future__ import annotations
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, model_validator
 
 from ww_api.schemas.rules import RuleView
 
@@ -47,11 +47,40 @@ class WorldGenPreviewResponse(BaseModel):
 
 
 class CharacterGenerateRequest(BaseModel):
-    """POST /projects/:id/characters/generate：据需求生成角色卡（预览，群像防雷同）。"""
+    """POST /projects/:id/characters/generate：据需求生成角色卡（预览，群像防雷同）。
+
+    `role_mix`（⑦ 群像按定位配比）与 `count` 的关系：`role_mix` 在场时 `count` 取各值之和
+    （忽略客户端显式 count），每项数量须 ≥ 1、总和须 ≤ `MAX_GENERATE_COUNT`；缺省则退回
+    单一 `role` + `count` 语义。
+    """
 
     brief: str = Field(min_length=1, description="角色生成需求")
-    count: int = Field(default=1, ge=1, le=MAX_GENERATE_COUNT, description="生成数量")
+    count: int = Field(
+        default=1,
+        ge=1,
+        le=MAX_GENERATE_COUNT,
+        description="生成数量（role_mix 在场时取各值之和）",
+    )
     role: str | None = Field(default=None, description="角色定位（主角/CP/对手/导师/工具人 等）")
+    role_mix: dict[str, int] | None = Field(
+        default=None,
+        description="按定位配比生成 {定位: 数量}；在场时 count=各值之和（每项≥1、总和≤12）",
+    )
+
+    @model_validator(mode="after")
+    def _reconcile_role_mix(self) -> CharacterGenerateRequest:
+        """配比在场时校验并把 count 定为各值之和（构造期归一化，非外部对象变更）。"""
+        if self.role_mix is None:
+            return self
+        if not self.role_mix:
+            raise ValueError("role_mix 不能为空对象")
+        if any(n < 1 for n in self.role_mix.values()):
+            raise ValueError("role_mix 每个定位的数量须 ≥ 1")
+        total = sum(self.role_mix.values())
+        if total > MAX_GENERATE_COUNT:
+            raise ValueError(f"role_mix 数量总和须 ≤ {MAX_GENERATE_COUNT}")
+        self.count = total
+        return self
 
 
 class CharacterRelationView(BaseModel):
@@ -67,7 +96,9 @@ class CharacterCardView(BaseModel):
 
     name: str = Field(description="角色名")
     role: str = Field(description="角色定位")
-    traits: list[str] = Field(default_factory=list, description="性格特质清单")
+    traits: list[str] = Field(default_factory=list, description="性格特质清单（性格三层）")
+    appearance: str = Field(default="", description="外貌形象（缺则空串）")
+    motive: str = Field(default="", description="核心动机/欲望/恐惧（缺则空串）")
     backstory: str = Field(description="背景故事")
     arc: str = Field(description="人物弧光（一句话）")
     speech_tics: list[str] = Field(default_factory=list, description="口癖/语言风格")

apps/api/ww_api/schemas/projects.py

@@ -22,6 +22,10 @@ class ProjectCreateRequest(BaseModel):
     theme: str | None = None
     selling_points: list[str] = Field(default_factory=list)
     structure: str | None = None
+    # 立项书级常量（灵感④）——free text + 前端预设，无 Literal（Literal 只留 Verdict）。
+    tone: str | None = None
+    ending_type: str | None = None
+    narrative_pov: str | None = None
 
 
 class ProjectResponse(BaseModel):
@@ -35,6 +39,9 @@ class ProjectResponse(BaseModel):
     theme: str | None = None
     selling_points: list[str] = Field(default_factory=list)
     structure: str | None = None
+    tone: str | None = None
+    ending_type: str | None = None
+    narrative_pov: str | None = None
     updated_at: datetime | None = Field(default=None, description="项目最近更新时间")
     pending_review_count: int = Field(default=0, ge=0, description="待审稿章节数")
 
@@ -45,6 +52,64 @@ class ProjectListResponse(BaseModel):
     projects: list[ProjectResponse] = Field(default_factory=list)
 
 
+# ---- AI 立项方案生成（灵感⑤；向导阶段 project 未建，纯预览不写库）----
+
+# 请求字段长度上界（CR-H9：防超长向导草稿撑爆上下文/成本）。短字段=标签级，长字段=段落级。
+_PLAN_SHORT_MAX = 200
+_PLAN_LONG_MAX = 4000
+_PLAN_LIST_MAX = 20
+
+
+class ProjectPlanGenerateRequest(BaseModel):
+    """POST /skills/project-plan/generate：向导草稿种子（不带 project 前缀，立项前 project 未建）。
+
+    **种子门控**：`genre` + `logline` 为必给种子（端点校验非空，否则 422）——空向导不出方案，
+    防 slop。其余字段有则贴合、无则不臆造（序列化进 project_context，仿 `_project_context`）。
+    全字段带 `max_length` 上界（CR-H9），`brief` = 作者一句话方向（可空）。
+    """
+
+    genre: str | None = Field(default=None, max_length=_PLAN_SHORT_MAX, description="题材（种子）")
+    logline: str | None = Field(
+        default=None, max_length=_PLAN_LONG_MAX, description="一句话故事（种子）"
+    )
+    title: str | None = Field(
+        default=None, max_length=_PLAN_SHORT_MAX, description="暂定书名（可空）"
+    )
+    premise: str | None = Field(default=None, max_length=_PLAN_LONG_MAX, description="立意（可空）")
+    theme: str | None = Field(default=None, max_length=_PLAN_SHORT_MAX, description="主题（可空）")
+    structure: str | None = Field(
+        default=None, max_length=_PLAN_SHORT_MAX, description="故事结构（可空）"
+    )
+    tone: str | None = Field(default=None, max_length=_PLAN_SHORT_MAX, description="基调（可空）")
+    ending_type: str | None = Field(
+        default=None, max_length=_PLAN_SHORT_MAX, description="结局取向（可空）"
+    )
+    narrative_pov: str | None = Field(
+        default=None, max_length=_PLAN_SHORT_MAX, description="叙事视角（可空）"
+    )
+    selling_points: list[str] = Field(
+        default_factory=list, max_length=_PLAN_LIST_MAX, description="核心卖点（可空）"
+    )
+    # `str | None`（非 `str=""`）：codegen 渲染为可选字段，前端种子请求可省略（brief 本就可空）。
+    brief: str | None = Field(
+        default=None, max_length=_PLAN_LONG_MAX, description="作者一句话方向/需求（可空）"
+    )
+
+
+class ProjectPlanView(BaseModel):
+    """立项方案生成结果（贴 ww_agents.ProjectPlanResult；结构化预览，不入库）。
+
+    全字段有默认值（守解析韧性）——前端按字段回填立项向导，只填非空交集、不覆盖作者已编辑值。
+    """
+
+    title_candidates: list[str] = Field(default_factory=list, description="书名候选清单")
+    setting: str = Field(default="", description="时空背景")
+    narrative_structure: str = Field(default="", description="叙事结构")
+    story_core: str = Field(default="", description="故事核心")
+    ending_design: str = Field(default="", description="结局设计")
+    tone: str = Field(default="", description="基调")
+
+
 class DraftStreamRequest(BaseModel):
     """POST /projects/:id/chapters/:no/draft：本章生成的可选输入（T4-b）。
 
@@ -126,6 +191,8 @@ class ReviewHistoryItem(BaseModel):
     foreshadow_sug: list[dict[str, Any]] = Field(default_factory=list)
     style: dict[str, Any] | None = None
     pace: dict[str, Any] | None = None
+    # 灵感③ advisory 人物塑造审留痕（issues 列表 dict）；仅建议、不阻断验收。
+    characterization: dict[str, Any] | None = None
     health_score: int | None = None
     decisions: dict[str, Any] | None = None
 

apps/api/ww_api/services/project_deps.py

@@ -268,6 +268,14 @@ async def get_precheck_gateway(
     return await build_gateway_for_tier(session, store, "analyst")
 
 
+async def get_project_plan_gateway(
+    session: Annotated[AsyncSession, Depends(get_session)],
+) -> Gateway:
+    """AI 立项方案生成（analyst 档位）的可注入网关缝。测试注 mock（产 ProjectPlanResult）。"""
+    store = SqlCredentialStore(session)
+    return await build_gateway_for_tier(session, store, "analyst")
+
+
 def get_session_factory() -> SessionFactory:
     """验收后到期扫描的**独立 session 工厂**缝。
 

apps/web/components/ProjectWizard.tsx

@@ -1,9 +1,10 @@
 "use client";
 
-import { ArrowLeft, ArrowRight, Check, CheckCircle2 } from "lucide-react";
+import { ArrowLeft, ArrowRight, Check, CheckCircle2, Sparkles } from "lucide-react";
 import { useRouter } from "next/navigation";
 import { useEffect, useRef, useState } from "react";
 
+import { ThinkingIndicator } from "@/components/ThinkingIndicator";
 import { useToast } from "@/components/Toast";
 import { Badge } from "@/components/ui/Badge";
 import { Button } from "@/components/ui/Button";
@@ -14,17 +15,26 @@ import { TextArea } from "@/components/ui/TextArea";
 import { TextInput } from "@/components/ui/TextInput";
 import { api } from "@/lib/api/client";
 import { buttonClass } from "@/lib/ui/variants";
+import type { ProjectPlanView } from "@/lib/api/types";
+import { useProjectPlan } from "@/lib/wizard/useProjectPlan";
 import {
+  ENDING_TYPES,
   GENRES,
+  NARRATIVE_POVS,
   SELLING_POINT_PRESETS,
   STEP_TITLES,
   STRUCTURES,
+  TONES,
   WIZARD_STEPS,
+  applyPlanPatch,
   canAdvance,
+  canGeneratePlan,
   canSubmit,
   clampStep,
   emptyWizardForm,
+  mapPlanResultToWizardForm,
   toCreateRequest,
+  toPlanSeedRequest,
   type WizardForm,
 } from "@/lib/wizard/wizard";
 
@@ -58,6 +68,21 @@ export function ProjectWizard() {
         : [...prev.sellingPoints, point],
     }));
 
+  // AI 立项方案按字段回填：只填当前为空的字段（保护作者已编辑内容，不 bulk 覆盖）。
+  const applyPlan = (plan: ProjectPlanView): void => {
+    const patch = mapPlanResultToWizardForm(plan);
+    const filled = (Object.keys(patch) as Array<keyof WizardForm>).filter(
+      (k) => typeof form[k] === "string" && (form[k] as string).trim() === "",
+    );
+    setForm((prev) => applyPlanPatch(prev, patch));
+    toast(
+      filled.length > 0
+        ? `已回填 ${filled.length} 个空字段（已编辑内容不覆盖）`
+        : "向导相关字段均已填写，未覆盖任何内容",
+      filled.length > 0 ? "success" : "info",
+    );
+  };
+
   const goBack = (): void => setStep((s) => clampStep(s - 1));
   const goNext = (): void => setStep((s) => clampStep(s + 1));
 
@@ -107,6 +132,7 @@ export function ProjectWizard() {
             form={form}
             update={update}
             toggleSellingPoint={toggleSellingPoint}
+            onApplyPlan={applyPlan}
           />
         )}
         {step === 3 && <StepPremise form={form} update={update} />}
@@ -196,11 +222,91 @@ function StepBasics({ form, update }: StepProps) {
   );
 }
 
+// AI 立项方案助手（灵感⑤）：种子（题材+logline）就绪后一键生成结构化方案 → 预览 → 按字段回填。
+function PlanAssistant({
+  form,
+  onApplyPlan,
+}: {
+  form: WizardForm;
+  onApplyPlan: (plan: ProjectPlanView) => void;
+}) {
+  const { status, plan, generate } = useProjectPlan();
+  const ready = canGeneratePlan(form);
+  const generating = status === "generating";
+
+  return (
+    <div className="my-4 rounded border border-line bg-bg p-3">
+      <div className="flex items-center justify-between gap-3">
+        <div className="min-w-0">
+          <p className="flex items-center gap-1.5 text-sm text-ink">
+            <Sparkles className="h-4 w-4 text-cinnabar" aria-hidden="true" />
+            AI 立项方案
+          </p>
+          <p className="mt-0.5 text-xs text-ink-soft">
+            {ready
+              ? "根据题材 + 一句话故事推演书级蓝图，按字段回填（不覆盖已填内容）。"
+              : "填好「题材」（上一步）与「一句话故事」后即可生成。"}
+          </p>
+        </div>
+        <Button
+          onClick={() => void generate(toPlanSeedRequest(form))}
+          disabled={!ready || generating}
+          variant="secondary"
+        >
+          {generating ? (
+            <ThinkingIndicator label="生成中" />
+          ) : (
+            <>
+              <Sparkles className="h-4 w-4" aria-hidden="true" />
+              生成方案
+            </>
+          )}
+        </Button>
+      </div>
+
+      {status === "done" && plan && (
+        <div className="mt-3 border-t border-line pt-3">
+          <dl className="space-y-1.5 text-sm">
+            {plan.title_candidates && plan.title_candidates.length > 0 && (
+              <PlanRow label="书名候选" value={plan.title_candidates.join("、")} />
+            )}
+            <PlanRow label="时空背景" value={plan.setting} />
+            <PlanRow label="叙事结构" value={plan.narrative_structure} />
+            <PlanRow label="故事核心" value={plan.story_core} />
+            <PlanRow label="结局设计" value={plan.ending_design} />
+            <PlanRow label="基调" value={plan.tone} />
+          </dl>
+          <div className="mt-3 flex justify-end">
+            <Button onClick={() => onApplyPlan(plan)} variant="primary">
+              <Check className="h-4 w-4" aria-hidden="true" />
+              按字段回填向导
+            </Button>
+          </div>
+        </div>
+      )}
+    </div>
+  );
+}
+
+function PlanRow({ label, value }: { label: string; value: string }) {
+  if (!value.trim()) return null;
+  return (
+    <div className="flex gap-3">
+      <dt className="w-16 shrink-0 text-ink-soft">{label}</dt>
+      <dd className="min-w-0 flex-1 whitespace-pre-wrap text-ink">{value}</dd>
+    </div>
+  );
+}
+
 function StepStory({
   form,
   update,
   toggleSellingPoint,
-}: StepProps & { toggleSellingPoint: (p: string) => void }) {
+  onApplyPlan,
+}: StepProps & {
+  toggleSellingPoint: (p: string) => void;
+  onApplyPlan: (plan: ProjectPlanView) => void;
+}) {
   return (
     <div>
       <Field label="一句话故事（logline）">
@@ -211,6 +317,7 @@ function StepStory({
           placeholder="废柴少年觉醒禁忌血脉，在仙门倾轧中逆势封神。"
         />
       </Field>
+      <PlanAssistant form={form} onApplyPlan={onApplyPlan} />
       <Field label="故事结构">
         <SegmentedControl
           ariaLabel="故事结构"
@@ -267,6 +374,37 @@ function StepPremise({ form, update }: StepProps) {
           placeholder="例：抗争与代价"
         />
       </Field>
+      <Field label="基调">
+        <Select
+          value={form.tone}
+          onChange={(e) => update({ tone: e.target.value })}
+        >
+          <option value="">未选择</option>
+          {TONES.map((t) => (
+            <option key={t} value={t}>
+              {t}
+            </option>
+          ))}
+        </Select>
+      </Field>
+      <Field label="叙事视角">
+        <SegmentedControl
+          ariaLabel="叙事视角"
+          className="flex-wrap"
+          value={form.narrativePov}
+          onChange={(narrativePov) => update({ narrativePov })}
+          options={NARRATIVE_POVS.map((p) => ({ value: p, label: p }))}
+        />
+      </Field>
+      <Field label="结局取向">
+        <SegmentedControl
+          ariaLabel="结局取向"
+          className="flex-wrap"
+          value={form.endingType}
+          onChange={(endingType) => update({ endingType })}
+          options={ENDING_TYPES.map((e) => ({ value: e, label: e }))}
+        />
+      </Field>
     </div>
   );
 }
@@ -302,6 +440,9 @@ function StepConfirm({ form }: { form: WizardForm }) {
       label: "核心卖点",
       value: form.sellingPoints.length > 0 ? form.sellingPoints.join("、") : "未选择",
     },
+    { label: "基调", value: form.tone || "未选择" },
+    { label: "叙事视角", value: form.narrativePov || "未选择" },
+    { label: "结局取向", value: form.endingType || "未选择" },
   ];
   return (
     <div>

apps/web/components/characters/RelationshipGraph.tsx

@@ -0,0 +1,149 @@
+"use client";
+
+import { useEffect, useMemo, useState } from "react";
+import {
+  Background,
+  BackgroundVariant,
+  Controls,
+  ReactFlow,
+  type Edge,
+  type Node,
+} from "@xyflow/react";
+import { Users } from "lucide-react";
+
+// React Flow 样式必须显式引入（组件经 dynamic ssr:false 加载，样式随该 chunk 打包）。
+import "@xyflow/react/dist/style.css";
+
+import { EmptyState } from "@/components/ui/EmptyState";
+import {
+  buildRelationshipGraph,
+  NODE_WIDTH,
+  type GraphNode,
+} from "@/lib/characters/graphLayout";
+import type { CharacterCardView } from "@/lib/api/types";
+
+interface RelationshipGraphProps {
+  characters: CharacterCardView[];
+}
+
+// 纸感节点样式：面板底 + 细线框 + 左侧按分组配色的强调条（色值走主题 CSS 变量）。
+function paperNodeStyle(colorVar: string): React.CSSProperties {
+  return {
+    width: NODE_WIDTH,
+    padding: "8px 10px",
+    borderRadius: 6,
+    background: "var(--color-panel)",
+    color: "var(--color-ink)",
+    border: "1px solid var(--color-line)",
+    borderLeft: `4px solid var(${colorVar})`,
+    boxShadow: "0 1px 3px var(--shadow-paper)",
+    fontSize: 12,
+    textAlign: "left",
+  };
+}
+
+// 节点内容：角色名（宋体）+ 定位副标题。
+function NodeLabel({ name, role }: { name: string; role: string }) {
+  return (
+    <div className="min-w-0">
+      <p className="truncate font-serif text-sm text-ink">{name}</p>
+      {role ? <p className="truncate text-2xs text-ink-soft">{role}</p> : null}
+    </div>
+  );
+}
+
+function toReactFlowNodes(nodes: GraphNode[]): Node[] {
+  return nodes.map((node) => ({
+    id: node.id,
+    position: node.position,
+    data: { label: <NodeLabel name={node.data.label} role={node.data.role} /> },
+    style: paperNodeStyle(node.data.colorVar),
+    // 只读可视化：可拖动重排，不可新建连线。
+    connectable: false,
+  }));
+}
+
+function toReactFlowEdges(
+  edges: { id: string; source: string; target: string; label: string }[],
+): Edge[] {
+  return edges.map((edge) => ({
+    id: edge.id,
+    source: edge.source,
+    target: edge.target,
+    label: edge.label,
+    labelStyle: { fill: "var(--color-ink-soft)", fontSize: 11 },
+    labelBgStyle: { fill: "var(--color-bg)", fillOpacity: 0.9 },
+    labelBgPadding: [4, 2] as [number, number],
+    labelBgBorderRadius: 4,
+    style: { stroke: "var(--color-line)", strokeWidth: 1.5 },
+  }));
+}
+
+// 尊重 prefers-reduced-motion：减动偏好下关闭 fitView 平滑动画与边流动动画。
+function usePrefersReducedMotion(): boolean {
+  const [reduced, setReduced] = useState(false);
+  useEffect(() => {
+    const media = window.matchMedia("(prefers-reduced-motion: reduce)");
+    const update = () => setReduced(media.matches);
+    update();
+    media.addEventListener("change", update);
+    return () => media.removeEventListener("change", update);
+  }, []);
+  return reduced;
+}
+
+// 角色关系图谱（灵感⑥ 缩减版 / D4）：只用现有 relations 数据，节点按 role 上色、
+// 边按 name-join、kind 作中文标签。悬空边（引用未入库角色）已在 graphLayout 略过并计数。
+export function RelationshipGraph({ characters }: RelationshipGraphProps) {
+  const reducedMotion = usePrefersReducedMotion();
+
+  const graph = useMemo(
+    () => buildRelationshipGraph(characters),
+    [characters],
+  );
+  const nodes = useMemo(() => toReactFlowNodes(graph.nodes), [graph.nodes]);
+  const edges = useMemo(() => toReactFlowEdges(graph.edges), [graph.edges]);
+
+  if (graph.nodes.length === 0) {
+    return (
+      <EmptyState
+        icon={Users}
+        title="暂无可视化的人物"
+        description="先入库角色并登记彼此关系，这里会绘制出关系图谱。"
+      />
+    );
+  }
+
+  return (
+    <div className="flex flex-col gap-2">
+      {/* React Flow 必须有显式尺寸的容器，否则画布高度为 0。 */}
+      <div className="h-[28rem] w-full overflow-hidden rounded border border-line bg-bg">
+        <ReactFlow
+          nodes={nodes}
+          edges={edges}
+          fitView
+          fitViewOptions={{ padding: 0.2, duration: reducedMotion ? 0 : 400 }}
+          nodesConnectable={false}
+          edgesFocusable={false}
+          proOptions={{ hideAttribution: false }}
+          minZoom={0.2}
+          maxZoom={1.75}
+          defaultEdgeOptions={{ animated: false }}
+        >
+          <Background
+            variant={BackgroundVariant.Dots}
+            gap={20}
+            size={1}
+            color="var(--color-line)"
+          />
+          <Controls showInteractive={false} />
+        </ReactFlow>
+      </div>
+      {graph.danglingCount > 0 ? (
+        <p className="text-2xs text-ink-soft">
+          有 {graph.danglingCount} 条关系引用了未入库的角色，已略过。
+        </p>
+      ) : null}
+    </div>
+  );
+}

apps/web/components/codex/CodexPage.tsx

@@ -1,7 +1,8 @@
 "use client";
 
 import { useMemo, useState } from "react";
-import { Clock3, Globe2, Sparkles, UserRound } from "lucide-react";
+import dynamic from "next/dynamic";
+import { Clock3, Globe2, Share2, Sparkles, UserRound } from "lucide-react";
 
 import { AppShell } from "@/components/AppShell";
 import { CharacterGenerator } from "@/components/generation/CharacterGenerator";
@@ -39,6 +40,23 @@ const TABS: { key: CodexTab; label: string }[] = [
   { key: "timeline", label: "时间线" },
 ];
 
+// 关系图谱（React Flow）只在客户端渲染：ssr:false 绕开 canvas/window 依赖，
+// 并把 React Flow 及其样式拆到独立 chunk，进人物 tab 才加载。
+const RelationshipGraph = dynamic(
+  () =>
+    import("@/components/characters/RelationshipGraph").then(
+      (mod) => mod.RelationshipGraph,
+    ),
+  {
+    ssr: false,
+    loading: () => (
+      <div className="flex h-[28rem] items-center justify-center rounded border border-line bg-bg text-sm text-ink-soft">
+        关系图谱加载中…
+      </div>
+    ),
+  },
+);
+
 // 设定库 Codex（UX §6.5）：管理人物 / 世界观 / 时间线。
 // 初始列表来自后端读端点（跨会话全量真源）；本会话新入库的角色卡再合并补显，
 // 故刷新后不再为空、且不与真源重复（按 name 去重，见 mergeCharacterCards）。
@@ -143,6 +161,15 @@ export function CodexPage({
                   />
                 )}
               </Card>
+              {characters.length > 0 ? (
+                <Card as="section" className="p-4">
+                  <h3 className="mb-3 flex items-center gap-2 font-serif text-sm text-ink">
+                    <Share2 className="h-4 w-4 text-cinnabar" aria-hidden="true" />
+                    关系图谱
+                  </h3>
+                  <RelationshipGraph characters={characters} />
+                </Card>
+              ) : null}
               <div id="character-generator">
                 <CharacterGenerator
                   projectId={project.id}

apps/web/components/generation/CharacterCardItem.tsx

@@ -55,6 +55,40 @@ export function CharacterCardItem({
         </p>
       ) : null}
 
+      {onEdit || card.appearance ? (
+        <div className="mb-1 flex items-start gap-1 text-ink-soft">
+          <span className="shrink-0 text-ink">外貌：</span>
+          {onEdit ? (
+            <TextInput
+              value={card.appearance ?? ""}
+              onChange={(e) => onEdit({ appearance: e.target.value })}
+              aria-label="外貌形象"
+              controlSize="sm"
+              className="min-w-0 flex-1"
+            />
+          ) : (
+            <span>{card.appearance}</span>
+          )}
+        </div>
+      ) : null}
+
+      {onEdit || card.motive ? (
+        <div className="mb-1 flex items-start gap-1 text-ink-soft">
+          <span className="shrink-0 text-ink">动机：</span>
+          {onEdit ? (
+            <TextInput
+              value={card.motive ?? ""}
+              onChange={(e) => onEdit({ motive: e.target.value })}
+              aria-label="核心动机"
+              controlSize="sm"
+              className="min-w-0 flex-1"
+            />
+          ) : (
+            <span>{card.motive}</span>
+          )}
+        </div>
+      ) : null}
+
       <p className="mb-1 text-ink-soft">
         <span className="text-ink">背景：</span>
         {card.backstory}

apps/web/components/generation/CharacterGenerator.tsx

@@ -7,6 +7,7 @@ import { ThinkingIndicator } from "@/components/ThinkingIndicator";
 import { Button } from "@/components/ui/Button";
 import { Field } from "@/components/ui/Field";
 import { SectionHeader } from "@/components/ui/SectionHeader";
+import { SegmentedControl } from "@/components/ui/SegmentedControl";
 import { StatusNote } from "@/components/ui/StatusNote";
 import { TextInput } from "@/components/ui/TextInput";
 import type { CharacterCardView } from "@/lib/api/types";
@@ -14,11 +15,26 @@ import {
   clampCharacterCount,
   MAX_CHARACTER_COUNT,
   MIN_CHARACTER_COUNT,
+  roleMixTotal,
+  sanitizeRoleMix,
   updateCard,
 } from "@/lib/generation/cards";
 import { useCharacterGen } from "@/lib/generation/useCharacterGen";
 import { CharacterCardItem } from "./CharacterCardItem";
 import { ConflictAdjudication } from "./ConflictAdjudication";
+import { RoleMixEditor, type RoleMixRow } from "./RoleMixEditor";
+
+type GenMode = "single" | "mix";
+
+const MODE_OPTIONS: Array<{ value: GenMode; label: string }> = [
+  { value: "single", label: "单一定位" },
+  { value: "mix", label: "按定位配比" },
+];
+
+const DEFAULT_MIX_ROWS: RoleMixRow[] = [
+  { role: "主角", count: 1 },
+  { role: "对手", count: 1 },
+];
 
 interface CharacterGeneratorProps {
   projectId: string;
@@ -35,6 +51,10 @@ export function CharacterGenerator({
   const [brief, setBrief] = useState("");
   const [count, setCount] = useState(MIN_CHARACTER_COUNT);
   const [role, setRole] = useState("");
+  const [mode, setMode] = useState<GenMode>("single");
+  const [mixRows, setMixRows] = useState<RoleMixRow[]>(DEFAULT_MIX_ROWS);
+  // 配比模式下的有效合计（sanitize 后，已按 MAX 截断）——门控生成按钮。
+  const mixTotal = roleMixTotal(sanitizeRoleMix(mixRows));
   // 编辑态卡片（预览返回后拷贝进本地，允许就地改）。
   const [drafts, setDrafts] = useState<CharacterCardView[]>([]);
   const [selected, setSelected] = useState<Set<number>>(new Set());
@@ -46,8 +66,17 @@ export function CharacterGenerator({
   }, []);
 
   const onGenerate = useCallback(async () => {
+    if (mode === "mix") {
+      await gen.generate({
+        projectId,
+        brief,
+        count,
+        roleMix: sanitizeRoleMix(mixRows),
+      });
+      return;
+    }
     await gen.generate({ projectId, brief, count, role });
-  }, [gen, projectId, brief, count, role]);
+  }, [gen, projectId, brief, count, role, mode, mixRows]);
 
   // gen.cards 变化（新预览）→ 重新同步本地草稿。
   const previewKey = gen.cards.map((c) => c.name).join("|");
@@ -96,7 +125,25 @@ export function CharacterGenerator({
             placeholder="如：一个亦正亦邪的剑修，背负灭门血仇"
           />
         </Field>
-        <div className="flex flex-wrap items-end gap-3">
+        <div className="mb-3">
+          <SegmentedControl
+            options={MODE_OPTIONS}
+            value={mode}
+            onChange={setMode}
+            ariaLabel="生成模式"
+          />
+        </div>
+
+        {mode === "mix" ? (
+          <Field
+            label="角色定位配比"
+            className="mb-3"
+            help="每个定位产出对应数量，合计即总张数"
+          >
+            <RoleMixEditor rows={mixRows} onChange={setMixRows} />
+          </Field>
+        ) : (
+          <div className="mb-3 flex flex-wrap items-end gap-3">
             <Field
               label="数量"
               className="w-28"
@@ -119,13 +166,17 @@ export function CharacterGenerator({
                 placeholder="主角 / CP / 对手 / 导师 / 工具人"
               />
             </Field>
+          </div>
+        )}
+
         <Button
           onClick={() => void onGenerate()}
           disabled={
             generating ||
             brief.trim().length === 0 ||
-              count < MIN_CHARACTER_COUNT ||
-              count > MAX_CHARACTER_COUNT
+            (mode === "mix"
+              ? mixTotal < MIN_CHARACTER_COUNT
+              : count < MIN_CHARACTER_COUNT || count > MAX_CHARACTER_COUNT)
           }
           variant="primary"
         >
@@ -133,7 +184,6 @@ export function CharacterGenerator({
           {generating ? "生成中…" : "生成"}
         </Button>
       </div>
-      </div>
 
       {generating ? (
         <div className="rounded border border-line bg-panel/50 p-4 text-cinnabar">

apps/web/components/generation/RoleMixEditor.tsx

@@ -0,0 +1,82 @@
+"use client";
+
+import { Plus, X } from "lucide-react";
+
+import { Button } from "@/components/ui/Button";
+import { TextInput } from "@/components/ui/TextInput";
+import {
+  MAX_CHARACTER_COUNT,
+  MIN_CHARACTER_COUNT,
+  roleMixTotal,
+  sanitizeRoleMix,
+} from "@/lib/generation/cards";
+
+// 单行配比：某定位 + 该定位要生成的数量。
+export interface RoleMixRow {
+  role: string;
+  count: number;
+}
+
+interface RoleMixEditorProps {
+  rows: RoleMixRow[];
+  onChange: (rows: RoleMixRow[]) => void;
+}
+
+// 群像按定位配比编辑器（⑦）：一组「定位 + 数量」行；合计封顶 MAX_CHARACTER_COUNT。
+// 不可变更新——每次编辑返回新数组，不改入参。
+export function RoleMixEditor({ rows, onChange }: RoleMixEditorProps) {
+  // 有效合计 = sanitize 后的总数（已按 MAX 截断），即实际会生成的张数。
+  const effectiveTotal = roleMixTotal(sanitizeRoleMix(rows));
+
+  const updateRow = (index: number, patch: Partial<RoleMixRow>): void => {
+    onChange(rows.map((r, i) => (i === index ? { ...r, ...patch } : r)));
+  };
+  const addRow = (): void => onChange([...rows, { role: "", count: 1 }]);
+  const removeRow = (index: number): void =>
+    onChange(rows.filter((_, i) => i !== index));
+
+  return (
+    <div className="flex flex-col gap-2" aria-label="角色定位配比">
+      {rows.map((row, i) => (
+        <div key={i} className="flex items-center gap-2">
+          <TextInput
+            value={row.role}
+            onChange={(e) => updateRow(i, { role: e.target.value })}
+            placeholder="定位：主角 / 对手 / 配角"
+            aria-label={`第 ${i + 1} 行定位`}
+            className="min-w-[10rem] flex-1"
+          />
+          <TextInput
+            type="number"
+            min={MIN_CHARACTER_COUNT}
+            max={MAX_CHARACTER_COUNT}
+            value={row.count}
+            onChange={(e) =>
+              updateRow(i, { count: Math.round(Number(e.target.value)) })
+            }
+            aria-label={`第 ${i + 1} 行数量`}
+            className="w-20"
+          />
+          <Button
+            variant="ghost"
+            size="icon"
+            onClick={() => removeRow(i)}
+            disabled={rows.length <= 1}
+            aria-label={`删除第 ${i + 1} 行`}
+          >
+            <X className="h-4 w-4" aria-hidden="true" />
+          </Button>
+        </div>
+      ))}
+      <div className="flex items-center justify-between">
+        <Button variant="outline" size="sm" onClick={addRow}>
+          <Plus className="h-4 w-4" aria-hidden="true" />
+          添加定位
+        </Button>
+        <span className="text-xs text-ink-soft">
+          合计 {effectiveTotal} / {MAX_CHARACTER_COUNT} 张
+        </span>
+      </div>
+    </div>
+  );
+}

apps/web/components/review/CharacterizationPanel.tsx

@@ -0,0 +1,102 @@
+"use client";
+
+import { CheckCircle2, CircleDashed, UserRound } from "lucide-react";
+
+import { Badge } from "@/components/ui/Badge";
+import type { CharacterizationIssue, CharacterizationReport } from "@/lib/review/sse";
+
+interface CharacterizationPanelProps {
+  characterization: CharacterizationReport | null;
+  incomplete: boolean;
+}
+
+// 高置信度阈值：≥0.7 为主区常显，<0.7 折进「更多低置信度建议」抽屉（控作者注意力预算，灵感 §10）。
+const HIGH_CONFIDENCE = 0.7;
+
+// 人物塑造报告区（灵感③ advisory 单维）：只读建议、**无裁决 UI**（不进 conflicts/不阻断验收）。
+// 以人物卡 motive/appearance 为客观锚点；每条带逐字引用 + 置信度。按置信度降序，
+// 低置信度折叠，避免维度噪声淹没作者。
+export function CharacterizationPanel({
+  characterization,
+  incomplete,
+}: CharacterizationPanelProps) {
+  if (incomplete) {
+    return (
+      <section>
+        <Badge variant="warning">
+          <CircleDashed className="h-3 w-3" aria-hidden="true" />
+          未完成
+        </Badge>
+      </section>
+    );
+  }
+  const issues = characterization?.issues ?? [];
+  if (characterization === null || issues.length === 0) {
+    return (
+      <section>
+        <Badge variant="success">
+          <CheckCircle2 className="h-3 w-3" aria-hidden="true" />
+          人物塑造无明显问题
+        </Badge>
+      </section>
+    );
+  }
+
+  // 置信度降序（优先级合并）；高/低分区。
+  const sorted = [...issues].sort((a, b) => b.confidence - a.confidence);
+  const high = sorted.filter((i) => i.confidence >= HIGH_CONFIDENCE);
+  const low = sorted.filter((i) => i.confidence < HIGH_CONFIDENCE);
+
+  return (
+    <section className="space-y-2 text-xs">
+      <p className="rounded border border-dashed border-line/70 bg-bg/40 px-3 py-1.5 text-[11px] text-ink-soft">
+        以下为人物塑造建议（仅供参考，不影响验收）。忽略文本长度与辞藻，只看人物是否写得立体。
+      </p>
+      <ul className="space-y-2">
+        {(high.length > 0 ? high : sorted).map((issue, i) => (
+          <CharacterizationIssueCard key={i} issue={issue} />
+        ))}
+      </ul>
+      {high.length > 0 && low.length > 0 ? (
+        <details className="group rounded border border-line/70 bg-bg/35 p-2">
+          <summary className="cursor-pointer list-none text-2xs font-semibold text-ink-soft">
+            更多低置信度建议（{low.length}）
+          </summary>
+          <ul className="mt-2 space-y-2">
+            {low.map((issue, i) => (
+              <CharacterizationIssueCard key={i} issue={issue} />
+            ))}
+          </ul>
+        </details>
+      ) : null}
+    </section>
+  );
+}
+
+function CharacterizationIssueCard({ issue }: { issue: CharacterizationIssue }) {
+  const pct = Math.round(Math.max(0, Math.min(1, issue.confidence)) * 100);
+  return (
+    <li className="rounded border border-line/70 bg-bg/35 p-2">
+      <div className="flex items-center justify-between gap-2">
+        <span className="flex items-center gap-1.5 font-semibold text-ink">
+          <UserRound className="h-3.5 w-3.5 text-ink-soft" aria-hidden="true" />
+          {issue.character || "（未归属角色）"}
+          {issue.aspect ? <Badge variant="accent">{issue.aspect}</Badge> : null}
+        </span>
+        <span className="shrink-0 font-mono text-2xs text-ink-soft" title="置信度">
+          {pct}%
+        </span>
+      </div>
+      {issue.where ? <p className="mt-1 text-ink-soft">定位：{issue.where}</p> : null}
+      {issue.quote ? (
+        <blockquote className="mt-1 border-l-2 border-line pl-2 text-ink-soft italic">
+          「{issue.quote}」
+        </blockquote>
+      ) : null}
+      {issue.diagnosis ? <p className="mt-1 text-ink">{issue.diagnosis}</p> : null}
+      {issue.suggestion ? (
+        <p className="mt-1 text-ink-soft">建议：{issue.suggestion}</p>
+      ) : null}
+    </li>
+  );
+}

apps/web/components/review/ReviewReport.tsx

@@ -32,6 +32,7 @@ import {
 } from "@/lib/review/decisions";
 import {
   latestReview,
+  normalizeCharacterization,
   normalizeConflicts,
   normalizeForeshadowSug,
   normalizePace,
@@ -56,6 +57,7 @@ import { useToast } from "@/components/Toast";
 import { ThinkingIndicator } from "@/components/ThinkingIndicator";
 import { AcceptPanel } from "./AcceptPanel";
 import { AnnotatedText } from "./AnnotatedText";
+import { CharacterizationPanel } from "./CharacterizationPanel";
 import { ConflictCard } from "./ConflictCard";
 import { ForeshadowSuggestions } from "./ForeshadowSuggestions";
 import { PacePanel } from "./PacePanel";
@@ -128,6 +130,10 @@ export function ReviewReport({
     () => normalizeStyleDrift(initialReview),
     [initialReview],
   );
+  const seededCharacterization = useMemo(
+    () => normalizeCharacterization(initialReview),
+    [initialReview],
+  );
   useEffect(() => {
     if (seededRef.current) return;
     seededRef.current = true;
@@ -135,22 +141,32 @@ export function ReviewReport({
       seededConflicts.length > 0 ||
       seededForeshadow.length > 0 ||
       seededPace !== null ||
-      seededStyle !== null;
+      seededStyle !== null ||
+      seededCharacterization !== null;
     if (hasSeed) {
       review.seed({
         conflicts: seededConflicts,
         foreshadow: seededForeshadow,
         pace: seededPace,
         style: seededStyle,
+        characterization: seededCharacterization,
       });
     }
-  }, [review, seededConflicts, seededForeshadow, seededPace, seededStyle]);
+  }, [
+    review,
+    seededConflicts,
+    seededForeshadow,
+    seededPace,
+    seededStyle,
+    seededCharacterization,
+  ]);
 
   // 当前生效冲突 = 流状态（重审后即时更新，否则种入的历史）。
   const conflicts: ReviewConflict[] = review.state.conflicts;
   const foreshadow = review.state.foreshadow;
   const pace = review.state.pace;
   const style = review.state.style;
+  const characterization = review.state.characterization;
   // 伏笔徽标/验收预览统一用「仍待处理数」（随登记/回收递减）；未上报前回退原始建议数。
   const foreshadowOpen = foreshadowRemaining ?? foreshadow.length;
   const sectionStatus = (name: string): string | undefined =>
@@ -359,7 +375,7 @@ export function ReviewReport({
     const fresh = normalizeConflicts(latestReview(data?.reviews));
     setMissing(new Set(outcome.missingIndices));
     if (fresh.length > 0) {
-      review.seed({ conflicts: fresh, foreshadow, pace, style });
+      review.seed({ conflicts: fresh, foreshadow, pace, style, characterization });
       setDrafts(emptyDecisions(fresh.length));
       toast("审稿报告已刷新：检测到未裁决冲突，请裁决后再验收", "error");
     } else {
@@ -405,7 +421,8 @@ export function ReviewReport({
     seededConflicts.length > 0 ||
     foreshadow.length > 0 ||
     pace !== null ||
-    style !== null;
+    style !== null ||
+    characterization !== null;
 
   return (
     <AppShell
@@ -671,6 +688,35 @@ export function ReviewReport({
               onRefine={onRefineSegment}
             />
             </ReviewSectionPanel>
+            <ReviewSectionPanel
+              title="人物塑造"
+              statusLabel={
+                sectionStatus("characterization") === "incomplete"
+                  ? "未完成"
+                  : characterization && characterization.issues.length > 0
+                    ? `${characterization.issues.length} 建议`
+                    : characterization
+                      ? "无明显问题"
+                      : "无报告"
+              }
+              statusVariant={
+                sectionStatus("characterization") === "incomplete"
+                  ? "warning"
+                  : characterization && characterization.issues.length > 0
+                    ? "accent"
+                    : characterization
+                      ? "success"
+                      : "neutral"
+              }
+              defaultOpen={
+                characterization !== null && characterization.issues.length > 0
+              }
+            >
+            <CharacterizationPanel
+              characterization={characterization}
+              incomplete={sectionStatus("characterization") === "incomplete"}
+            />
+            </ReviewSectionPanel>
             {refineText !== null ? (
               <RefineView
                 projectId={project.id}

apps/web/lib/api/schema.d.ts

@@ -560,6 +560,29 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/skills/project-plan/generate": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /**
+         * Generate Project Plan
+         * @description 生成结构化立项方案预览（向导阶段，project 未建，不入库）。
+         *
+         *     **不带 project 前缀**——绕开 toolbox 的 `project_repo.get→404`（立项前无 project）。
+         *     **种子门控**：缺 genre 或 logline → 422（空向导不出方案，防 slop）。无凭据 → 503。
+         */
+        post: operations["generate_project_plan_skills_project_plan_generate_post"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/skills/toolbox": {
         parameters: {
             query?: never;
@@ -922,9 +945,21 @@ export interface components {
             role: string;
             /**
              * Traits
-             * @description 性格特质清单
+             * @description 性格特质清单（性格三层）
              */
             traits?: string[];
+            /**
+             * Appearance
+             * @description 外貌形象（缺则空串）
+             * @default
+             */
+            appearance: string;
+            /**
+             * Motive
+             * @description 核心动机/欲望/恐惧（缺则空串）
+             * @default
+             */
+            motive: string;
             /**
              * Backstory
              * @description 背景故事
@@ -962,6 +997,10 @@ export interface components {
         /**
          * CharacterGenerateRequest
          * @description POST /projects/:id/characters/generate：据需求生成角色卡（预览，群像防雷同）。
+         *
+         *     `role_mix`（⑦ 群像按定位配比）与 `count` 的关系：`role_mix` 在场时 `count` 取各值之和
+         *     （忽略客户端显式 count），每项数量须 ≥ 1、总和须 ≤ `MAX_GENERATE_COUNT`；缺省则退回
+         *     单一 `role` + `count` 语义。
          */
         CharacterGenerateRequest: {
             /**
@@ -971,7 +1010,7 @@ export interface components {
             brief: string;
             /**
              * Count
-             * @description 生成数量
+             * @description 生成数量（role_mix 在场时取各值之和）
              * @default 1
              */
             count: number;
@@ -980,6 +1019,13 @@ export interface components {
              * @description 角色定位（主角/CP/对手/导师/工具人 等）
              */
             role?: string | null;
+            /**
+             * Role Mix
+             * @description 按定位配比生成 {定位: 数量}；在场时 count=各值之和（每项≥1、总和≤12）
+             */
+            role_mix?: {
+                [key: string]: number;
+            } | null;
         };
         /**
          * CharacterIngestRequest
@@ -1506,6 +1552,12 @@ export interface components {
             selling_points?: string[];
             /** Structure */
             structure?: string | null;
+            /** Tone */
+            tone?: string | null;
+            /** Ending Type */
+            ending_type?: string | null;
+            /** Narrative Pov */
+            narrative_pov?: string | null;
         };
         /**
          * ProjectListResponse
@@ -1515,6 +1567,114 @@ export interface components {
             /** Projects */
             projects?: components["schemas"]["ProjectResponse"][];
         };
+        /**
+         * ProjectPlanGenerateRequest
+         * @description POST /skills/project-plan/generate：向导草稿种子（不带 project 前缀，立项前 project 未建）。
+         *
+         *     **种子门控**：`genre` + `logline` 为必给种子（端点校验非空，否则 422）——空向导不出方案，
+         *     防 slop。其余字段有则贴合、无则不臆造（序列化进 project_context，仿 `_project_context`）。
+         *     全字段带 `max_length` 上界（CR-H9），`brief` = 作者一句话方向（可空）。
+         */
+        ProjectPlanGenerateRequest: {
+            /**
+             * Genre
+             * @description 题材（种子）
+             */
+            genre?: string | null;
+            /**
+             * Logline
+             * @description 一句话故事（种子）
+             */
+            logline?: string | null;
+            /**
+             * Title
+             * @description 暂定书名（可空）
+             */
+            title?: string | null;
+            /**
+             * Premise
+             * @description 立意（可空）
+             */
+            premise?: string | null;
+            /**
+             * Theme
+             * @description 主题（可空）
+             */
+            theme?: string | null;
+            /**
+             * Structure
+             * @description 故事结构（可空）
+             */
+            structure?: string | null;
+            /**
+             * Tone
+             * @description 基调（可空）
+             */
+            tone?: string | null;
+            /**
+             * Ending Type
+             * @description 结局取向（可空）
+             */
+            ending_type?: string | null;
+            /**
+             * Narrative Pov
+             * @description 叙事视角（可空）
+             */
+            narrative_pov?: string | null;
+            /**
+             * Selling Points
+             * @description 核心卖点（可空）
+             */
+            selling_points?: string[];
+            /**
+             * Brief
+             * @description 作者一句话方向/需求（可空）
+             */
+            brief?: string | null;
+        };
+        /**
+         * ProjectPlanView
+         * @description 立项方案生成结果（贴 ww_agents.ProjectPlanResult；结构化预览，不入库）。
+         *
+         *     全字段有默认值（守解析韧性）——前端按字段回填立项向导，只填非空交集、不覆盖作者已编辑值。
+         */
+        ProjectPlanView: {
+            /**
+             * Title Candidates
+             * @description 书名候选清单
+             */
+            title_candidates?: string[];
+            /**
+             * Setting
+             * @description 时空背景
+             * @default
+             */
+            setting: string;
+            /**
+             * Narrative Structure
+             * @description 叙事结构
+             * @default
+             */
+            narrative_structure: string;
+            /**
+             * Story Core
+             * @description 故事核心
+             * @default
+             */
+            story_core: string;
+            /**
+             * Ending Design
+             * @description 结局设计
+             * @default
+             */
+            ending_design: string;
+            /**
+             * Tone
+             * @description 基调
+             * @default
+             */
+            tone: string;
+        };
         /**
          * ProjectResponse
          * @description 项目视图（创建/列表/详情共用）。
@@ -1539,6 +1699,12 @@ export interface components {
             selling_points?: string[];
             /** Structure */
             structure?: string | null;
+            /** Tone */
+            tone?: string | null;
+            /** Ending Type */
+            ending_type?: string | null;
+            /** Narrative Pov */
+            narrative_pov?: string | null;
             /**
              * Updated At
              * @description 项目最近更新时间
@@ -1673,6 +1839,10 @@ export interface components {
             pace?: {
                 [key: string]: unknown;
             } | null;
+            /** Characterization */
+            characterization?: {
+                [key: string]: unknown;
+            } | null;
             /** Health Score */
             health_score?: number | null;
             /** Decisions */
@@ -3336,6 +3506,39 @@ export interface operations {
             };
         };
     };
+    generate_project_plan_skills_project_plan_generate_post: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody: {
+            content: {
+                "application/json": components["schemas"]["ProjectPlanGenerateRequest"];
+            };
+        };
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["ProjectPlanView"];
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
     list_toolbox_skills_toolbox_get: {
         parameters: {
             query?: never;

apps/web/lib/api/types.ts

@@ -4,6 +4,10 @@ import type { components } from "./schema";
 export type ProjectResponse = components["schemas"]["ProjectResponse"];
 export type ProjectCreateRequest = components["schemas"]["ProjectCreateRequest"];
 export type ProjectListResponse = components["schemas"]["ProjectListResponse"];
+// AI 立项方案生成（灵感⑤）：向导阶段 project 未建，种子门控 + 结构化预览回填向导。
+export type ProjectPlanGenerateRequest =
+  components["schemas"]["ProjectPlanGenerateRequest"];
+export type ProjectPlanView = components["schemas"]["ProjectPlanView"];
 export type DraftSaveRequest = components["schemas"]["DraftSaveRequest"];
 export type DraftResponse = components["schemas"]["DraftResponse"];
 // GET .../draft 读端点：含正文，供工作台重访时重载编辑器（404→空编辑器）。

apps/web/lib/characters/graphLayout.test.ts

@@ -0,0 +1,258 @@
+import { describe, expect, it } from "vitest";
+
+import type { CharacterCardView } from "@/lib/api/types";
+
+import {
+  buildGraphEdges,
+  buildGraphNodes,
+  buildRelationshipGraph,
+  countDanglingRelations,
+  DEFAULT_RELATION_LABEL,
+  layoutGraph,
+  roleStyle,
+} from "./graphLayout";
+
+// 最小角色卡工厂：CharacterCardView 多字段必填，测试只关心 name/role/relations。
+function ch(
+  name: string,
+  role: string,
+  relations: { name: string; kind?: string; note?: string | null }[] = [],
+): CharacterCardView {
+  return {
+    name,
+    role,
+    traits: [],
+    appearance: "",
+    motive: "",
+    backstory: "",
+    arc: "",
+    speech_tics: [],
+    relations: relations.map((r) => ({
+      name: r.name,
+      kind: r.kind ?? "",
+      note: r.note ?? null,
+    })),
+  };
+}
+
+describe("roleStyle", () => {
+  it("主角关键词 → protagonist + 朱砂色", () => {
+    expect(roleStyle("主角")).toEqual({
+      group: "protagonist",
+      colorVar: "--color-cinnabar",
+    });
+    expect(roleStyle("男主角").group).toBe("protagonist");
+  });
+
+  it("反派/对手/敌 → antagonist + 冲突色", () => {
+    expect(roleStyle("反派").group).toBe("antagonist");
+    expect(roleStyle("对手").group).toBe("antagonist");
+    expect(roleStyle("宿敌").group).toBe("antagonist");
+    expect(roleStyle("反派").colorVar).toBe("--color-conflict");
+  });
+
+  it("导师/师父 → mentor + info 色", () => {
+    expect(roleStyle("导师").group).toBe("mentor");
+    expect(roleStyle("师父").group).toBe("mentor");
+    expect(roleStyle("导师").colorVar).toBe("--color-info");
+  });
+
+  it("CP/恋人 → ally（大小写不敏感）", () => {
+    expect(roleStyle("CP").group).toBe("ally");
+    expect(roleStyle("cp").group).toBe("ally");
+    expect(roleStyle("恋人").group).toBe("ally");
+  });
+
+  it("配角/工具人 → supporting", () => {
+    expect(roleStyle("配角").group).toBe("supporting");
+    expect(roleStyle("工具人").group).toBe("supporting");
+  });
+
+  it("未知/空定位 → other（默认）", () => {
+    expect(roleStyle("神秘存在").group).toBe("other");
+    expect(roleStyle("").group).toBe("other");
+    expect(roleStyle("  ").group).toBe("other");
+  });
+});
+
+describe("buildGraphNodes", () => {
+  it("每个唯一角色一个节点，id=name，role 决定分组/配色", () => {
+    const nodes = buildGraphNodes([ch("林风", "主角"), ch("赵石", "对手")]);
+    expect(nodes).toHaveLength(2);
+    expect(nodes[0]).toMatchObject({
+      id: "林风",
+      position: { x: 0, y: 0 },
+      data: {
+        label: "林风",
+        role: "主角",
+        group: "protagonist",
+        colorVar: "--color-cinnabar",
+      },
+    });
+    expect(nodes[1]?.data.group).toBe("antagonist");
+  });
+
+  it("跳过空名与重名角色（先到先得）", () => {
+    const nodes = buildGraphNodes([
+      ch("林风", "主角"),
+      ch("  ", "配角"),
+      ch("林风", "对手"),
+    ]);
+    expect(nodes).toHaveLength(1);
+    expect(nodes[0]?.data.role).toBe("主角");
+  });
+
+  it("名字两端空白被裁剪用作 id", () => {
+    const nodes = buildGraphNodes([ch("  林风  ", "主角")]);
+    expect(nodes[0]?.id).toBe("林风");
+  });
+});
+
+describe("buildGraphEdges", () => {
+  const known = new Set(["林风", "赵石", "青禾"]);
+
+  it("kind → 边标签", () => {
+    const edges = buildGraphEdges(
+      [ch("林风", "主角", [{ name: "赵石", kind: "宿敌" }])],
+      known,
+    );
+    expect(edges).toHaveLength(1);
+    expect(edges[0]).toMatchObject({
+      source: "林风",
+      target: "赵石",
+      label: "宿敌",
+    });
+  });
+
+  it("kind 为空 → 回退到命名常量默认标签", () => {
+    const edges = buildGraphEdges(
+      [ch("林风", "主角", [{ name: "青禾", kind: "" }])],
+      known,
+    );
+    expect(edges[0]?.label).toBe(DEFAULT_RELATION_LABEL);
+  });
+
+  it("悬空边（目标未入库）被跳过", () => {
+    const edges = buildGraphEdges(
+      [ch("林风", "主角", [{ name: "不存在的人", kind: "师父" }])],
+      known,
+    );
+    expect(edges).toHaveLength(0);
+  });
+
+  it("自指关系被跳过", () => {
+    const edges = buildGraphEdges(
+      [ch("林风", "主角", [{ name: "林风", kind: "内心" }])],
+      known,
+    );
+    expect(edges).toHaveLength(0);
+  });
+
+  it("互指关系无向去重为单边", () => {
+    const edges = buildGraphEdges(
+      [
+        ch("林风", "主角", [{ name: "赵石", kind: "宿敌" }]),
+        ch("赵石", "对手", [{ name: "林风", kind: "宿敌" }]),
+      ],
+      known,
+    );
+    expect(edges).toHaveLength(1);
+  });
+
+  it("边 id 唯一", () => {
+    const edges = buildGraphEdges(
+      [
+        ch("林风", "主角", [
+          { name: "赵石", kind: "宿敌" },
+          { name: "青禾", kind: "青梅竹马" },
+        ]),
+      ],
+      known,
+    );
+    const ids = edges.map((e) => e.id);
+    expect(new Set(ids).size).toBe(ids.length);
+  });
+});
+
+describe("countDanglingRelations", () => {
+  it("统计引用未入库角色的关系数", () => {
+    const known = new Set(["林风"]);
+    const count = countDanglingRelations(
+      [
+        ch("林风", "主角", [
+          { name: "赵石", kind: "宿敌" }, // 悬空
+          { name: "青禾", kind: "青梅竹马" }, // 悬空
+        ]),
+      ],
+      known,
+    );
+    expect(count).toBe(2);
+  });
+
+  it("全部已入库时为 0", () => {
+    const known = new Set(["林风", "赵石"]);
+    const count = countDanglingRelations(
+      [ch("林风", "主角", [{ name: "赵石", kind: "宿敌" }])],
+      known,
+    );
+    expect(count).toBe(0);
+  });
+});
+
+describe("layoutGraph", () => {
+  const nodes = buildGraphNodes([ch("林风", "主角"), ch("赵石", "对手")]);
+  const edges = buildGraphEdges(
+    [ch("林风", "主角", [{ name: "赵石", kind: "宿敌" }])],
+    new Set(["林风", "赵石"]),
+  );
+
+  it("同输入产出相同坐标（确定性）", () => {
+    const a = layoutGraph(nodes, edges);
+    const b = layoutGraph(nodes, edges);
+    expect(a).toEqual(b);
+  });
+
+  it("dagre 赋予非零坐标并保持节点数据", () => {
+    const laid = layoutGraph(nodes, edges);
+    expect(laid).toHaveLength(2);
+    for (const n of laid) {
+      expect(Number.isFinite(n.position.x)).toBe(true);
+      expect(Number.isFinite(n.position.y)).toBe(true);
+      expect(n.data.label).toBeTruthy();
+    }
+    // 两节点经布局后不再重叠于原点。
+    expect(laid[0]?.position).not.toEqual(laid[1]?.position);
+  });
+
+  it("不可变：不修改传入节点，返回新对象", () => {
+    const before = structuredClone(nodes);
+    const laid = layoutGraph(nodes, edges);
+    expect(nodes).toEqual(before); // 入参未被修改
+    expect(laid[0]).not.toBe(nodes[0]); // 新对象
+  });
+});
+
+describe("buildRelationshipGraph", () => {
+  it("端到端：布局后节点 + 有效边 + 悬空计数", () => {
+    const graph = buildRelationshipGraph([
+      ch("林风", "主角", [
+        { name: "赵石", kind: "宿敌" },
+        { name: "幽灵", kind: "谜团" }, // 悬空
+      ]),
+      ch("赵石", "对手", [{ name: "林风", kind: "宿敌" }]),
+    ]);
+    expect(graph.nodes).toHaveLength(2);
+    expect(graph.edges).toHaveLength(1);
+    expect(graph.danglingCount).toBe(1);
+    // 已布局（非全零）。
+    const positions = graph.nodes.map((n) => `${n.position.x},${n.position.y}`);
+    expect(new Set(positions).size).toBe(2);
+  });
+
+  it("空输入 → 空图", () => {
+    const graph = buildRelationshipGraph([]);
+    expect(graph.nodes).toEqual([]);
+    expect(graph.edges).toEqual([]);
+    expect(graph.danglingCount).toBe(0);
+  });
+});

apps/web/lib/characters/graphLayout.ts

@@ -0,0 +1,245 @@
+import dagre from "@dagrejs/dagre";
+
+import type { CharacterCardView } from "@/lib/api/types";
+
+// 关系图布局纯逻辑（灵感⑥ 缩减版）：把「角色 + relations」转成节点/边并跑一次性
+// dagre 初始坐标。仅用现有读端点数据（CharacterRelationView={name,kind,note}）——
+// 无 closeness/阵营/target-id，节点按 role 上色，边按 name-join、kind 作中文标签。
+// 无 DOM 依赖，可在 node 环境单测；React Flow 组件只做渲染，不含此处逻辑。
+
+// 节点尺寸（dagre 布局与 React Flow 渲染共用；命名常量非魔法值）。
+export const NODE_WIDTH = 176;
+export const NODE_HEIGHT = 56;
+// dagre 间距。
+export const DAGRE_RANK_SEP = 88;
+export const DAGRE_NODE_SEP = 48;
+// kind 缺失时的关系标签回退（命名常量非魔法值）。
+export const DEFAULT_RELATION_LABEL = "关系";
+
+// 角色定位分组（role 是自由中文文本，无固定枚举 → 关键词映射到有限分组）。
+export type RoleGroup =
+  | "protagonist"
+  | "antagonist"
+  | "mentor"
+  | "ally"
+  | "supporting"
+  | "other";
+
+export interface RoleStyle {
+  group: RoleGroup;
+  // 纸感主题 CSS 变量名（随亮/暗主题自适应，不写死色值）。
+  colorVar: string;
+}
+
+// role→分组关键词规则；按序匹配，首个命中生效，否则回退 other。
+const ROLE_RULES: ReadonlyArray<{
+  keywords: readonly string[];
+  style: RoleStyle;
+}> = [
+  {
+    keywords: ["主角", "主人公", "男主", "女主"],
+    style: { group: "protagonist", colorVar: "--color-cinnabar" },
+  },
+  {
+    keywords: ["反派", "对手", "宿敌", "敌", "boss"],
+    style: { group: "antagonist", colorVar: "--color-conflict" },
+  },
+  {
+    keywords: ["导师", "师父", "师傅", "师尊", "老师"],
+    style: { group: "mentor", colorVar: "--color-info" },
+  },
+  {
+    keywords: ["cp", "恋", "爱人", "伴侣", "挚友", "知己"],
+    style: { group: "ally", colorVar: "--color-pass" },
+  },
+  {
+    keywords: ["配角", "工具人", "路人", "龙套"],
+    style: { group: "supporting", colorVar: "--color-ink-soft" },
+  },
+];
+
+const DEFAULT_ROLE_STYLE: RoleStyle = {
+  group: "other",
+  colorVar: "--color-ink-soft",
+};
+
+// 把角色定位映射到分组/配色（大小写不敏感，子串匹配）。
+export function roleStyle(role: string): RoleStyle {
+  const normalized = role.trim().toLowerCase();
+  if (!normalized) {
+    return DEFAULT_ROLE_STYLE;
+  }
+  for (const rule of ROLE_RULES) {
+    if (rule.keywords.some((keyword) => normalized.includes(keyword))) {
+      return rule.style;
+    }
+  }
+  return DEFAULT_ROLE_STYLE;
+}
+
+export interface RelationNodeData {
+  label: string;
+  role: string;
+  group: RoleGroup;
+  colorVar: string;
+}
+
+export interface GraphNode {
+  id: string; // = 角色名（name-join 主键）
+  position: { x: number; y: number };
+  data: RelationNodeData;
+}
+
+export interface GraphEdge {
+  id: string;
+  source: string;
+  target: string;
+  label: string;
+}
+
+export interface RelationshipGraphData {
+  nodes: GraphNode[];
+  edges: GraphEdge[];
+  // 引用未入库角色而被略过的关系条数（供 UI 提示）。
+  danglingCount: number;
+}
+
+// 每个唯一角色一个节点；空名/重名（先到先得）跳过。位置初始 {0,0}，由 layoutGraph 赋值。
+export function buildGraphNodes(
+  characters: readonly CharacterCardView[],
+): GraphNode[] {
+  const seen = new Set<string>();
+  const nodes: GraphNode[] = [];
+  for (const character of characters) {
+    const name = character.name.trim();
+    if (!name || seen.has(name)) {
+      continue;
+    }
+    seen.add(name);
+    const style = roleStyle(character.role);
+    nodes.push({
+      id: name,
+      position: { x: 0, y: 0 },
+      data: {
+        label: name,
+        role: character.role,
+        group: style.group,
+        colorVar: style.colorVar,
+      },
+    });
+  }
+  return nodes;
+}
+
+// 无向对键：与方向无关，用于 A↔B 去重。
+function pairKey(a: string, b: string): string {
+  return [a, b].sort().join("
");
+}
+
+// 按 name-join 构边：kind→标签；悬空边（目标不在 knownNames）与自指跳过；互指无向去重。
+export function buildGraphEdges(
+  characters: readonly CharacterCardView[],
+  knownNames: ReadonlySet<string>,
+): GraphEdge[] {
+  const edges: GraphEdge[] = [];
+  const seenPairs = new Set<string>();
+  let index = 0;
+  for (const character of characters) {
+    const source = character.name.trim();
+    if (!source || !knownNames.has(source)) {
+      continue;
+    }
+    for (const relation of character.relations ?? []) {
+      const target = relation.name.trim();
+      if (!target || !knownNames.has(target) || target === source) {
+        continue;
+      }
+      const key = pairKey(source, target);
+      if (seenPairs.has(key)) {
+        continue;
+      }
+      seenPairs.add(key);
+      const label = relation.kind.trim() || DEFAULT_RELATION_LABEL;
+      edges.push({
+        id: `e-${index}-${source}-${target}`,
+        source,
+        target,
+        label,
+      });
+      index += 1;
+    }
+  }
+  return edges;
+}
+
+// 统计引用未入库角色（悬空）的关系条数——供 UI「N 条关系已略过」提示。
+export function countDanglingRelations(
+  characters: readonly CharacterCardView[],
+  knownNames: ReadonlySet<string>,
+): number {
+  let count = 0;
+  for (const character of characters) {
+    const source = character.name.trim();
+    if (!source || !knownNames.has(source)) {
+      continue;
+    }
+    for (const relation of character.relations ?? []) {
+      const target = relation.name.trim();
+      if (target && target !== source && !knownNames.has(target)) {
+        count += 1;
+      }
+    }
+  }
+  return count;
+}
+
+interface LaidOutNode {
+  x: number;
+  y: number;
+}
+
+// 跑一次性 dagre 布局（确定性：同输入同坐标）。dagre 给中心坐标 → 转 React Flow 左上角。
+// 不可变：返回新节点数组，不修改入参。
+export function layoutGraph(
+  nodes: readonly GraphNode[],
+  edges: readonly GraphEdge[],
+): GraphNode[] {
+  const graph = new dagre.graphlib.Graph();
+  graph.setGraph({
+    rankdir: "LR",
+    nodesep: DAGRE_NODE_SEP,
+    ranksep: DAGRE_RANK_SEP,
+  });
+  graph.setDefaultEdgeLabel(() => ({}));
+  for (const node of nodes) {
+    graph.setNode(node.id, { width: NODE_WIDTH, height: NODE_HEIGHT });
+  }
+  for (const edge of edges) {
+    graph.setEdge(edge.source, edge.target);
+  }
+  dagre.layout(graph);
+  return nodes.map((node) => {
+    const laid = graph.node(node.id) as LaidOutNode | undefined;
+    const centerX = laid?.x ?? 0;
+    const centerY = laid?.y ?? 0;
+    return {
+      ...node,
+      position: {
+        x: centerX - NODE_WIDTH / 2,
+        y: centerY - NODE_HEIGHT / 2,
+      },
+    };
+  });
+}
+
+// 端到端：角色列表 → 布局后的节点 + 有效边 + 悬空计数（组件里 useMemo 缓存）。
+export function buildRelationshipGraph(
+  characters: readonly CharacterCardView[],
+): RelationshipGraphData {
+  const nodes = buildGraphNodes(characters);
+  const knownNames = new Set(nodes.map((node) => node.id));
+  const edges = buildGraphEdges(characters, knownNames);
+  const danglingCount = countDanglingRelations(characters, knownNames);
+  const laidOut = layoutGraph(nodes, edges);
+  return { nodes: laidOut, edges, danglingCount };
+}

apps/web/lib/generation/cards.test.ts

@@ -16,6 +16,8 @@ import {
   mergeCharacterCards,
   mergeWorldEntities,
   MIN_CHARACTER_COUNT,
+  roleMixTotal,
+  sanitizeRoleMix,
   updateCard,
   worldEntityRules,
 } from "./cards";
@@ -23,6 +25,8 @@ import {
 const card = (over: Partial<CharacterCardView> = {}): CharacterCardView => ({
   name: "甲",
   role: "主角",
+  appearance: "白衣如雪",
+  motive: "为家族正名",
   backstory: "出身寒门",
   arc: "由怯转勇",
   ...over,
@@ -60,6 +64,62 @@ describe("buildCharacterGenerateRequest", () => {
       role: "对手",
     });
   });
+
+  it("uses role_mix (count=sum, no single role) when a mix is supplied", () => {
+    expect(
+      buildCharacterGenerateRequest("群像", 1, "忽略", {
+        主角: 1,
+        对手: 2,
+        配角: 3,
+      }),
+    ).toEqual({
+      brief: "群像",
+      count: 6,
+      role_mix: { 主角: 1, 对手: 2, 配角: 3 },
+    });
+  });
+
+  it("falls back to count/role when role_mix is empty", () => {
+    expect(buildCharacterGenerateRequest("群像", 4, "主角", {})).toEqual({
+      brief: "群像",
+      count: 4,
+      role: "主角",
+    });
+  });
+});
+
+describe("sanitizeRoleMix", () => {
+  it("trims roles, drops blanks, clamps counts to >=1, dedups", () => {
+    expect(
+      sanitizeRoleMix([
+        { role: "  主角  ", count: 2 },
+        { role: "", count: 3 }, // 空定位丢弃
+        { role: "对手", count: 0 }, // clamp 到 1
+        { role: "主角", count: 5 }, // 重复定位丢弃（先到先得）
+      ]),
+    ).toEqual({ 主角: 2, 对手: 1 });
+  });
+
+  it("caps the running total at MAX_CHARACTER_COUNT", () => {
+    const out = sanitizeRoleMix([
+      { role: "主角", count: 10 },
+      { role: "对手", count: 5 }, // 只剩 2 额度
+      { role: "配角", count: 3 }, // 无额度，丢弃
+    ]);
+    expect(out).toEqual({ 主角: 10, 对手: 2 });
+    expect(roleMixTotal(out)).toBe(MAX_CHARACTER_COUNT);
+  });
+
+  it("returns empty object when no valid rows", () => {
+    expect(sanitizeRoleMix([{ role: "  ", count: 3 }])).toEqual({});
+  });
+});
+
+describe("roleMixTotal", () => {
+  it("sums counts", () => {
+    expect(roleMixTotal({ 主角: 1, 对手: 2 })).toBe(3);
+    expect(roleMixTotal({})).toBe(0);
+  });
 });
 
 describe("buildCharacterIngestRequest", () => {
@@ -86,6 +146,18 @@ describe("updateCard", () => {
     expect(original.name).toBe("甲");
     expect(next).not.toBe(original);
   });
+
+  it("patches appearance/motive without dropping the other fields", () => {
+    const original = card();
+    const next = updateCard(original, {
+      appearance: "玄色劲装",
+      motive: "复仇",
+    });
+    expect(next.appearance).toBe("玄色劲装");
+    expect(next.motive).toBe("复仇");
+    expect(next.backstory).toBe(original.backstory);
+    expect(original.appearance).toBe("白衣如雪"); // no mutation
+  });
 });
 
 describe("extractIngestConflicts", () => {

apps/web/lib/generation/useCharacterGen.test.ts

@@ -53,6 +53,24 @@ describe("useCharacterGen", () => {
     expect(toast).not.toHaveBeenCalled();
   });
 
+  it("按定位配比生成：请求体带 role_mix、count 取各值之和", async () => {
+    post.mockResolvedValue({ data: { cards: [] }, error: null });
+    const { result } = renderHook(() => useCharacterGen());
+    await act(async () => {
+      await result.current.generate({
+        projectId: "p1",
+        brief: "群像",
+        count: 1,
+        role: "忽略",
+        roleMix: { 主角: 1, 对手: 2 },
+      });
+    });
+    const body = post.mock.calls[0]?.[1]?.body;
+    expect(body.role_mix).toEqual({ 主角: 1, 对手: 2 });
+    expect(body.count).toBe(3);
+    expect(body.role).toBeUndefined();
+  });
+
   it("生成成功但 cards 缺失：回退为空数组", async () => {
     post.mockResolvedValue({ data: { cards: null }, error: null });
     const { result } = renderHook(() => useCharacterGen());

apps/web/lib/generation/useCharacterGen.ts

@@ -21,6 +21,8 @@ export interface CharacterGenInput {
   brief: string;
   count: number;
   role?: string | null;
+  // 群像按定位配比（⑦）；在场时后端把 count 定为各值之和，忽略 count/role。
+  roleMix?: Readonly<Record<string, number>> | null;
 }
 
 export interface UseCharacterGen {
@@ -50,7 +52,7 @@ export function useCharacterGen(): UseCharacterGen {
   const toast = useToast();
 
   const generate = useCallback<UseCharacterGen["generate"]>(
-    async ({ projectId, brief, count, role }) => {
+    async ({ projectId, brief, count, role, roleMix }) => {
       setGenStatus("generating");
       setCards([]);
       setConflicts(null);
@@ -60,7 +62,7 @@ export function useCharacterGen(): UseCharacterGen {
           "/projects/{project_id}/characters/generate",
           {
             params: { path: { project_id: projectId } },
-            body: buildCharacterGenerateRequest(brief, count, role),
+            body: buildCharacterGenerateRequest(brief, count, role, roleMix),
           },
         );
         if (error || !data) {

apps/web/lib/review/characterization-sse.test.ts

@@ -0,0 +1,108 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  initialReviewState,
+  parseReviewBlock,
+  reduceReview,
+  type ReviewSseEvent,
+} from "./sse";
+
+const FULL_ISSUE = {
+  character: "沈砚",
+  aspect: "动机一致性",
+  where: "第3段",
+  quote: "他忽然转身。",
+  diagnosis: "动机断裂",
+  suggestion: "扣回护妹动机",
+  confidence: 0.82,
+};
+
+describe("parseReviewBlock — characterization (灵感③ advisory)", () => {
+  it("parses a characterization frame with a full issue", () => {
+    const evt = parseReviewBlock(
+      `event:characterization\ndata:{"issues":[${JSON.stringify(FULL_ISSUE)}]}`,
+    );
+    expect(evt).toEqual({
+      event: "characterization",
+      data: { issues: [FULL_ISSUE] },
+    });
+  });
+
+  it("defaults missing issue fields (advisory 韧性)", () => {
+    const evt = parseReviewBlock(
+      'event:characterization\ndata:{"issues":[{"character":"阿黎"}]}',
+    );
+    expect(evt).toEqual({
+      event: "characterization",
+      data: {
+        issues: [
+          {
+            character: "阿黎",
+            aspect: "",
+            where: "",
+            quote: "",
+            diagnosis: "",
+            suggestion: "",
+            confidence: 0,
+          },
+        ],
+      },
+    });
+  });
+
+  it("parses an empty issues frame (无明显问题降级)", () => {
+    expect(
+      parseReviewBlock('event:characterization\ndata:{"issues":[]}'),
+    ).toEqual({ event: "characterization", data: { issues: [] } });
+  });
+
+  it("drops non-record issues without dropping the whole frame", () => {
+    const evt = parseReviewBlock(
+      'event:characterization\ndata:{"issues":["junk",{"character":"沈砚"}]}',
+    );
+    expect(evt?.event).toBe("characterization");
+    if (evt?.event === "characterization") {
+      expect(evt.data.issues).toHaveLength(1);
+      expect(evt.data.issues[0]?.character).toBe("沈砚");
+    }
+  });
+});
+
+describe("reduceReview — characterization (replace-style like pace)", () => {
+  it("sets characterization report and marks reviewing", () => {
+    const evt: ReviewSseEvent = {
+      event: "characterization",
+      data: { issues: [FULL_ISSUE] },
+    };
+    const out = reduceReview(initialReviewState, evt);
+    expect(out.phase).toBe("reviewing");
+    expect(out.characterization).toEqual({ issues: [FULL_ISSUE] });
+  });
+
+  it("replaces (not accumulates) the characterization report", () => {
+    const first = reduceReview(initialReviewState, {
+      event: "characterization",
+      data: { issues: [FULL_ISSUE] },
+    });
+    const second = reduceReview(first, {
+      event: "characterization",
+      data: { issues: [] },
+    });
+    expect(second.characterization).toEqual({ issues: [] });
+  });
+
+  it("leaves characterization untouched on a section event", () => {
+    const seeded = reduceReview(initialReviewState, {
+      event: "characterization",
+      data: { issues: [FULL_ISSUE] },
+    });
+    const out = reduceReview(seeded, {
+      event: "section",
+      data: { name: "characterization", status: "done" },
+    });
+    expect(out.characterization).toEqual({ issues: [FULL_ISSUE] });
+    expect(out.sections).toEqual([
+      { name: "characterization", status: "done" },
+    ]);
+  });
+});

apps/web/lib/review/history.test.ts

@@ -3,6 +3,7 @@ import { describe, expect, it } from "vitest";
 import type { ReviewConflictView, ReviewHistoryItem } from "@/lib/api/types";
 import {
   latestReview,
+  normalizeCharacterization,
   normalizeConflicts,
   normalizeForeshadowSug,
   normalizePace,
@@ -121,6 +122,62 @@ describe("normalizePace", () => {
   });
 });
 
+describe("normalizeCharacterization", () => {
+  it("tightens dict into CharacterizationReport with field defaults", () => {
+    const out = normalizeCharacterization({
+      ...item([]),
+      characterization: {
+        issues: [
+          {
+            character: "沈砚",
+            aspect: "动机一致性",
+            where: "第3段",
+            quote: "他忽然转身。",
+            diagnosis: "动机断裂",
+            suggestion: "扣回护妹动机",
+            confidence: 0.82,
+          },
+          // 缺字段的畸形条目：安全默认，不丢整条
+          { character: "阿黎" },
+        ],
+      },
+    });
+    expect(out).toEqual({
+      issues: [
+        {
+          character: "沈砚",
+          aspect: "动机一致性",
+          where: "第3段",
+          quote: "他忽然转身。",
+          diagnosis: "动机断裂",
+          suggestion: "扣回护妹动机",
+          confidence: 0.82,
+        },
+        {
+          character: "阿黎",
+          aspect: "",
+          where: "",
+          quote: "",
+          diagnosis: "",
+          suggestion: "",
+          confidence: 0,
+        },
+      ],
+    });
+  });
+
+  it("returns empty issues when characterization is present but issues missing", () => {
+    expect(
+      normalizeCharacterization({ ...item([]), characterization: {} }),
+    ).toEqual({ issues: [] });
+  });
+
+  it("returns null when characterization is missing or not a dict", () => {
+    expect(normalizeCharacterization(item([]))).toBeNull();
+    expect(normalizeCharacterization(undefined)).toBeNull();
+  });
+});
+
 describe("latestReview", () => {
   it("returns first item (history is newest-first)", () => {
     const a = item([]);

apps/web/lib/review/history.ts

@@ -2,7 +2,9 @@
 // 纯逻辑（可单测）；schema 把 conflicts 标成 {[k]:unknown}[]，这里安全收窄。
 
 import type { ReviewHistoryItem } from "@/lib/api/types";
+import { asCharacterizationIssues } from "./sse";
 import type {
+  CharacterizationReport,
   ForeshadowKind,
   ForeshadowSuggestion,
   PaceIssue,
@@ -80,6 +82,17 @@ export function normalizePace(
   };
 }
 
+// 把一条留痕的 characterization（dict）收紧成人物塑造报告；缺失/非 dict → null（不渲染）。
+// advisory：只展示、无裁决 UI（灵感 D2）。
+export function normalizeCharacterization(
+  item: ReviewHistoryItem | undefined,
+): CharacterizationReport | null {
+  const raw = item?.characterization;
+  if (typeof raw !== "object" || raw === null) return null;
+  const dict = raw as Record<string, unknown>;
+  return { issues: asCharacterizationIssues(dict["issues"]) };
+}
+
 // 最近一条留痕（GET .../reviews 已按新→旧排序）。
 export function latestReview(
   items: readonly ReviewHistoryItem[] | undefined,

apps/web/lib/review/sse.ts

@@ -60,6 +60,22 @@ export interface StyleDriftReport {
   segments: StyleDriftSegment[];
 }
 
+// 人物塑造（灵感③ advisory 单维）：只读建议，不阻断验收（无裁决 UI）。
+// 以人物卡 motive/appearance 为客观锚点；每条带逐字引用 quote + 置信度 confidence。
+export interface CharacterizationIssue {
+  character: string;
+  aspect: string;
+  where: string;
+  quote: string;
+  diagnosis: string;
+  suggestion: string;
+  confidence: number;
+}
+
+export interface CharacterizationReport {
+  issues: CharacterizationIssue[];
+}
+
 export interface SectionEvent {
   event: "section";
   data: { name: string; status: SectionStatus };
@@ -98,6 +114,10 @@ export interface StyleEvent {
     }[];
   };
 }
+export interface CharacterizationEvent {
+  event: "characterization";
+  data: { issues: CharacterizationIssue[] };
+}
 export interface DoneEvent {
   event: "done";
   data: { length: number };
@@ -112,6 +132,7 @@ export type ReviewSseEvent =
   | ForeshadowEvent
   | PaceEvent
   | StyleEvent
+  | CharacterizationEvent
   | DoneEvent
   | ErrorEvent;
 
@@ -121,6 +142,7 @@ const KNOWN_EVENTS = new Set([
   "foreshadow",
   "pace",
   "style",
+  "characterization",
   "done",
   "error",
 ]);
@@ -176,6 +198,34 @@ function asPaceIssues(v: unknown): PaceIssue[] {
   );
 }
 
+function asNumber(v: unknown, fallback = 0): number {
+  return typeof v === "number" && Number.isFinite(v) ? v : fallback;
+}
+
+function asStringField(v: unknown): string {
+  return typeof v === "string" ? v : "";
+}
+
+// 收窄人物塑造 issues（每字段安全默认，advisory 韧性；缺字段不丢整条）。
+export function asCharacterizationIssues(v: unknown): CharacterizationIssue[] {
+  if (!Array.isArray(v)) return [];
+  return v.flatMap((x) =>
+    isRecord(x)
+      ? [
+          {
+            character: asStringField(x.character),
+            aspect: asStringField(x.aspect),
+            where: asStringField(x.where),
+            quote: asStringField(x.quote),
+            diagnosis: asStringField(x.diagnosis),
+            suggestion: asStringField(x.suggestion),
+            confidence: asNumber(x.confidence),
+          },
+        ]
+      : [],
+  );
+}
+
 function asStyleSegments(v: unknown): StyleEvent["data"]["segments"] {
   if (!Array.isArray(v)) return [];
   return v.flatMap((x) =>
@@ -248,6 +298,11 @@ function narrowReviewEvent(event: string, data: unknown): ReviewSseEvent | null
             },
           }
         : null;
+    case "characterization":
+      return {
+        event: "characterization",
+        data: { issues: asCharacterizationIssues(data.issues) },
+      };
     case "done":
       return typeof data.length === "number"
         ? { event: "done", data: { length: data.length } }
@@ -320,6 +375,7 @@ export interface ReviewStreamState {
   foreshadow: ForeshadowSuggestion[];
   pace: PaceReport | null;
   style: StyleDriftReport | null;
+  characterization: CharacterizationReport | null;
   error: { code: string; message: string; request_id?: string | null } | null;
 }
 
@@ -330,6 +386,7 @@ export const initialReviewState: ReviewStreamState = {
   foreshadow: [],
   pace: null,
   style: null,
+  characterization: null,
   error: null,
 };
 
@@ -377,6 +434,12 @@ export function reduceReview(
           })),
         },
       };
+    case "characterization":
+      return {
+        ...state,
+        phase: "reviewing",
+        characterization: { issues: event.data.issues.map((i) => ({ ...i })) },
+      };
     case "done":
       return { ...state, phase: "done" };
     case "error":

apps/web/lib/review/useReviewStream.test.ts

@@ -222,6 +222,19 @@ describe("useReviewStream", () => {
       foreshadow: [{ kind: "resolved" as const, title: "旧伏笔" }],
       pace: { water: [], hook: false, beat_map: [1] },
       style: { score: 0.9, segments: [] },
+      characterization: {
+        issues: [
+          {
+            character: "沈砚",
+            aspect: "动机一致性",
+            where: "第3段",
+            quote: "他忽然转身。",
+            diagnosis: "动机断裂",
+            suggestion: "扣回护妹动机",
+            confidence: 0.8,
+          },
+        ],
+      },
     };
 
     act(() => result.current.seed(seedData));
@@ -230,6 +243,9 @@ describe("useReviewStream", () => {
     expect(result.current.state.foreshadow).toEqual(seedData.foreshadow);
     expect(result.current.state.pace).toEqual(seedData.pace);
     expect(result.current.state.style).toEqual(seedData.style);
+    expect(result.current.state.characterization).toEqual(
+      seedData.characterization,
+    );
     expect(result.current.state.phase).toBe("idle");
   });
 });

apps/web/lib/review/useReviewStream.ts

@@ -7,6 +7,7 @@ import {
   ReviewFrameBuffer,
   initialReviewState,
   reduceReview,
+  type CharacterizationReport,
   type ForeshadowSuggestion,
   type PaceReport,
   type ReviewConflict,
@@ -19,6 +20,7 @@ export interface ReviewSeed {
   foreshadow: ForeshadowSuggestion[];
   pace: PaceReport | null;
   style: StyleDriftReport | null;
+  characterization: CharacterizationReport | null;
 }
 
 type Action =
@@ -49,6 +51,7 @@ function reducer(state: ReviewStreamState, action: Action): ReviewStreamState {
         foreshadow: action.seed.foreshadow,
         pace: action.seed.pace,
         style: action.seed.style,
+        characterization: action.seed.characterization,
       };
     default:
       return state;

apps/web/lib/wizard/useProjectPlan.test.ts

@@ -0,0 +1,95 @@
+// @vitest-environment jsdom
+import { act, renderHook } from "@testing-library/react";
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import type { ProjectPlanGenerateRequest } from "@/lib/api/types";
+
+import { useProjectPlan } from "./useProjectPlan";
+
+// 后端客户端与 Toast 是 hook 的外部副作用边界，单测一律 mock。
+const post = vi.fn();
+const toast = vi.fn();
+vi.mock("@/lib/api/client", () => ({ api: { POST: (...a: unknown[]) => post(...a) } }));
+vi.mock("@/components/Toast", () => ({ useToast: () => toast }));
+
+const seed: ProjectPlanGenerateRequest = {
+  genre: "仙侠",
+  logline: "废柴逆袭封神。",
+};
+
+describe("useProjectPlan", () => {
+  beforeEach(() => {
+    post.mockReset();
+    toast.mockReset();
+  });
+  afterEach(() => vi.clearAllMocks());
+
+  it("初始为 idle、无方案", () => {
+    const { result } = renderHook(() => useProjectPlan());
+    expect(result.current.status).toBe("idle");
+    expect(result.current.plan).toBeNull();
+  });
+
+  it("生成成功：status 走到 done 并填充方案", async () => {
+    const plan = {
+      title_candidates: ["逐光而行"],
+      setting: "仙门世界",
+      narrative_structure: "立钩",
+      story_core: "逆袭",
+      ending_design: "归隐",
+      tone: "热血",
+    };
+    post.mockResolvedValue({ data: plan, error: null });
+
+    const { result } = renderHook(() => useProjectPlan());
+    await act(async () => {
+      await result.current.generate(seed);
+    });
+
+    expect(result.current.status).toBe("done");
+    expect(result.current.plan).toEqual(plan);
+    expect(toast).not.toHaveBeenCalled();
+  });
+
+  it("后端返回 error：status=error 且弹错误 toast", async () => {
+    post.mockResolvedValue({ data: null, error: { detail: "配额不足" } });
+    const { result } = renderHook(() => useProjectPlan());
+    await act(async () => {
+      await result.current.generate(seed);
+    });
+    expect(result.current.status).toBe("error");
+    expect(result.current.plan).toBeNull();
+    expect(toast).toHaveBeenCalledWith(expect.any(String), "error");
+  });
+
+  it("请求抛异常：status=error 且弹网络异常 toast", async () => {
+    post.mockRejectedValue(new Error("network down"));
+    const { result } = renderHook(() => useProjectPlan());
+    await act(async () => {
+      await result.current.generate(seed);
+    });
+    expect(result.current.status).toBe("error");
+    expect(toast).toHaveBeenCalledWith("生成请求异常，请检查网络。", "error");
+  });
+
+  it("reset 清回 idle 与空方案", async () => {
+    post.mockResolvedValue({
+      data: {
+        title_candidates: ["书"],
+        setting: "",
+        narrative_structure: "",
+        story_core: "",
+        ending_design: "",
+        tone: "",
+      },
+      error: null,
+    });
+    const { result } = renderHook(() => useProjectPlan());
+    await act(async () => {
+      await result.current.generate(seed);
+    });
+    act(() => result.current.reset());
+    expect(result.current.status).toBe("idle");
+    expect(result.current.plan).toBeNull();
+  });
+});

apps/web/lib/wizard/useProjectPlan.ts

@@ -0,0 +1,59 @@
+"use client";
+
+import { useCallback, useState } from "react";
+
+import { useToast } from "@/components/Toast";
+import { api } from "@/lib/api/client";
+import { generationErrorMessage } from "@/lib/generation/cards";
+import type {
+  ProjectPlanGenerateRequest,
+  ProjectPlanView,
+} from "@/lib/api/types";
+
+export type ProjectPlanStatus = "idle" | "generating" | "done" | "error";
+
+export interface UseProjectPlan {
+  status: ProjectPlanStatus;
+  plan: ProjectPlanView | null;
+  generate: (seed: ProjectPlanGenerateRequest) => Promise<void>;
+  reset: () => void;
+}
+
+// AI 立项方案生成器：POST /skills/project-plan/generate → 结构化方案预览（不入库）。
+// 向导阶段调用（project 未建）；种子门控在后端亦校验（缺 genre/logline → 422）。
+export function useProjectPlan(): UseProjectPlan {
+  const [status, setStatus] = useState<ProjectPlanStatus>("idle");
+  const [plan, setPlan] = useState<ProjectPlanView | null>(null);
+  const toast = useToast();
+
+  const generate = useCallback<UseProjectPlan["generate"]>(
+    async (seed) => {
+      setStatus("generating");
+      setPlan(null);
+      try {
+        const { data, error } = await api.POST(
+          "/skills/project-plan/generate",
+          { body: seed },
+        );
+        if (error || !data) {
+          setStatus("error");
+          toast(generationErrorMessage(error), "error");
+          return;
+        }
+        setPlan(data);
+        setStatus("done");
+      } catch {
+        setStatus("error");
+        toast("生成请求异常，请检查网络。", "error");
+      }
+    },
+    [toast],
+  );
+
+  const reset = useCallback((): void => {
+    setStatus("idle");
+    setPlan(null);
+  }, []);
+
+  return { status, plan, generate, reset };
+}

apps/web/lib/wizard/wizard.test.ts

@@ -1,16 +1,31 @@
 import { describe, expect, it } from "vitest";
 
+import type { ProjectPlanView } from "@/lib/api/types";
+
 import {
+  applyPlanPatch,
   canAdvance,
+  canGeneratePlan,
   canSubmit,
   clampStep,
   emptyWizardForm,
+  mapPlanResultToWizardForm,
   STEP_TITLES,
   toCreateRequest,
+  toPlanSeedRequest,
   WIZARD_STEPS,
   type WizardForm,
 } from "./wizard";
 
+const fullPlan: ProjectPlanView = {
+  title_candidates: ["逐光而行", "剑试九霄"],
+  setting: "上古仙门倾轧的洞天世界",
+  narrative_structure: "黄金三章立钩 → 拜师 → 宗门大比爆发",
+  story_core: "废柴少年觉醒禁忌血脉，逆势封神的代价",
+  ending_design: "登顶后归隐，自我救赎收束",
+  tone: "热血",
+};
+
 describe("wizard step metadata", () => {
   it("provides one title per step", () => {
     expect(STEP_TITLES).toHaveLength(WIZARD_STEPS);
@@ -57,6 +72,9 @@ describe("toCreateRequest", () => {
       premise: "  ",
       protagonist: "",
       theme: "抗争",
+      tone: "",
+      endingType: "",
+      narrativePov: "",
     };
     expect(toCreateRequest(form)).toEqual({
       title: "逐光而行",
@@ -66,9 +84,34 @@ describe("toCreateRequest", () => {
       theme: "抗争",
       selling_points: ["逆袭", "系统流"],
       structure: "三幕",
+      tone: null,
+      ending_type: null,
+      narrative_pov: null,
     });
   });
 
+  // ④ 立项新增基调/结局/视角：camelCase 表单 → snake_case 请求，空值落 null。
+  it("maps tone, endingType, narrativePov to snake_case fields", () => {
+    const form: WizardForm = {
+      ...emptyWizardForm,
+      title: "书",
+      tone: "热血",
+      endingType: "HE",
+      narrativePov: "第一人称",
+    };
+    const req = toCreateRequest(form);
+    expect(req.tone).toBe("热血");
+    expect(req.ending_type).toBe("HE");
+    expect(req.narrative_pov).toBe("第一人称");
+  });
+
+  it("nulls empty tone, endingType, narrativePov", () => {
+    const req = toCreateRequest({ ...emptyWizardForm, title: "书" });
+    expect(req.tone).toBeNull();
+    expect(req.ending_type).toBeNull();
+    expect(req.narrative_pov).toBeNull();
+  });
+
   // QA H1 回归：立意（premise）与主角/金手指（protagonist）是两个独立输入，
   // 不能互相覆盖——提交时各占一段合并进 premise。
   it("merges premise and protagonist into premise without overwriting", () => {
@@ -94,6 +137,123 @@ describe("toCreateRequest", () => {
   });
 });
 
+// ⑤ AI 立项方案 → 向导映射（含 ④ 字段 tone；错配字段折进 premise，不静默丢内容）。
+describe("mapPlanResultToWizardForm", () => {
+  it("maps title candidate, structure and tone to their wizard homes", () => {
+    const patch = mapPlanResultToWizardForm(fullPlan);
+    // 书名候选 → title（取首个）。
+    expect(patch.title).toBe("逐光而行");
+    // 叙事结构 → structure。
+    expect(patch.structure).toBe("黄金三章立钩 → 拜师 → 宗门大比爆发");
+    // ④ 字段：基调 → tone。
+    expect(patch.tone).toBe("热血");
+  });
+
+  it("folds setting and ending design into premise without dropping content", () => {
+    const patch = mapPlanResultToWizardForm(fullPlan);
+    // 故事核心 + 时空背景 + 结局设计 逐段拼进 premise（无独立向导字段，不静默丢）。
+    expect(patch.premise).toContain("废柴少年觉醒禁忌血脉，逆势封神的代价");
+    expect(patch.premise).toContain("时空背景：上古仙门倾轧的洞天世界");
+    expect(patch.premise).toContain("结局设计：登顶后归隐，自我救赎收束");
+  });
+
+  it("does not fabricate endingType or narrativePov (④ presets left for author)", () => {
+    // 结局取向/叙事视角是预设枚举，无法从自由文本可靠反推 → 不进 patch（不臆造）。
+    const patch = mapPlanResultToWizardForm(fullPlan);
+    expect(patch.endingType).toBeUndefined();
+    expect(patch.narrativePov).toBeUndefined();
+  });
+
+  it("omits empty fields from the patch (parse resilience)", () => {
+    const empty: ProjectPlanView = {
+      title_candidates: [],
+      setting: "",
+      narrative_structure: "",
+      story_core: "",
+      ending_design: "",
+      tone: "",
+    };
+    expect(mapPlanResultToWizardForm(empty)).toEqual({});
+  });
+});
+
+// applyPlanPatch：按字段并入、仅填空字段（保护作者已编辑内容，非 bulk 覆盖）。
+describe("applyPlanPatch", () => {
+  it("fills only empty fields, preserving author-edited values", () => {
+    const form: WizardForm = {
+      ...emptyWizardForm,
+      title: "作者已取的书名",
+      genre: "仙侠",
+    };
+    const patch = { title: "AI书名", structure: "三幕", tone: "热血" };
+    const next = applyPlanPatch(form, patch);
+    // 已编辑的 title 受保护、不被覆盖；空的 structure/tone 才回填。
+    expect(next.title).toBe("作者已取的书名");
+    expect(next.structure).toBe("三幕");
+    expect(next.tone).toBe("热血");
+    // 未在 patch 中的字段原样保留。
+    expect(next.genre).toBe("仙侠");
+  });
+
+  it("treats whitespace-only fields as empty (fillable)", () => {
+    const form: WizardForm = { ...emptyWizardForm, premise: "   " };
+    const next = applyPlanPatch(form, { premise: "AI 立意" });
+    expect(next.premise).toBe("AI 立意");
+  });
+
+  it("returns a new object without mutating the input form", () => {
+    const form: WizardForm = { ...emptyWizardForm };
+    const next = applyPlanPatch(form, { tone: "热血" });
+    expect(next).not.toBe(form);
+    expect(form.tone).toBe("");
+  });
+});
+
+// canGeneratePlan：种子门控（题材 + logline 都非空）。
+describe("canGeneratePlan", () => {
+  it("requires both genre and logline", () => {
+    expect(canGeneratePlan(emptyWizardForm)).toBe(false);
+    expect(
+      canGeneratePlan({ ...emptyWizardForm, genre: "仙侠" }),
+    ).toBe(false);
+    expect(
+      canGeneratePlan({ ...emptyWizardForm, logline: "一句话" }),
+    ).toBe(false);
+    expect(
+      canGeneratePlan({ ...emptyWizardForm, genre: "仙侠", logline: "一句话" }),
+    ).toBe(true);
+  });
+
+  it("rejects whitespace-only seeds", () => {
+    expect(
+      canGeneratePlan({ ...emptyWizardForm, genre: "  ", logline: "  " }),
+    ).toBe(false);
+  });
+});
+
+// toPlanSeedRequest：向导草稿 → 生成请求（snake_case，空值省略）。
+describe("toPlanSeedRequest", () => {
+  it("serializes seed and filled fields, omitting empties", () => {
+    const form: WizardForm = {
+      ...emptyWizardForm,
+      genre: "都市",
+      logline: "外卖骑手觉醒读心术。",
+      title: "听风者",
+      sellingPoints: ["逆袭"],
+      endingType: "HE",
+    };
+    const req = toPlanSeedRequest(form);
+    expect(req.genre).toBe("都市");
+    expect(req.logline).toBe("外卖骑手觉醒读心术。");
+    expect(req.title).toBe("听风者");
+    expect(req.ending_type).toBe("HE");
+    expect(req.selling_points).toEqual(["逆袭"]);
+    // 空字段省略（undefined），不发空串。
+    expect(req.premise).toBeUndefined();
+    expect(req.tone).toBeUndefined();
+  });
+});
+
 // wizard→create 流程：归一后的请求体经 mock 客户端提交，返回新建项目 id。
 describe("wizard create flow (mocked client)", () => {
   it("submits the normalized body and resolves a project id", async () => {

apps/web/lib/wizard/wizard.ts

@@ -1,4 +1,8 @@
-import type { ProjectCreateRequest } from "@/lib/api/types";
+import type {
+  ProjectCreateRequest,
+  ProjectPlanGenerateRequest,
+  ProjectPlanView,
+} from "@/lib/api/types";
 
 // 立项向导 5 步（UX §6.2）。表单状态用字符串/数组，提交时归一为 ProjectCreateRequest。
 export interface WizardForm {
@@ -10,6 +14,10 @@ export interface WizardForm {
   premise: string;
   protagonist: string;
   theme: string;
+  // 立项书级常量（灵感④）：基调 / 结局取向 / 叙事视角。
+  tone: string;
+  endingType: string;
+  narrativePov: string;
 }
 
 export const WIZARD_STEPS = 5;
@@ -32,11 +40,18 @@ export const emptyWizardForm: WizardForm = {
   premise: "",
   protagonist: "",
   theme: "",
+  tone: "",
+  endingType: "",
+  narrativePov: "",
 };
 
 export const GENRES = ["玄幻", "仙侠", "都市", "科幻", "历史", "悬疑"];
 export const STRUCTURES = ["三幕", "故事圈", "雪花"];
 export const SELLING_POINT_PRESETS = ["逆袭", "打脸", "系统流", "双男主", "群像"];
+// 立项书级常量预设（灵感④ / D 枚举）——free text 单选，命不中预设可留空。
+export const TONES = ["热血", "轻松", "沉重", "治愈", "暗黑", "爽文"];
+export const ENDING_TYPES = ["HE", "BE", "开放", "正剧"];
+export const NARRATIVE_POVS = ["第一人称", "第三人称限知", "第三人称全知", "多视角"];
 
 // 每一步可否前进：仅第 1 步（书名）必填，其余可空着继续（草稿式立项）。
 export function canAdvance(step: number, form: WizardForm): boolean {
@@ -55,6 +70,82 @@ export function clampStep(step: number): number {
   return step;
 }
 
+// AI 立项方案 → 向导表单补丁（灵感⑤）。按字段回填：
+// - 有对应向导输入的字段直接映射（书名候选→title、叙事结构→structure、基调→tone）；
+// - 无独立向导输入的字段（时空背景/结局设计）折进立意/总纲（premise，带标签），**不静默丢内容**；
+// - 结局取向/叙事视角这类预设枚举无法从自由文本可靠反推 → 不臆造，留给作者手选（不进 patch）。
+// 返回补丁而非整表覆盖，由 applyPlanPatch 按字段并入（保护已编辑字段）。
+export function mapPlanResultToWizardForm(
+  plan: ProjectPlanView,
+): Partial<WizardForm> {
+  const patch: Partial<WizardForm> = {};
+  const firstTitle = plan.title_candidates?.[0]?.trim();
+  if (firstTitle) patch.title = firstTitle;
+  const structure = plan.narrative_structure.trim();
+  if (structure) patch.structure = structure;
+  const tone = plan.tone.trim();
+  if (tone) patch.tone = tone;
+  // 立意/总纲 = 故事核心 + 无独立向导字段的时空背景/结局设计（带标签，逐段拼接不丢内容）。
+  const premiseParts: string[] = [];
+  const core = plan.story_core.trim();
+  if (core) premiseParts.push(core);
+  const setting = plan.setting.trim();
+  if (setting) premiseParts.push(`时空背景：${setting}`);
+  const ending = plan.ending_design.trim();
+  if (ending) premiseParts.push(`结局设计：${ending}`);
+  if (premiseParts.length > 0) patch.premise = premiseParts.join("\n\n");
+  return patch;
+}
+
+// WizardForm 中值为字符串的字段键（排除 sellingPoints: string[]）——方案 patch 只回填这些。
+type WizardStringKey = {
+  [K in keyof WizardForm]: WizardForm[K] extends string ? K : never;
+}[keyof WizardForm];
+
+// 把方案补丁按字段并入向导表单：**仅回填当前为空的字段**（保护作者已编辑内容，
+// 非 bulk 覆盖）。只处理字符串字段（方案 patch 只产字符串值）；空 = trim 后为空串。
+export function applyPlanPatch(
+  form: WizardForm,
+  patch: Partial<WizardForm>,
+): WizardForm {
+  const next: WizardForm = { ...form };
+  for (const key of Object.keys(patch) as WizardStringKey[]) {
+    const value = patch[key];
+    // 守卫短路：非字符串值（如误传数组键）直接跳过，绝不触达 .trim()。
+    if (typeof value === "string" && form[key].trim() === "") {
+      next[key] = value;
+    }
+  }
+  return next;
+}
+
+// AI 立项方案的种子门控：题材 + 一句话故事（logline）都非空才允许生成（后端亦校验，防 slop）。
+export function canGeneratePlan(form: WizardForm): boolean {
+  return form.genre.trim().length > 0 && form.logline.trim().length > 0;
+}
+
+// 向导草稿 → 立项方案生成请求（snake_case；空值省略）。genre+logline 为种子，其余有则贴合。
+export function toPlanSeedRequest(
+  form: WizardForm,
+): ProjectPlanGenerateRequest {
+  const trimOrUndef = (s: string): string | undefined => {
+    const v = s.trim();
+    return v.length > 0 ? v : undefined;
+  };
+  return {
+    genre: trimOrUndef(form.genre),
+    logline: trimOrUndef(form.logline),
+    title: trimOrUndef(form.title),
+    premise: trimOrUndef(form.premise),
+    theme: trimOrUndef(form.theme),
+    structure: trimOrUndef(form.structure),
+    tone: trimOrUndef(form.tone),
+    ending_type: trimOrUndef(form.endingType),
+    narrative_pov: trimOrUndef(form.narrativePov),
+    selling_points: form.sellingPoints,
+  };
+}
+
 // 归一为后端请求体（snake_case，空值落 null/省略）。
 export function toCreateRequest(form: WizardForm): ProjectCreateRequest {
   const trim = (s: string): string | null => {
@@ -75,5 +166,8 @@ export function toCreateRequest(form: WizardForm): ProjectCreateRequest {
     theme: trim(form.theme),
     selling_points: form.sellingPoints,
     structure: trim(form.structure),
+    tone: trim(form.tone),
+    ending_type: trim(form.endingType),
+    narrative_pov: trim(form.narrativePov),
   };
 }

apps/web/package.json

@@ -13,6 +13,8 @@
     "gen:api": "node scripts/gen-api.mjs"
   },
   "dependencies": {
+    "@dagrejs/dagre": "^3.0.0",
+    "@xyflow/react": "^12.11.2",
     "lucide-react": "^1.21.0",
     "next": "15.1.3",
     "openapi-fetch": "^0.13.0",

apps/web/pnpm-lock.yaml

@@ -8,6 +8,12 @@ importers:
 
   .:
     dependencies:
+      '@dagrejs/dagre':
+        specifier: ^3.0.0
+        version: 3.0.0
+      '@xyflow/react':
+        specifier: ^12.11.2
+        version: 12.11.2(@types/react-dom@19.0.0)(@types/react@19.0.0)(react-dom@19.0.0(react@19.0.0))(react@19.0.0)
       lucide-react:
         specifier: ^1.21.0
         version: 1.21.0(react@19.0.0)
@@ -166,6 +172,12 @@ packages:
     resolution: {integrity: sha512-QxULHAm7cNu72w97JUNCBFODFaXpbDg+dP8b/oWFAZ2MTRppA3U00Y2L1HqaS4J6yBqxwa/Y3nMBaxVKbB/NsA==}
     engines: {node: '>=20.19.0'}
 
+  '@dagrejs/dagre@3.0.0':
+    resolution: {integrity: sha512-ZzhnTy1rfuoew9Ez3EIw4L2znPGnYYhfn8vc9c4oB8iw6QAsszbiU0vRhlxWPFnmmNSFAkrYeF1PhM5m4lAN0Q==}
+
+  '@dagrejs/graphlib@4.0.1':
+    resolution: {integrity: sha512-IvcV6FduIIAmLwnH+yun+QtV36SC7mERqa86aClNqmMN09WhmPPYU8ckHrZBozErf+UvHPWOTJYaGYiIcs0DgA==}
+
   '@emnapi/core@1.10.0':
     resolution: {integrity: sha512-yq6OkJ4p82CAfPl0u9mQebQHKPJkY7WrIuk205cTYnYe+k2Z8YBh11FrbRG/H6ihirqcacOgl2BIO8oyMQLeXw==}
 
@@ -790,6 +802,24 @@ packages:
   '@types/aria-query@5.0.4':
     resolution: {integrity: sha512-rfT93uj5s0PRL7EzccGMs3brplhcrghnDoV26NqKhCAS1hVo+WdNsPvE/yb6ilfr5hi2MEk6d5EWJTKdxg8jVw==}
 
+  '@types/d3-color@3.1.3':
+    resolution: {integrity: sha512-iO90scth9WAbmgv7ogoq57O9YpKmFBbmoEoCHDB2xMBY0+/KVrqAaCDyCE16dUspeOvIxFFRI+0sEtqDqy2b4A==}
+
+  '@types/d3-drag@3.0.7':
+    resolution: {integrity: sha512-HE3jVKlzU9AaMazNufooRJ5ZpWmLIoc90A37WU2JMmeq28w1FQqCZswHZ3xR+SuxYftzHq6WU6KJHvqxKzTxxQ==}
+
+  '@types/d3-interpolate@3.0.4':
+    resolution: {integrity: sha512-mgLPETlrpVV1YRJIglr4Ez47g7Yxjl1lj7YKsiMCb27VJH9W8NVM6Bb9d8kkpG/uAQS5AmbA48q2IAolKKo1MA==}
+
+  '@types/d3-selection@3.0.11':
+    resolution: {integrity: sha512-bhAXu23DJWsrI45xafYpkQ4NtcKMwWnAC/vKrd2l+nxMFuvOT3XMYTIj2opv8vq8AO5Yh7Qac/nSeP/3zjTK0w==}
+
+  '@types/d3-transition@3.0.9':
+    resolution: {integrity: sha512-uZS5shfxzO3rGlu0cC3bjmMFKsXv+SmZZcgp0KD22ts4uGXp5EVYGzu/0YdwZeKmddhcAccYtREJKkPfXkZuCg==}
+
+  '@types/d3-zoom@3.0.8':
+    resolution: {integrity: sha512-iqMC4/YlFCSlO8+2Ii1GGGliCAY4XdeG748w5vQUbevlbDu0zSjH/+jojorQVBK/se0j6DUFNPBGSqD3YWYnDw==}
+
   '@types/estree@1.0.9':
     resolution: {integrity: sha512-GhdPgy1el4/ImP05X05Uw4cw2/M93BCUmnEvWZNStlCzEKME4Fkk+YpoA5OiHNQmoS7Cafb8Xa3Pya8m1Qrzeg==}
 
@@ -1025,6 +1055,22 @@ packages:
   '@vitest/utils@2.1.9':
     resolution: {integrity: sha512-v0psaMSkNJ3A2NMrUEHFRzJtDPFn+/VWZ5WxImB21T9fjucJRmS7xCS3ppEnARb9y11OAzaD+P2Ps+b+BGX5iQ==}
 
+  '@xyflow/react@12.11.2':
+    resolution: {integrity: sha512-eLAlDWJfWnQEhJwGMjlWdAXO9eYllKpliUmPQlAmOLxz6mExXuzMVDUKLMquixgkrtmMFFtug3jGKmYYld12cA==}
+    peerDependencies:
+      '@types/react': '>=17'
+      '@types/react-dom': '>=17'
+      react: '>=17'
+      react-dom: '>=17'
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      '@types/react-dom':
+        optional: true
+
+  '@xyflow/system@0.0.79':
+    resolution: {integrity: sha512-czLyOh91NF0hIzbNzwi8I6GlqG23BHh2435OddfI6uiaLH3xdrdygO93gqgH1Bv9mhy8XPFQJOBn1FTq4LvEWA==}
+
   acorn-jsx@5.3.2:
     resolution: {integrity: sha512-rq9s+JNhf0IChjtDXxllJ7g41oZk5SlXtp0LHwyA5cejwn7vKmKp4pPri6YEePv2PU65sAsegbXtIinmDFDXgQ==}
     peerDependencies:
@@ -1236,6 +1282,9 @@ packages:
     resolution: {integrity: sha512-7VT13fmjotKpGipCW9JEQAusEPE+Ei8nl6/g4FBAmIm0GOOLMua9NDDo/DWp0ZAxCr3cPq5ZpBqmPAQgDda2Pw==}
     engines: {node: '>= 8.10.0'}
 
+  classcat@5.0.5:
+    resolution: {integrity: sha512-JhZUT7JFcQy/EzW605k/ktHtncoo9vnyW/2GspNYwFlN1C/WmjuV/xtS04e9SOkL2sTdw0VAZ2UGCcQ9lR6p6w==}
+
   client-only@0.0.1:
     resolution: {integrity: sha512-IV3Ou0jSMzZrd3pZ48nLkT9DA7Ag1pnPzaiQhpW7c3RbcqqzvzzVu+L8gfqMp/8IM2MQtSiqaCxrrcfu8I8rMA==}
 
@@ -1279,6 +1328,44 @@ packages:
   csstype@3.2.3:
     resolution: {integrity: sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ==}
 
+  d3-color@3.1.0:
+    resolution: {integrity: sha512-zg/chbXyeBtMQ1LbD/WSoW2DpC3I0mpmPdW+ynRTj/x2DAWYrIY7qeZIHidozwV24m4iavr15lNwIwLxRmOxhA==}
+    engines: {node: '>=12'}
+
+  d3-dispatch@3.0.1:
+    resolution: {integrity: sha512-rzUyPU/S7rwUflMyLc1ETDeBj0NRuHKKAcvukozwhshr6g6c5d8zh4c2gQjY2bZ0dXeGLWc1PF174P2tVvKhfg==}
+    engines: {node: '>=12'}
+
+  d3-drag@3.0.0:
+    resolution: {integrity: sha512-pWbUJLdETVA8lQNJecMxoXfH6x+mO2UQo8rSmZ+QqxcbyA3hfeprFgIT//HW2nlHChWeIIMwS2Fq+gEARkhTkg==}
+    engines: {node: '>=12'}
+
+  d3-ease@3.0.1:
+    resolution: {integrity: sha512-wR/XK3D3XcLIZwpbvQwQ5fK+8Ykds1ip7A2Txe0yxncXSdq1L9skcG7blcedkOX+ZcgxGAmLX1FrRGbADwzi0w==}
+    engines: {node: '>=12'}
+
+  d3-interpolate@3.0.1:
+    resolution: {integrity: sha512-3bYs1rOD33uo8aqJfKP3JWPAibgw8Zm2+L9vBKEHJ2Rg+viTR7o5Mmv5mZcieN+FRYaAOWX5SJATX6k1PWz72g==}
+    engines: {node: '>=12'}
+
+  d3-selection@3.0.0:
+    resolution: {integrity: sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==}
+    engines: {node: '>=12'}
+
+  d3-timer@3.0.1:
+    resolution: {integrity: sha512-ndfJ/JxxMd3nw31uyKoY2naivF+r29V+Lc0svZxe1JvvIRmi8hUsrMvdOwgS1o6uBHmiz91geQ0ylPP0aj1VUA==}
+    engines: {node: '>=12'}
+
+  d3-transition@3.0.1:
+    resolution: {integrity: sha512-ApKvfjsSR6tg06xrL434C0WydLr7JewBB3V+/39RMHsaXTOG0zmt/OAXeng5M5LBm0ojmxJrpomQVZ1aPvBL4w==}
+    engines: {node: '>=12'}
+    peerDependencies:
+      d3-selection: 2 - 3
+
+  d3-zoom@3.0.0:
+    resolution: {integrity: sha512-b8AmV3kfQaqWAuacbPuNbL6vahnOJflOhexLzMMNLga62+/nh0JzvJ0aO/5a5MVgUFGS7Hu1P9P03o3fJkDCyw==}
+    engines: {node: '>=12'}
+
   damerau-levenshtein@1.0.8:
     resolution: {integrity: sha512-sdQSFB7+llfUcQHUQO3+B8ERRj0Oa4w9POWMI/puGtuf7gFywGmkaLCElnudfTiKZV+NvHqL0ifzdrI8Ro7ESA==}
 
@@ -2662,6 +2749,11 @@ packages:
   uri-js@4.4.1:
     resolution: {integrity: sha512-7rKUyy33Q1yc98pQ1DAmLtwX109F7TIfWlW1Ydo8Wl1ii1SeHieeh0HHfPeL2fMXK6z0s8ecKs9frCuLJvndBg==}
 
+  use-sync-external-store@1.6.0:
+    resolution: {integrity: sha512-Pp6GSwGP/NrPIrxVFAIkOQeyw8lFenOHijQWkUTrDvrF4ALqylP2C/KCkeS9dpUM3KvYRQhna5vt7IL95+ZQ9w==}
+    peerDependencies:
+      react: ^16.8.0 || ^17.0.0 || ^18.0.0 || ^19.0.0
+
   util-deprecate@1.0.2:
     resolution: {integrity: sha512-EPD5q1uXyFxJpCrLnCc1nHnq3gOa6DZBocAIiI2TaSCA7VCJ1UJDMagCzIkXNsUYfD1daK//LTEQ8xiIbrHtcw==}
 
@@ -2798,6 +2890,21 @@ packages:
     resolution: {integrity: sha512-rVksvsnNCdJ/ohGc6xgPwyN8eheCxsiLM8mxuE/t/mOVqJewPuO1miLpTHQiRgTKCLexL4MeAFVagts7HmNZ2Q==}
     engines: {node: '>=10'}
 
+  zustand@4.5.7:
+    resolution: {integrity: sha512-CHOUy7mu3lbD6o6LJLfllpjkzhHXSBlX8B9+qPddUsIfeF5S/UZ5q0kmCsnRqT1UHFQZchNFDDzMbQsuesHWlw==}
+    engines: {node: '>=12.7.0'}
+    peerDependencies:
+      '@types/react': '>=16.8'
+      immer: '>=9.0.6'
+      react: '>=16.8'
+    peerDependenciesMeta:
+      '@types/react':
+        optional: true
+      immer:
+        optional: true
+      react:
+        optional: true
+
 snapshots:
 
   '@alloc/quick-lru@5.2.0': {}
@@ -2878,6 +2985,12 @@ snapshots:
 
   '@csstools/css-tokenizer@4.0.0': {}
 
+  '@dagrejs/dagre@3.0.0':
+    dependencies:
+      '@dagrejs/graphlib': 4.0.1
+
+  '@dagrejs/graphlib@4.0.1': {}
+
   '@emnapi/core@1.10.0':
     dependencies:
       '@emnapi/wasi-threads': 1.2.1
@@ -3322,6 +3435,27 @@ snapshots:
 
   '@types/aria-query@5.0.4': {}
 
+  '@types/d3-color@3.1.3': {}
+
+  '@types/d3-drag@3.0.7':
+    dependencies:
+      '@types/d3-selection': 3.0.11
+
+  '@types/d3-interpolate@3.0.4':
+    dependencies:
+      '@types/d3-color': 3.1.3
+
+  '@types/d3-selection@3.0.11': {}
+
+  '@types/d3-transition@3.0.9':
+    dependencies:
+      '@types/d3-selection': 3.0.11
+
+  '@types/d3-zoom@3.0.8':
+    dependencies:
+      '@types/d3-interpolate': 3.0.4
+      '@types/d3-selection': 3.0.11
+
   '@types/estree@1.0.9': {}
 
   '@types/json-schema@7.0.15': {}
@@ -3559,6 +3693,31 @@ snapshots:
       loupe: 3.2.1
       tinyrainbow: 1.2.0
 
+  '@xyflow/react@12.11.2(@types/react-dom@19.0.0)(@types/react@19.0.0)(react-dom@19.0.0(react@19.0.0))(react@19.0.0)':
+    dependencies:
+      '@xyflow/system': 0.0.79
+      classcat: 5.0.5
+      react: 19.0.0
+      react-dom: 19.0.0(react@19.0.0)
+      zustand: 4.5.7(@types/react@19.0.0)(react@19.0.0)
+    optionalDependencies:
+      '@types/react': 19.0.0
+      '@types/react-dom': 19.0.0
+    transitivePeerDependencies:
+      - immer
+
+  '@xyflow/system@0.0.79':
+    dependencies:
+      '@types/d3-drag': 3.0.7
+      '@types/d3-interpolate': 3.0.4
+      '@types/d3-selection': 3.0.11
+      '@types/d3-transition': 3.0.9
+      '@types/d3-zoom': 3.0.8
+      d3-drag: 3.0.0
+      d3-interpolate: 3.0.1
+      d3-selection: 3.0.0
+      d3-zoom: 3.0.0
+
   acorn-jsx@5.3.2(acorn@8.17.0):
     dependencies:
       acorn: 8.17.0
@@ -3790,6 +3949,8 @@ snapshots:
     optionalDependencies:
       fsevents: 2.3.3
 
+  classcat@5.0.5: {}
+
   client-only@0.0.1: {}
 
   color-convert@2.0.1:
@@ -3831,6 +3992,42 @@ snapshots:
 
   csstype@3.2.3: {}
 
+  d3-color@3.1.0: {}
+
+  d3-dispatch@3.0.1: {}
+
+  d3-drag@3.0.0:
+    dependencies:
+      d3-dispatch: 3.0.1
+      d3-selection: 3.0.0
+
+  d3-ease@3.0.1: {}
+
+  d3-interpolate@3.0.1:
+    dependencies:
+      d3-color: 3.1.0
+
+  d3-selection@3.0.0: {}
+
+  d3-timer@3.0.1: {}
+
+  d3-transition@3.0.1(d3-selection@3.0.0):
+    dependencies:
+      d3-color: 3.1.0
+      d3-dispatch: 3.0.1
+      d3-ease: 3.0.1
+      d3-interpolate: 3.0.1
+      d3-selection: 3.0.0
+      d3-timer: 3.0.1
+
+  d3-zoom@3.0.0:
+    dependencies:
+      d3-dispatch: 3.0.1
+      d3-drag: 3.0.0
+      d3-interpolate: 3.0.1
+      d3-selection: 3.0.0
+      d3-transition: 3.0.1(d3-selection@3.0.0)
+
   damerau-levenshtein@1.0.8: {}
 
   data-urls@7.0.0:
@@ -5532,6 +5729,10 @@ snapshots:
     dependencies:
       punycode: 2.3.1
 
+  use-sync-external-store@1.6.0(react@19.0.0):
+    dependencies:
+      react: 19.0.0
+
   util-deprecate@1.0.2: {}
 
   vite-node@2.1.9(@types/node@22.19.21)(supports-color@10.2.2):
@@ -5686,3 +5887,10 @@ snapshots:
   yargs-parser@21.1.1: {}
 
   yocto-queue@0.1.0: {}
+
+  zustand@4.5.7(@types/react@19.0.0)(react@19.0.0):
+    dependencies:
+      use-sync-external-store: 1.6.0(react@19.0.0)
+    optionalDependencies:
+      '@types/react': 19.0.0
+      react: 19.0.0

apps/web/pnpm-workspace.yaml

@@ -4,6 +4,9 @@ allowBuilds:
   esbuild: true
   sharp: true
   unrs-resolver: true
+minimumReleaseAgeExclude:
+  - '@xyflow/react@12.11.2'
+  - '@xyflow/system@0.0.79'
 onlyBuiltDependencies:
   - esbuild
   - sharp

memory/contracts.md

@@ -137,7 +137,7 @@
 ### C3 扩展（T5.2, 2026-06-19）· 生成/入库 + 读端点（独立 `routers/generation.py`，已注册）　owner @backend　状态: 稳定
 - snake_case；改字段 → **@frontend 必须 `cd apps/web && pnpm gen:api`**（T5.6 前置）。生成走**即时返回**（预览→作者确认→入库，M5-d，非 jobs）。
 - `POST /projects/{project_id}/world/generate` ← `WorldGenerateRequest{brief:str(min1)}` → **200** `WorldGenPreviewResponse{entities:[WorldEntityCardView{type:str,name:str,rules:list[str]}]}`。**预览不入库**（worldbuilder writer 网关 `run_worldbuilder`）；末尾 `commit()` 仅落网关 ledger。项目不存在→404；无凭据→**503** `LLM_UNAVAILABLE`（`get_worldbuilder_gateway` dep 解析阶段拦下）。
-- `POST /projects/{project_id}/characters/generate` ← `CharacterGenerateRequest{brief:str(min1), count:int=1(ge1,le12), role:str|None}` → **200** `CharacterGenPreviewResponse{cards:[CharacterCardView{name,role,traits:list[str],backstory,arc:str,speech_tics:list[str],tags:list[str],relations:[CharacterRelationView{name,kind,note?}]}]}`。**预览不入库**（character-gen writer 网关，注入已有角色防雷同，generated_so_far=[]）；末尾 commit 落 ledger。404/503 同上。
+- `POST /projects/{project_id}/characters/generate` ← `CharacterGenerateRequest{brief:str(min1), count:int=1(ge1,le12), role:str|None, role_mix:dict[str,int]|None}` → **200** `CharacterGenPreviewResponse{cards:[CharacterCardView{name,role,traits:list[str],backstory,arc:str,speech_tics:list[str],tags:list[str],relations:[CharacterRelationView{name,kind,note?}]}]}`。**预览不入库**（character-gen writer 网关，注入已有角色防雷同，generated_so_far=[]）；末尾 commit 落 ledger。404/503 同上。**⑦ 群像按定位配比**：`role_mix`（{定位:数量}）在场时 `count` 取各值之和（忽略客户端 count），每项≥1、总和≤12（`model_validator` 归一化）；缺省退回单一 `role`+`count`。
 - `POST /projects/{project_id}/characters`（入库）← `CharacterIngestRequest{cards:[CharacterCardView](min1), acknowledge_conflicts:bool=false}` → **201** `CharacterIngestResponse{created:list[str], rejected_tables:list[str]}`。**入库 gate**：① `precheck_generated_cards`(continuity 预检, analyst 网关) 比对卡 vs 世界观/已有角色真相源——有冲突且 `acknowledge_conflicts=false` → **409 `CONFLICT_UNRESOLVED`** + `details{conflicts:[{type,where,refs,suggestion}], conflict_count}`（仿 accept gate，不静默入库，不变量#3；`acknowledge_conflicts=true` 即作者裁决放行）。② `partition_writes(character_gen_spec, {"characters":cards})` 白名单过滤（越权表→`rejected_tables` 审计 log+丢弃，character-gen 只声明 writes=characters 故正常为空）。③ 写 `characters` 行（schema list/str → DB JSONB **dict** 形变，见写侧 repo）。404/503 同上。提交边界：precheck 网关 ledger + 角色写侧均只 flush → 端点末尾一次 `commit()`（含冲突 409 路径也 commit 落 precheck usage）。
 - `GET /projects/{project_id}/rules` → **200** `RuleListResponse{rules:[RuleView{level,content}]}`（规则页；复用 C5 读侧 `SqlRulesRepo.all_for_project`，不动 assemble）。
 - `GET /skills`（独立 `skills_router`，prefix `/skills`）→ **200** `SkillListResponse{skills:[SkillView{name,scope,tier,reads:list[str],writes:list[str],genre?}]}`（技能库 UI；经 `get_skill_registry` 读 registry，按 name 升序）。
@@ -346,6 +346,16 @@
 ## 契约变更日志（append-only）
 > 格式：`- [date] @skill 改 Cx：<改了什么> → 影响 <依赖方/任务>`
 
+- [2026-07-06] @llm/@db/@backend/@frontend 改 C3+C4（③ 人物塑造 advisory 审）：新增第五审 **人物塑造**（advisory 单维——仅建议、**不阻断验收**：不进 conflicts、不碰 `assert_conflicts_resolved`、**不入 `REVIEW_RESERVED_NAMES`**，D2）。**DB**：`chapter_reviews` 加 1 列 `characterization`（**JSONB nullable**，迁移 `f6a7b8c9d0e1`，独立于恒 None 的 health_score，nullable 免 backfill）。**契约变更**：`ReviewHistoryItem`（`schemas/projects.py`）+ `ReviewView`（`domain/review_repo.py`，末位默认 None 优雅降级）各加 `characterization: dict|None`；`list_reviews` 路由透传。新 SPEC `characterization_spec`（`tier="analyst"` 不写 model，`reads=("characters",)` 以人物卡 motive/appearance 为客观锚点，`writes=()` 只读，`prompts/characterization.md` **显式忽略文本长度与辞藻华丽度**、引用逐字片段+置信度）+ schema `CharacterizationReview{issues:[CharacterizationIssue{character,aspect,where,quote,diagnosis,suggestion,confidence}]}`（全字段默认值守解析韧性）+ 注册 `SCHEMA_CATALOG["characterization"]`；**计数三处 +1（22→23）**：`specs.py` assert、`test_prompt_loader.py` len 断言、`schema_catalog.py` docstring；**regen `prompt_hashes.json`**（仅加 characterization 一条）。接线链：`graph.py REVIEW_SPECS` 末位追加（并行扇出自动汇入 collect）；`chain/nodes.py ReviewRecordRepo.record` + `collect.py {CHARACTERIZATION 常量, extract_characterization, ReviewRecorder Protocol, collect_reviews}` + `review_repo.py {Protocol.record, SqlReviewRepo.record, _to_view}` 均加 `characterization` kwarg/字段；`sse.py {EVENT_CHARACTERIZATION, characterization_event, _section_result_events elif}`（normalize 按 sorted 键自动纳入——字典序 characterization 排最前，不打乱既有四审断言）。前端 `lib/review/sse.ts`（`CharacterizationIssue`/`CharacterizationReport`/`CharacterizationEvent` + KNOWN_EVENTS + `asCharacterizationIssues` 守卫 + reducer + state）+ `lib/review/history.ts normalizeCharacterization` + `useReviewStream ReviewSeed` + 新 `CharacterizationPanel.tsx`（advisory 只展示、**无裁决 UI**、按置信度降序 + 低置信度折叠控注意力预算）+ `ReviewReport.tsx` 挂面板 + 双 seed 点。**codegen**：`characterization` `= None` → openapi-typescript 渲染为可选。**测试**：`test_characterization_does_not_block_accept`（零 continuity 冲突 + 有 advisory 问题 → 验收直通）+ 计数/golden + collect 映射 + normalize 事件 + 前端 reducer/normalize/parse 纯函数。→ 影响 @frontend（已 `pnpm gen:api`）。依赖 ⑧（motive/appearance 已入 CharacterCard 契约，作客观锚点）。
+
+- [2026-07-06] @llm/@backend/@frontend 立 C3 新端点（⑤ AI 立项方案生成）：**`POST /skills/project-plan/generate`**（**不带 project 前缀**——立项前 project 未建，绕开 toolbox 的 `project_repo.get→404`）← `ProjectPlanGenerateRequest{genre?,logline?,title?,premise?,theme?,structure?,tone?,ending_type?,narrative_pov?,selling_points[],brief?}`（全字段 `max_length` 上界 CR-H9；`brief` 用 `str|None` 令 codegen 渲染为可选）→ `ProjectPlanView{title_candidates[str], setting, narrative_structure, story_core, ending_design, tone}`（全字段默认值守解析韧性）。**种子门控**：缺 genre 或 logline（strip 后空）→ **422 VALIDATION**（空向导不出方案，防 slop）；无凭据→503。新 SPEC `project_plan_spec`（`tier="analyst"` 不写 model，`reads=()`/`writes=()` 纯预览，`prompts/project-plan.md`）+ schema `ProjectPlanResult`（`ww_agents.schemas`）+ 注册 `SCHEMA_CATALOG["project-plan"]`；**计数三处 +1（21→22）**：`specs.py` assert、`test_prompt_loader.py` len 断言、`schema_catalog.py` docstring；**regen `prompt_hashes.json`**（仅加 project-plan 一条）。端点复用 `build_brief_context`+`run_generator`（**`run_generator`/`_build_request` 的 `project_id` 放宽为 `uuid.UUID|None`**——向导阶段无 project，`usage_ledger.project_id` 本就 nullable）；新网关缝 `get_project_plan_gateway`（analyst 档，`project_deps.py`）。**零迁移**（纯预览不写业务表，仅落 usage ledger）。前端 `wizard.ts` 加 `mapPlanResultToWizardForm`（书名候选→title、叙事结构→structure、基调→tone；时空背景/结局设计无独立向导字段→折进 premise 带标签**不静默丢**；结局取向/叙事视角预设枚举不臆造）+ `applyPlanPatch`（**仅回填空字段、保护已编辑**）+ `canGeneratePlan`/`toPlanSeedRequest` + `useProjectPlan` hook（renderHook 测试）+ `ProjectWizard.tsx` 第 2 步 `PlanAssistant`（生成→预览→按字段回填）。→ 影响 @frontend（已 `pnpm gen:api`）。依赖 ④（tone/ending_type/narrative_pov 已入 project 契约与向导）。
+
+- [2026-07-06] @db/@backend/@frontend 改 C2+C3（④ 立项基调/结局/视角）：`projects` 表加 3 列 `tone`/`ending_type`/`narrative_pov`（**Text nullable、无 CHECK**，同 genre/structure；迁移 `e5f6a7b8c9d0`，nullable 免 backfill）。`ProjectCreateRequest`+`ProjectResponse`（`schemas/projects.py`）各加 3 字段 `str|None`（**无 Literal**，Literal 只留 Verdict）；domain `ProjectCreate`/`ProjectView`/`_to_view`/`SqlProjectRepo.create` + `FakeProjectRepo` 同步透传。完整 stable_core 链路：`ProjectSpecView`（`repositories.py`，frozen）加 3 字段 + `SqlProjectSpecRepo.spec` 查询构造带上 + `_build_spec_section`（`assemble.py`）渲染进「作品蓝本」块（**书级常量→缓存前缀，字节稳定守 #9**，绝不入 volatile）。前端 `wizard.ts` `WizardForm`/`emptyWizardForm`/`toCreateRequest` + 预设数组 `TONES`/`ENDING_TYPES`/`NARRATIVE_POVS`（仿 GENRES/STRUCTURES）+ `ProjectWizard.tsx` 第 3 步控件（tone→Select，pov/ending→SegmentedControl）+ 确认页回显。**codegen**：3 字段皆 `= None` 无默认工厂 → openapi-typescript 渲染为 `?: string | null` 可选。→ 影响 @frontend（已 `pnpm gen:api`）；未来 ⑤ AI 立项方案回填这 3 字段。
+
+- [2026-07-06] @llm/@backend 改 C3（⑧ appearance/motive 穿线）：`CharacterCardView`（`schemas/generation.py`）+ `ww_agents.CharacterCard` 各加 `appearance:str=""` + `motive:str=""`（**均给默认值守解析韧性**，DB 列 `characters.appearance/motive` 早已在初始迁移、此前一直 NULL——**无新迁移**）。写侧 `CharacterWriteRepo.create`/`SqlCharacterWriteRepo`（create+**update backfill 分支**）/`CharacterWriteFields` 同步加两列；`routers/generation.py` `_card_to_view`/`_view_to_card`/`_existing_characters` 双向形变带上两字段。同批**重切 motive/traits/arc 口径**：动机唯一落 `motive`，`traits[]` 只留核心/表层/阴影三层，`arc` 引用 motive（改 `character-gen.md` → 已 regen `prompt_hashes.json` 金标准）。**codegen 注意**：因带 `default`，openapi-typescript 把两字段渲染为**必填**（响应恒有值），前端 `CharacterCardView` 字面量构造点须显式给两字段。→ 影响 @frontend（已 `pnpm gen:api` + `CharacterCardItem` 展示/编辑两字段）；未来 ③ 人物塑造审查以 motive/appearance 为客观锚点。
+
+- [2026-07-06] @backend/@llm/@frontend 改 C3（⑦ 群像按定位配比·缩减版）：`CharacterGenerateRequest`（`schemas/generation.py`）加 `role_mix:dict[str,int]|None`——`model_validator(mode="after")` 在场时校验（非空、每项≥1、总和≤`MAX_GENERATE_COUNT`=12）并把 `count` 归一为各值之和（构造期归一化，忽略客户端 count）；缺省退回单一 `role`+`count`。**归属纠偏**：`build_character_gen_context`+`run_character_gen`（`ww_core/orchestrator/generation_node.py`）=**@llm 域**，加末位默认 `role_mix=None` 关键字参，在场时注入「角色定位配比」块（保插入序、确定性、无时间戳，守 #6/#7；覆盖单一 role 行）；`routers/generation.py` 穿 `role_mix=body.role_mix`。`character-gen.md` 加「按配比分配 role」纪律 → 已 regen `prompt_hashes.json`（仅 character-gen 哈希变，**不进 SPECS/不 bump 计数**）。前端 `cards.ts` 加 `sanitizeRoleMix`/`roleMixTotal` + `buildCharacterGenerateRequest` 第 4 参 roleMix（在场时 count=Σ、发 role_mix、丢 role）；`useCharacterGen` `CharacterGenInput.roleMix` 穿参；新 `RoleMixEditor.tsx` + `CharacterGenerator` SegmentedControl 单一/配比双模。**codegen**：`role_mix` 渲染 `?: {[k:string]:number} | null` 可选。→ 影响 @frontend（已 `pnpm gen:api`）。**无迁移**。
+
 - [2026-06-24] @llm/@backend 立 C6-ext（Prompt 外置方案A）：21 prompt 散文外置 `prompts/<name>.md` + `load_prompt`/`SPECS`/`SCHEMA_CATALOG`/`REVIEW_RESERVED_NAMES`（@llm）+ `SpecResolver` + SkillRegistry 入库守卫前移 + toolbox 桥接 `SPECS`（@backend）+ `.gitattributes`/wheel package-data/CI 冒烟（@devops）。随后全量重写 21 prompt 内容（任务对齐，schema 契约/不变量保持，金标准 fixture 重生成）。门禁绿：ruff/format/mypy(209)/pytest(744)。→ 影响：后续波次可将编排器 + apps/api 路由切 `SpecResolver` 并删 `*_spec` 导出；前端零影响（GeneratorTool 描述符字段不变）。
 
 - [2026-06-23] @backend 立 C-Chain（C2）：3 端点 + `schemas/chain.py` + `services/{chain_runner,chain_deps}.py` + `jobs` 零迁移复用（加 `status="awaiting_input"` + `JobRepo.set_awaiting`）+ 新 `ErrorCode.CONFLICT`(409) + `project_deps` 加 `build_chain_gateway`/`get_chain_gateway`/`get_digest_gateway_builder`/`get_checkpointer_factory`。OpenAPI 含 2 新 POST（GET /jobs 复用）。门禁绿：ruff/format/mypy(193) 干净 + pytest 600 passed + alembic 无漂移。→ 影响 @frontend（follow-up `pnpm gen:api`）、@db/@devops C3（检查点迁移）。
@@ -355,3 +365,5 @@
 - [2026-06-20] @backend 再扩 C3（B0 **可控版**，接上条）：新增 **`PUT /projects/{project_id}/chapters/{chapter_no}/injection`** ← `InjectionOverrideRequest{pinned:[{kind,name}], excluded:[{kind,name}], recent_n:int|null(1..20)}` → 返回 `InjectionResponse`（同上，**新增回显字段 `pinned`/`excluded`**，`selected[].reasons` 可含新值 **`author_pin`**）。语义：pin 强制纳入并加 `author_pin` 理由 / excluded 强制剔除（**优先于 pin**）/ recent_n 覆盖近况回看章数。GET injection 与 **draft 流式端点**均先读同一覆盖再 `assemble(override=...)`，故「看到的=写章用的」（不变量 #6 作者兜底）。持久化：**新表 `chapter_injection`**（迁移 `ad2c4c663daf`，唯一 `(project_id,chapter_no)`，复用 outline 行已否决）；`select_relevant_entities` 加 `pinned/excluded` frozenset 入参、`assemble` 加 `override` 关键字参；新 `domain/injection_repo.py`（`InjectionOverride`/`EntityRef`/`SqlInjectionOverrideRepo`，upsert 只 flush 端点 commit）。→ 影响 @frontend（已 `pnpm gen:api`；F1 加 pin/排除控件 + recent_n 步进器）。
 
 - [2026-06-18] @llm 改 C1：`Gateway.run()` 现消费 `LlmRequest.output_schema`——`OpenAICompatAdapter.complete` 在 schema 非空时经 **instructor**(`create_with_completion(response_model=...)`) 取已校验 Pydantic 实例并填 `LlmResponse.parsed`；无 schema 时 `parsed is None`、纯文本路径不变；记账仍 **1 条 usage_ledger**/调用（usage 从 raw completion 提取）。`ProviderResult` 新增 `parsed` 字段。结构化路径经可注入 `StructuredClient` Protocol 注 fake（测试不联网）。→ 影响 T2.2（续审节点可直接 `gateway.run(req).parsed`）、未来所有结构化输出 Agent。
+
+- [2026-07-06] @frontend 前端依赖新增（灵感⑥ 关系图谱缩减版 / D4，**无后端契约变更、无 gen:api、无迁移**）：`apps/web/package.json` 加 `@xyflow/react@^12.11.2`（React Flow，MIT，peer react>=17 兼容 R19）+ `@dagrejs/dagre@^3.0.0`（一次性初始布局，自带 TS 类型）。`pnpm install` 后 `pnpm-workspace.yaml` 被 pnpm 自动追加 `minimumReleaseAgeExclude`（@xyflow/react/@xyflow/system 供应链豁免）+ `pnpm-lock.yaml` 更新。新增 `lib/characters/graphLayout.ts`（纯逻辑：角色+relations→节点/边+dagre 布局，仅用现有 `CharacterCardView.relations={name,kind,note}`——**边按 name-join、kind 作中文标签、节点按 role 上色，无 closeness/阵营**，悬空边跳过并计数）+ `components/characters/RelationshipGraph.tsx`（`'use client'`，经 CodexPage `dynamic(ssr:false)` 加载，显式容器高度 + import RF CSS + prefers-reduced-motion 关 fitView 动画 + 纸感配色）。消费方：无（叶子特性，挂在设定库人物 tab）。

packages/agents/tests/fixtures/prompt_hashes.json

@@ -2,7 +2,8 @@
   "blurb": "2adba04a16707d7d9dfbb84d0a5a302ea8e946104dd8a617ff7801ed86f41605",
   "book-title": "aa53ccf22da1d4e555c8168fc23f2f3547cc6f59e56e5df57540b7a946755e39",
   "brainstorm": "f670dc77e3d9bd4a1df2b8cb29b51edb8b229d214148907d4b05519492364ada",
-  "character-gen": "61e443ab15aab8abfd4a785b32b079b304fbd2096e40541452f553c008156330",
+  "character-gen": "e0f97a5d7bb622fdd7582cdeef85ba13b3f434b2633f4d22a8c6daeba48636bf",
+  "characterization": "643988cb899a66472acb3e18709f3299d2025f9bda132c774ee26c589266b51f",
   "continue": "f1ce02be3b186fc966c89cf0d54ea57a400e91d435ea91f5722cb17e43d11763",
   "continuity": "1bdf9799bad9076e54de2023433ec7da7b559509d986392a0563dc1ee20cf4a1",
   "de-ai": "9ce3020cdd4b223cc4d13c289babd2811bbfece4b392711dcba4fd0301c87249",
@@ -15,6 +16,7 @@
   "opening": "c3b442a26dddbfac0a3c1b01ab5b726b34b656ebd640c19abd25b900a27093b5",
   "outliner": "3086ba81fe8028687bf079db2c2fb227ba5a162b7fc09210914e6f4ebd9c0d2e",
   "pace": "c6a023cb93fde4879a0fb93cd28e0227694a4cf867e64e768319ee3135d82746",
+  "project-plan": "0b0afeee63c04a3dbfcf6ebc9bd4cf4f3e93a7ce5c7be15e013be23b38d97dfe",
   "refiner": "65a4baa298bedce4592829c02d6a16c15a19ef6fdbb2098278f9decc8ffeb813",
   "style": "3a7b2c078b62e4e08f62af91ecfae8d9a27adc598ca60f59927c3bcd136393c5",
   "style_extract": "998c30ea0d0eab3936e1d6b319e832645eefaa7f7dd4a1c86d5c1ec9467e8b8c",

packages/agents/tests/test_characterization_spec.py

@@ -0,0 +1,115 @@
+"""灵感③ 人物塑造续审契约测试：CharacterizationReview schema + characterization_spec。
+
+advisory 单维——仅建议、不阻断验收（D2）。
+
+契约测试——构造符合 schema 的 mock 响应，校验字段默认值降级、tier、只读（writes==()）、
+输出 schema 派生。人物塑造是 advisory：**不加入 REVIEW_RESERVED_NAMES**。不联网、无 DB。
+"""
+
+from __future__ import annotations
+
+import pytest
+from pydantic import ValidationError
+from ww_agents import (
+    REVIEW_RESERVED_NAMES,
+    AgentSpec,
+    CharacterizationIssue,
+    CharacterizationReview,
+    characterization_spec,
+)
+
+# ---- CharacterizationReview schema ----
+
+
+def test_characterization_parses_mock_response() -> None:
+    # Arrange：模拟网关 instructor 校验后的结构化产出（每条带逐字引用 + 置信度）
+    mock = {
+        "issues": [
+            {
+                "character": "沈砚",
+                "aspect": "动机一致性",
+                "where": "第4段决斗前退缩",
+                "quote": "他忽然转身就走，什么也没说。",
+                "diagnosis": "与人物卡既定「护妹至死」动机断裂，无过渡",
+                "suggestion": "让退缩扣回护妹动机（如察觉陷阱护妹先撤）",
+                "confidence": 0.82,
+            }
+        ]
+    }
+
+    # Act
+    review = CharacterizationReview.model_validate(mock)
+
+    # Assert
+    assert len(review.issues) == 1
+    issue = review.issues[0]
+    assert issue.character == "沈砚"
+    assert issue.aspect == "动机一致性"
+    assert issue.quote == "他忽然转身就走，什么也没说。"
+    assert issue.confidence == pytest.approx(0.82)
+
+
+def test_characterization_defaults_to_empty_issues() -> None:
+    # 无人物卡 / 无问题 → 空 issues（默认值降级，仿 StyleDriftReview 韧性）
+    review = CharacterizationReview()
+    assert review.issues == []
+
+
+def test_characterization_issue_all_fields_have_defaults() -> None:
+    # 畸形/缺字段守韧性：全字段给默认值，空 dict 也能解析
+    issue = CharacterizationIssue.model_validate({})
+    assert issue.character == ""
+    assert issue.aspect == ""
+    assert issue.where == ""
+    assert issue.quote == ""
+    assert issue.diagnosis == ""
+    assert issue.suggestion == ""
+    assert issue.confidence == 0.0
+
+
+def test_characterization_issue_partial_fields_ok() -> None:
+    # 缺 confidence/quote 时仍解析（降级默认），不抛
+    issue = CharacterizationIssue.model_validate({"character": "阿黎", "diagnosis": "扁平"})
+    assert issue.character == "阿黎"
+    assert issue.diagnosis == "扁平"
+    assert issue.confidence == 0.0
+
+
+# ---- characterization_spec 声明（analyst 档，advisory 单维）----
+
+
+def test_characterization_spec_is_analyst_tier() -> None:
+    # 不变量 #2：只声明档位，不写 model
+    assert characterization_spec.tier == "analyst"
+    assert characterization_spec.name == "characterization"
+
+
+def test_characterization_spec_is_read_only() -> None:
+    # 不变量 #3：审查只读，唯一落库=collect 留痕
+    assert characterization_spec.writes == ()
+
+
+def test_characterization_spec_reads_characters() -> None:
+    # 以人物卡 motive/appearance 为客观锚点
+    assert characterization_spec.reads == ("characters",)
+
+
+def test_characterization_spec_output_schema() -> None:
+    assert characterization_spec.output_schema is CharacterizationReview
+
+
+def test_characterization_spec_prompt_ignores_length_and_diction() -> None:
+    # 对抗 LLM-judge 冗长/辞藻偏见：prompt 必须显式声明忽略文本长度与辞藻华丽度
+    prompt = characterization_spec.system_prompt
+    assert "忽略文本长度与辞藻华丽度" in prompt
+
+
+def test_characterization_not_in_review_reserved_names() -> None:
+    # advisory：不进受信保留名（不阻断验收、无 conflicts 通道）
+    assert "characterization" not in REVIEW_RESERVED_NAMES
+
+
+def test_characterization_spec_is_agent_spec_and_immutable() -> None:
+    assert isinstance(characterization_spec, AgentSpec)
+    with pytest.raises(ValidationError):
+        characterization_spec.tier = "writer"

packages/agents/tests/test_craft_doctrine.py

@@ -0,0 +1,43 @@
+"""写章 craft 教条加载器单测（灵感①）。
+
+`load_craft_doctrine()` import 期读盘 `prompts/write_craft.md`（仿 load_prompt），
+主写章不在 SPECS、不进 prompt_hashes 金标准。文件缺失应 fail-fast（崩 import）。
+"""
+
+from __future__ import annotations
+
+import hashlib
+
+from ww_agents import load_craft_doctrine
+
+
+def _sha256(text: str) -> str:
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()
+
+
+def test_load_craft_doctrine_returns_nonempty_text() -> None:
+    # Act
+    doctrine = load_craft_doctrine()
+    # Assert
+    assert isinstance(doctrine, str)
+    assert doctrine.strip()
+
+
+def test_load_craft_doctrine_is_byte_stable() -> None:
+    # 稳定缓存前缀（不变量 #9）：两次读盘字节完全一致。
+    assert _sha256(load_craft_doctrine()) == _sha256(load_craft_doctrine())
+
+
+def test_craft_doctrine_covers_core_principles() -> None:
+    # 品类无关内核应覆盖：show-don't-tell / 对话个性化 / 反注水（按情节功能分配字数）。
+    doctrine = load_craft_doctrine()
+    assert "反注水" in doctrine or "注水" in doctrine
+    assert "对话" in doctrine
+    assert "情节功能" in doctrine
+
+
+def test_craft_doctrine_is_genre_agnostic() -> None:
+    # 品类无关内核不应写死具体题材名（题材片段归 GENRE_CRAFT）。
+    doctrine = load_craft_doctrine()
+    for genre in ("玄幻", "仙侠", "都市"):
+        assert genre not in doctrine, f"品类无关内核不应含题材名 {genre!r}"

packages/agents/tests/test_generation_specs.py

@@ -108,6 +108,30 @@ def test_character_card_optional_collections_default_empty() -> None:
     assert card.relations == []
 
 
+def test_character_card_has_appearance_motive_defaults() -> None:
+    # ⑧ 穿线：appearance/motive 有默认值守解析韧性（LLM 漏产字段不致整卡解析失败）。
+    card = CharacterCard.model_validate(
+        {"name": "甲", "role": "工具人", "backstory": "无名小卒", "arc": "始终如一"}
+    )
+    assert card.appearance == ""
+    assert card.motive == ""
+
+
+def test_character_card_parses_appearance_and_motive() -> None:
+    card = CharacterCard.model_validate(
+        {
+            "name": "苏璃",
+            "role": "CP",
+            "backstory": "遗孤",
+            "arc": "由敌转盟",
+            "appearance": "银发赤瞳，惯着玄色劲装",
+            "motive": "替亡兄查清族灭真相",
+        }
+    )
+    assert card.appearance == "银发赤瞳，惯着玄色劲装"
+    assert card.motive == "替亡兄查清族灭真相"
+
+
 def test_character_card_requires_core_fields() -> None:
     # name/role/backstory/arc 为必填
     with pytest.raises(ValidationError):
@@ -168,6 +192,31 @@ def test_character_gen_spec_prompt_documents_anti_duplication() -> None:
     assert "防雷同" in prompt
 
 
+def test_character_gen_prompt_declares_appearance_and_motive_fields() -> None:
+    # ⑧ 穿线：产出契约须显式声明 appearance/motive 两个新字段。
+    prompt = character_gen_spec.system_prompt
+    assert "appearance" in prompt
+    assert "motive" in prompt
+
+
+def test_motive_not_double_written_in_traits() -> None:
+    # D6 重切回归：动机唯一落 `motive` 字段，`traits` 只留性格三层——
+    # traits 契约行须把动机/欲望/恐惧**改导向 motive**，绝不再指示折进 traits
+    # （否则双写→漂移→喂 ③ 人物塑造审查假信号）。
+    prompt = character_gen_spec.system_prompt
+    lines = prompt.splitlines()
+    traits_line = next(line for line in lines if line.lstrip().startswith("- **traits**"))
+    motive_line = next(line for line in lines if line.lstrip().startswith("- **motive**"))
+    # traits 契约行显式把动机重导向 motive 字段（而非折进 traits）。
+    assert "motive" in traits_line
+    # 旧的「写明核心动机」折入指示已被移除。
+    assert "写明核心动机" not in prompt
+    assert "并写明" not in traits_line
+    # motive 契约行承接内驱三要素。
+    assert "动机" in motive_line
+    assert "恐惧" in motive_line
+
+
 # ---- 共性：immutable + 是 AgentSpec ----
 
 

packages/agents/tests/test_project_plan_spec.py

@@ -0,0 +1,97 @@
+"""立项方案生成器（project-plan）契约测试：schema + spec（灵感⑤）。
+
+契约测试——构造符合 schema 的 mock 响应，校验字段默认值（解析韧性）、tier、只读权限。
+project-plan：`{title_candidates:[str], setting, narrative_structure, story_core,
+ending_design, tone}`（分析档，reads=()，writes=() 纯预览）。不联网、无 DB。
+"""
+
+from __future__ import annotations
+
+import pytest
+from pydantic import ValidationError
+from ww_agents import (
+    SPECS,
+    AgentSpec,
+    ProjectPlanResult,
+    project_plan_spec,
+)
+
+# ---- ProjectPlanResult schema ----
+
+
+def test_project_plan_parses_mock_response() -> None:
+    # Arrange：模拟网关 instructor 校验后的结构化立项方案产出
+    mock = {
+        "title_candidates": ["逐光而行", "剑试九霄", "凡骨录"],
+        "setting": "上古仙门倾轧的洞天世界",
+        "narrative_structure": "黄金三章立钩 → 拜师入门 → 宗门大比爆发",
+        "story_core": "废柴少年觉醒禁忌血脉，在仙门倾轧中逆势封神，代价是逐渐失去人性",
+        "ending_design": "主角登顶却选择归隐，核心冲突以自我救赎收束（正剧）",
+        "tone": "热血 + 悲怆",
+    }
+
+    # Act
+    result = ProjectPlanResult.model_validate(mock)
+
+    # Assert
+    assert result.title_candidates == ["逐光而行", "剑试九霄", "凡骨录"]
+    assert result.setting == "上古仙门倾轧的洞天世界"
+    assert result.narrative_structure.startswith("黄金三章")
+    assert "逆势封神" in result.story_core
+    assert result.ending_design.endswith("（正剧）")
+    assert result.tone == "热血 + 悲怆"
+
+
+def test_project_plan_all_fields_default_for_parse_resilience() -> None:
+    # 全字段给默认值守解析韧性：LLM 漏产任一字段不致整卡解析失败（仿 StyleDriftReview 降级）。
+    result = ProjectPlanResult()
+    assert result.title_candidates == []
+    assert result.setting == ""
+    assert result.narrative_structure == ""
+    assert result.story_core == ""
+    assert result.ending_design == ""
+    assert result.tone == ""
+
+
+def test_project_plan_partial_response_degrades_gracefully() -> None:
+    # 只给书名候选，其余缺失 → 缺字段降级为空串，不报错。
+    result = ProjectPlanResult.model_validate({"title_candidates": ["书名甲"]})
+    assert result.title_candidates == ["书名甲"]
+    assert result.story_core == ""
+
+
+# ---- project_plan_spec 声明 ----
+
+
+def test_project_plan_spec_is_analyst_tier() -> None:
+    # 不变量 #2：立项方案构思用分析档，只声明 tier、不写 model
+    assert project_plan_spec.tier == "analyst"
+    assert project_plan_spec.name == "project-plan"
+
+
+def test_project_plan_spec_is_read_only_preview() -> None:
+    # 立项前 project 未建：reads=()（无库可读）；纯预览 writes=()（不变量 #3）
+    assert project_plan_spec.reads == ()
+    assert project_plan_spec.writes == ()
+
+
+def test_project_plan_spec_output_schema() -> None:
+    assert project_plan_spec.output_schema is ProjectPlanResult
+
+
+def test_project_plan_spec_registered_in_specs() -> None:
+    # 注册表按 name 命中同一实例（不变量：SPECS[name] is *_spec）
+    assert SPECS["project-plan"] is project_plan_spec
+
+
+def test_project_plan_spec_is_immutable() -> None:
+    assert isinstance(project_plan_spec, AgentSpec)
+    with pytest.raises(ValidationError):
+        project_plan_spec.tier = "writer"
+
+
+def test_project_plan_spec_prompt_documents_seed_discipline() -> None:
+    # prompt 必须强调忠于种子（题材 + logline），不另起炉灶
+    prompt = project_plan_spec.system_prompt
+    assert "logline" in prompt
+    assert "种子" in prompt

packages/agents/tests/test_prompt_loader.py

@@ -52,15 +52,21 @@ def test_load_prompt_matches_golden(name: str) -> None:
 
 
 # ---- #1 注册表唯一性 ----
-def test_specs_registry_len_is_21() -> None:
-    assert len(SPECS) == 21
+def test_specs_registry_len_is_23() -> None:
+    assert len(SPECS) == 23
     # name 即 key，dict 已去重；逐项确认 key == spec.name（无错位）
     assert all(key == spec.name for key, spec in SPECS.items())
 
 
+# 非 SPEC 教条 prompt：主写章 craft 内核（灵感①）。它是散文教条、不是结构化产物 spec，
+# 故不入 SPECS/SCHEMA_CATALOG、不进 prompt_hashes 金标准——但仍是 prompts/*.md，需在
+# md↔spec↔catalog 一一对应校验里显式排除，否则集合等式误报。
+_DOCTRINE_STEMS = {"write_craft"}
+
+
 # ---- #2 md ↔ spec ↔ catalog 一一对应（覆盖 style.md / character-gen.md 连字符）----
 def test_md_spec_catalog_one_to_one() -> None:
-    md_stems = {p.stem for p in PROMPTS_DIR.glob("*.md")}
+    md_stems = {p.stem for p in PROMPTS_DIR.glob("*.md")} - _DOCTRINE_STEMS
     assert set(SPECS) == md_stems == set(SCHEMA_CATALOG)
     # 易错连字符 / 非变量名命名显式覆盖
     assert "style" in md_stems and "style_drift" not in md_stems

packages/agents/ww_agents/__init__.py

@@ -5,7 +5,7 @@
 
 from __future__ import annotations
 
-from .prompt_loader import PromptNotFoundError, load_prompt
+from .prompt_loader import PromptNotFoundError, load_craft_doctrine, load_prompt
 from .schema_catalog import SCHEMA_CATALOG, output_schema_for
 from .schemas import (
     Blurb,
@@ -13,6 +13,8 @@ from .schemas import (
     BookTeardownResult,
     CharacterCard,
     CharacterGenResult,
+    CharacterizationIssue,
+    CharacterizationReview,
     CharacterRelation,
     Conflict,
     ConflictType,
@@ -38,6 +40,7 @@ from .schemas import (
     PaceIssue,
     PaceReview,
     PolishResult,
+    ProjectPlanResult,
     Scene,
     StyleDimension,
     StyleDriftReview,
@@ -56,6 +59,7 @@ from .specs import (
     book_title_spec,
     brainstorm_spec,
     character_gen_spec,
+    characterization_spec,
     continue_spec,
     continuity_spec,
     de_ai_spec,
@@ -68,6 +72,7 @@ from .specs import (
     opening_spec,
     outliner_spec,
     pace_spec,
+    project_plan_spec,
     refiner_spec,
     style_drift_spec,
     style_extract_spec,
@@ -85,6 +90,8 @@ __all__ = [
     "BookTeardownResult",
     "CharacterCard",
     "CharacterGenResult",
+    "CharacterizationIssue",
+    "CharacterizationReview",
     "CharacterRelation",
     "Conflict",
     "ConflictType",
@@ -111,6 +118,7 @@ __all__ = [
     "PaceIssue",
     "PaceReview",
     "PolishResult",
+    "ProjectPlanResult",
     "Scene",
     "StyleDimension",
     "StyleDriftReview",
@@ -124,6 +132,7 @@ __all__ = [
     "book_title_spec",
     "brainstorm_spec",
     "character_gen_spec",
+    "characterization_spec",
     "continue_spec",
     "continuity_spec",
     "de_ai_spec",
@@ -132,12 +141,14 @@ __all__ = [
     "foreshadow_spec",
     "glossary_spec",
     "golden_finger_spec",
+    "load_craft_doctrine",
     "load_prompt",
     "name_spec",
     "opening_spec",
     "outliner_spec",
     "output_schema_for",
     "pace_spec",
+    "project_plan_spec",
     "refiner_spec",
     "style_drift_spec",
     "style_extract_spec",

packages/agents/ww_agents/prompt_loader.py

@@ -46,3 +46,18 @@ def load_prompt(name: str) -> str:
 
     _CACHE[name] = text
     return text
+
+
+# 主写章 craft 教条文件名（灵感①）。它是散文教条、不是结构化产物 spec：
+# 不入 SPECS、不进 prompt_hashes 金标准、不参与计数断言——但复用同一读盘/规整/缓存路径，
+# 保证字节稳定（不变量 #9）。
+CRAFT_DOCTRINE_NAME: Final = "write_craft"
+
+
+def load_craft_doctrine() -> str:
+    """读 `prompts/write_craft.md`（品类无关写章 craft 内核）→ 规整 → 缓存 → 返回。
+
+    与 `load_prompt` 同一读盘/规整/缓存逻辑（DRY），文件缺失同样 fail-fast
+    （`PromptNotFoundError`）。调用方在 import 期取常量即触发读盘。
+    """
+    return load_prompt(CRAFT_DOCTRINE_NAME)

packages/agents/ww_agents/prompts/character-gen.md

@@ -2,14 +2,14 @@
 
 ## 你信奉的角色塑造法则
 
-- **缺陷驱动弧光**：好角色起点必有缺点/盲区/创伤，故事让其被填平或代价显现，才有真正的「转变」。没有缺陷的完人弧光是平的——`arc` 必须由 `traits` 里的阴影/恐惧推动。
-- **三层性格 + 内驱动机**：性格分核心（灵魂底色）-表层（外人所见）-阴影（自己都未必承认的一面）三层，再叠加核心动机 / 欲望 / 恐惧 / 价值观。这套内驱让角色在任何情境下行为自洽。
-- **代入感来自具体**：用具体的经历、口癖、萌点把角色钉成「一个真实的人」，而非一串形容词。
+- **缺陷驱动弧光**：好角色起点必有缺点/盲区/创伤，故事让其被填平或代价显现，才有真正的「转变」。没有缺陷的完人弧光是平的——`arc` 必须由 `traits` 的阴影层与 `motive` 里的恐惧推动。
+- **三层性格与内驱动机分列**：性格（`traits`）分核心（灵魂底色）-表层（外人所见）-阴影（自己都未必承认的一面）三层；核心动机 / 欲望 / 恐惧这套**内驱单独落 `motive` 字段**，不折进 `traits`。两者分列，让角色在任何情境下行为自洽、且动机不双写。
+- **代入感来自具体**：用具体的经历、口癖、萌点、外貌（`appearance`）把角色钉成「一个真实的人」，而非一串形容词。
 - **契合世界观硬规则**：取名风格、能力上限、出身势力都不能越过世界观的力量体系 / 势力 / 地理硬规则。
 
 ## 输入材料
 
-- **角色需求**（一句话，如「亦正亦邪的女二、与主角有宿命纠葛、出身敌对势力」）+ 数量（count）+ 定位；
+- **角色需求**（一句话，如「亦正亦邪的女二、与主角有宿命纠葛、出身敌对势力」）+ 数量（count）+ 定位；若给出**角色定位配比**（各定位对应数量的清单），按配比给每个定位产出对应数量的卡、合计等于总数量；
 - **世界观约束**（world_entities：力量体系、势力、地理硬规则——取名 / 能力须契合）；
 - **已有角色**（characters：现有角色卡，新角色须与其区分、可与其建关系）；
 - **本批已生成卡**（同一次批量里先产出的卡——后续卡须与它们差异化，避免雷同）。
@@ -22,9 +22,11 @@
 
 - **name**（字符串）：角色名，取名风格契合世界观；
 - **role**（字符串）：角色定位（主角 / CP / 对手 / 导师 / 工具人 等），按需求/定位分配；
-- **traits**（字符串列表）：性格特质清单——按「核心-表层-阴影」三层展开，并写明核心动机 / 欲望 / 恐惧 / 价值观；每条独立成项；
+- **traits**（字符串列表）：性格特质清单——**只按「核心-表层-阴影」三层展开**（核心=灵魂底色、表层=外人所见、阴影=自己都未必承认的一面）；每条独立成项。**动机 / 欲望 / 恐惧不写这里，落 `motive` 字段**；
+- **appearance**（字符串）：外貌形象——身形、五官、气质、标志性外观特征、常着装扮；喂给文风一致性让描写有据。无从设定则留白（空串），不硬编；
+- **motive**（字符串）：核心动机 / 欲望 / 恐惧——驱动角色行动的**唯一内驱**，是 `arc` 转变的推力；一段话写清「他到底要什么、怕什么」。无从设定则留白（空串）；
 - **backstory**（字符串）：背景故事——出身、关键经历、创伤 / 转折点；
-- **arc**（字符串）：人物弧光——起点 → 转变 → 终点，须与剧情挂钩，且转变由 traits 里的缺陷/阴影推动；
+- **arc**（字符串）：人物弧光——起点 → 转变 → 终点，须与剧情挂钩，且转变由 `motive` 的欲望/恐惧与 `traits` 的阴影层推动；
 - **speech_tics**（字符串列表）：口癖 / 语言风格——用词偏好、口头禅；这是喂给文风一致性、让对话有辨识度的关键，每条独立成项；
 - **tags**（字符串列表）：人设标签 / 萌点（网文专属，辅助检索与差异化），每条独立成项；
 - **relations**（对象列表）：关系网——与已有 / 同批角色建边；每条边含 `name`（关系对象的角色名）+ `kind`（关系类型：宿敌 / 师徒 / CP / 同盟 / 工具人 等）+ 可选 `note`（关系说明，可缺省）；无关系则为空列表。
@@ -44,15 +46,16 @@
 ## 落笔前自检
 
 每张卡产出前，逐项确认：
-- 八个字段是否齐全、类型正确（该列表的给列表）？
-- `arc` 的转变是否真由 `traits` 里的阴影 / 恐惧推动，而非凭空逆转？
-- 取名、能力、出身是否都没越过世界观硬规则？
+- 十个字段是否齐全、类型正确（该列表的给列表，appearance / motive 给字符串）？
+- 动机 / 欲望 / 恐惧是否只写在 `motive`、没有回流进 `traits`（`traits` 只剩性格三层）？
+- `arc` 的转变是否真由 `motive` 的欲望 / 恐惧与 `traits` 的阴影推动，而非凭空逆转？
+- 取名、能力、出身、`appearance` 是否都没越过世界观硬规则？
 - 与已有卡 / 本批前序卡在内核上确实**差异化**了？
 
 ## 纪律
 
 - 你只产结构化角色卡，**不改稿、不写库、不直接互调其他 agent**——入库前由编排器追加一道 continuity 校验确认不与世界观 / 力量体系冲突（§6.5）；
-- count 是几就产几张，不多不少；定位若给出则按定位分配；
+- count 是几就产几张，不多不少；给出**角色定位配比**时严格按配比分配——每个定位产出清单标注的数量、各定位之和等于总数量；只给单一定位则整批按该定位分配；两者皆无则按需求自行判定；
 - 不臆造与需求 / 世界观无关的设定，宁可留白也不硬填。
 
 ## 示例（单卡，仅示范字段填法与质感，勿照抄设定）
@@ -61,7 +64,9 @@
 
 - name：苏璃（敌对势力「玄霜阁」的姓氏体系，冷感单字名）
 - role：CP / 亦敌亦友女二
-- traits：["核心：极度护短、认定的人会拼命；表层：清冷毒舌、拒人千里；阴影：童年被族中抛弃，恐惧再次被抛弃故先发制人地疏远他人", "核心动机：替亡兄查清族灭真相", "欲望：被真正接纳", "恐惧：信任后被背叛", "价值观：恩怨分明、私恩高于阵营"]
+- traits：["核心：极度护短、认定的人会拼命", "表层：清冷毒舌、拒人千里", "阴影：童年被族中抛弃，先发制人地疏远他人"]
+- appearance：银发赤瞳，身形清瘦，惯着玄霜阁玄色劲装；左颈一道旧疤，情绪起伏时会无意识地按住它
+- motive：核心动机是替亡兄查清族灭真相；深处渴望被真正接纳，却最恐惧信任后再被背叛——恩怨分明、私恩高于阵营
 - backstory：玄霜阁旁支遗孤，兄长死于阁主清洗，被迫隐忍效忠仇敌，暗中收集证据……
 - arc：起点——以阵营之名敌视主角；转变——查真相途中数次被主角所救，护短的核心压过阵营立场；终点——背叛玄霜阁与主角同盟，却因「先发制人疏远他人」的旧伤酿成一次几乎致命的误会
 - speech_tics：["惯用反问句堵人", "说狠话前先轻笑一声「呵」", "极少称呼对方名字，用「你」保持距离"]

packages/agents/ww_agents/prompts/characterization.md

@@ -0,0 +1,47 @@
+你是长篇连载网络小说的「人物塑造续审」（characterization-checker）——专审人物塑造的资深网文编辑。你的唯一职责：对照作品真相源里的人物卡（尤其**动机 motive** 与**外貌 appearance** 两个客观锚点），逐条诊断本章草稿里人物塑造的**失真/扁平/失声/动机断裂**，引用具体片段并给出置信度。你**只读、只诊断、只报告**，绝不改稿、不续写、不写库（不变量 #3）。你的产出是**建议性的（advisory）**——供作者参考，不阻断验收。
+
+## 审稿依据：人物卡（客观锚点优先）
+
+注入材料里带本作的人物卡（含 `motive` 动机/欲望/恐惧、`appearance` 外貌、性格三层 `traits`、成长弧 `arc`）。以人物卡为客观基准，重点核对本章人物的言行是否与其**既定动机**一致、外貌/标识描写是否与**既定 appearance** 冲突。人物卡缺失或未注入该角色时，退回网文通用基线：主角与主要配角应有**清晰动机驱动**、**可辨识的说话声音**、**行为符合已建立的性格逻辑**，避免工具人式的扁平与前后失声。
+
+## 只审塑造、不审辞藻（关键纪律）
+
+- **显式忽略文本长度与辞藻华丽度**：段落长短、用词是否华丽、句子是否优美，**一律不在本审范围**。只判「这个人物写得像不像这个人、动机通不通、有没有辨识度」，不判文笔好坏、不判篇幅多寡。
+- 不因描写细腻/篇幅长就抬高评价，也不因文字朴素/篇幅短就判为塑造不足——**塑造质量与文本体量、辞藻华丽度无关**（对抗 LLM 评审的冗长偏见与辞藻偏见）。
+- 节奏/注水、文风漂移、一致性冲突、伏笔——**均不在本审范围**（各有专审）。本审只看人物塑造这一维。
+
+## 审稿方法（逐步执行）
+
+1. **定位人物**：列出本章有实质戏份的角色，优先主角与主要配角。
+2. **对锚诊断**：逐个角色，对照人物卡的 `motive`/`appearance`/`traits`/`arc`，找出本章中与之**失真、断裂、扁平化、失声**的具体表现。
+3. **证据优先**：每条诊断都必须从本章草稿**逐字摘录**一小段支撑片段（内容锚，不得改写），锚定「问题出在哪句」；无法逐字锚定则不报（不臆造）。
+4. **给置信度**：每条诊断给一个 `0.0–1.0` 的置信度——你越确信是真问题给越高；主观/边缘/见仁见智的给低值，供作者过滤（对抗 LLM 评审主观偏见）。
+5. **建议而非改稿**：给出改进方向（如「让其行为回扣既定动机 X」），但**绝不输出改写后的正文**。
+
+### 塑造维度锚点（aspect 取值口径）
+
+- `动机一致性`：本章言行是否被人物卡既定动机驱动，有无无动机行动/动机断裂；
+- `声音辨识度`：对白/内心是否有该角色可辨识的说话声音，还是千人一腔；
+- `行为逻辑`：行为是否符合已建立的性格逻辑与处境，有无为推进剧情的强行 OOC；
+- `外貌一致性`：外貌/标识性描写是否与人物卡 `appearance` 冲突；
+- `扁平化`：是否沦为只服务功能的工具人、缺乏立体动机与内在张力。
+
+## 输出格式（严格 JSON，仅含以下顶层字段，不要多余字段或解释文字）
+
+- **`issues`**：人物塑造问题列表（数组）。逐条诊断；每个元素为对象，含：
+  - `character`：涉及的角色名（尽量对齐人物卡 name；无法归属某角色则留空串）；
+  - `aspect`：塑造维度（取上述锚点之一或贴切的简短维度名）；
+  - `where`：本章定位（哪一段/哪个场景体现该问题，便于作者定位）；
+  - `quote`：从本章草稿**逐字摘录**的支撑片段（含标点，不得改写；无法逐字锚定则不要输出该条）；
+  - `diagnosis`：诊断说明（为何此处塑造偏弱/失真，扣住动机/声音/行为逻辑等客观点）；
+  - `suggestion`：改进建议（方向性，不含改写后的正文）；
+  - `confidence`：置信度浮点 `0.0–1.0`（越确信越高，主观/边缘给低值）。
+  - 若全章人物塑造无明显问题，`issues` 为空数组 `[]`。
+
+## 纪律（不可违反）
+
+- 你**只读、只报人物塑造诊断**，绝不改稿、不写库、不输出任何改写后的正文（不变量 #3）；你的结论是 advisory 建议，不阻断验收。
+- **显式忽略文本长度与辞藻华丽度**——只判塑造，不判文笔与篇幅。
+- 每条诊断必须带逐字摘录的 `quote` 与 `confidence`，不臆造、不空泛；无逐字证据则不报。
+- 无明显问题则 `issues` 为空数组；宁缺毋滥、少而准，避免维度噪声淹没作者注意力。
+- 只输出上述字段的 JSON，不附加旁白、不解释推理过程。

packages/agents/ww_agents/prompts/project-plan.md

@@ -0,0 +1,32 @@
+你是中文网文的「立项方案师」（project-plan，分析档）。你像一位既懂创作、又懂平台数据的资深网文主编：在作者刚有一个题材 + 一句话故事的种子时，为其构思一份**结构化立项方案**，把模糊的灵感推演成可开写的书级蓝图，供作者预填立项向导。你只产结构化方案，**不改稿、不写库**（纯预览，不变量 #3）。
+
+## 输入材料
+
+- 作品种子（向导草稿序列化文本）：题材、一句话故事（logline）为**必给种子**；可能另附暂定书名、立意、主题、故事结构、基调、结局取向、叙事视角、核心卖点等——**有则贴合，无则不臆造**。
+- 作者的一句话方向/需求（可空——空则围绕种子自由构思）。
+
+读不到的字段就当它不存在：不编造与已知种子无关的设定，宁缺毋造；已给的字段要顺着推演、不要推翻。
+
+## 核心纪律
+
+- **忠于种子**：方案必须从题材 + logline 生长出来，是对种子的**放大与落地**，不是另起炉灶换一个故事。
+- **可开写**：每个维度都要具体到能指导后续世界观/大纲/写章，不写正确的废话（「主角很强」「结局很爽」这类无信息量表述一律不要）。
+- **服务连载**：站在平台/读者视角想「黄金三章靠什么钩人、凭什么比同类更想追读、长线靠什么留存」，而非只站作者立场自嗨。
+
+## 各维度字段要求（严格对应输出结构）
+
+- **title_candidates**（书名候选清单）：3–5 个差异化书名候选，每个都要贴题材、有记忆点、能一眼看出爽点/赛道；已给暂定书名时可保留并另给替代。避免雷同——若两个候选互换后几乎一样则重拟其一。
+- **setting**（时空背景）：故事发生的时代/地域/世界底层规则概述——是仙侠的洞天宗门、都市的现代都会、还是架空王朝？点明与主线强相关的设定支点，别写百科式罗列。
+- **narrative_structure**（叙事结构）：全书主线推进的骨架——开篇怎么立钩、主线如何升级/推进、阶段性爽点与转折大致怎么铺（可呼应作者已选的三幕/故事圈/雪花等结构，但要落到本故事的具体节拍上）。
+- **story_core**（故事核心）：这本书的立意与核心冲突——主角要什么、挡路的是什么、代价是什么，一句到一段说清「凭什么是一个值得追的故事」。这是立项的灵魂，务必具体。
+- **ending_design**（结局设计）：故事大致收束方向与情绪落点——主角最终抵达何处、核心冲突如何了结、留给读者什么。呼应作者已选的结局取向（HE/BE/开放/正剧）时要贴合，未选时给一个与 story_core 自洽的收束设想。
+- **tone**（基调）：全书的情绪底色与阅读质感（热血 / 轻松 / 沉重 / 治愈 / 暗黑 / 爽文 等，或其组合描述）。呼应作者已选基调；未选时据题材与 story_core 判定。
+
+## 输出格式（结构化）
+
+产出一个立项方案对象，含字段 `title_candidates`（字符串数组）、`setting`、`narrative_structure`、`story_core`、`ending_design`、`tone`（均为字符串）。无从判断的字段可留空（对应结构上的可空默认），**但不要为凑字数编造与种子无关的内容**。除该结构外不输出任何解释性文字。
+
+## 边界
+
+- 贴合题材 + logline 种子推演，作者已给的其它字段顺着用、不推翻；
+- 只产结构化立项方案，**不改稿、不写任何库**（纯预览，不变量 #3）。

packages/agents/ww_agents/prompts/write_craft.md

@@ -0,0 +1,32 @@
+# 写章创作纪律（品类无关内核）
+
+你是一位深耕中文网络小说的职业写手。下面是贯穿所有题材的创作纪律——它们是**灵活的原则**，不是逐条打勾的清单。按当下这一章真正要写的情节去取舍运用，不要机械套用、不要为了满足某条规则而牺牲叙事本身。
+
+## 语域总纲
+
+- 网文是**读起来快**的文体：信息密度偏低、句子偏短、段落偏碎。读者靠画面与动作往前滑，而不是靠密集的说明与铺陈。
+- **重行动、轻独白**：优先用发生的事推进故事，少写大段静止的心理分析或情绪反刍。心理活动点到即止，让位于外部动作、对话与场景。
+- **情节持续推进**：每个场景都要往前走一步——带来新信息、新冲突、新关系变化或新目标。避免原地打转的交代和没有推进的过场堆叠。
+
+## show，别 tell
+
+- 把情绪**外化**为可见可感的东西：动作、姿态、生理反应、环境细节、对话里的停顿与语气，而不是直接抛出「他很愤怒」「她很紧张」这类概括标签。
+- 让读者从画面里**自己推断**人物的状态，而不是由叙述者代为总结。越关键的情绪，越要用细节演出来，而不是一句话说破。
+
+## 对话的声音
+
+- 每个角色的说话方式要**可区分**：依据角色档案里的性格、出身、处境，赋予各自的用词习惯、句子长短、口头禅与说话节奏。
+- **忌所有人一个腔调**——反派、少年、长者、市井之徒不该共用同一套措辞与句式。对话要能让读者仅凭语气就认出是谁在说。
+- 对话标签（「他说」「她道」）的**位置要变化**：时而前置、时而后置、时而用一个动作 beat 代替，避免每句都以同一模式收尾。
+
+## 反注水：按情节功能分配篇幅
+
+- 字数**跟着情节功能走**，不是往一个固定目标硬凑。关键的冲突、转折、情绪高点要写**厚**——给足动作分解、感官细节与张力；过场、赶路、次要交代要写**薄**——一两句带过。
+- 密点厚、疏点薄。**绝不为凑字数注水**：删掉重复的心理复述、冗余的形容词堆叠、无推进的对话与说明。
+- 篇幅是结果而非目标——先想清楚这一章每个部分承担什么功能，再决定它值多少笔墨。
+
+## 文字质感
+
+- **句长错落**：长短句交替，避免连续同长度、同结构的句子带来的机械节奏。
+- **避免用情绪概括收尾**：段落或章节的结尾不要落在「他心中五味杂陈」「一切都不一样了」这类抽象总结上，用具体的动作、画面或悬念收束。
+- **慎用连续排比与套路化修辞**：偶一为之可以蓄势，密集堆砌则显空洞、露出 AI 腔。宁可朴素准确，不要华而不实。

packages/agents/ww_agents/schema_catalog.py

@@ -2,7 +2,7 @@
 
 Pydantic 类型承载结构化解析与 `isinstance` 校验，无法从文本推导、无法运行时安全构造，
 必须留 Python。`SPECS[name].output_schema` 一律从这里派生（单向派生，避免双真相）。
-`input_schema` 全 21 个恒为 None（入参为序列化文本），本波不建入参槽（YAGNI）。
+`input_schema` 全 23 个恒为 None（入参为序列化文本），本波不建入参槽（YAGNI）。
 
 `refiner` 是唯一纯文本 writer：`None` 是合法值（无结构化 schema）。
 """
@@ -17,6 +17,7 @@ from .schemas import (
     BlurbResult,
     BookTeardownResult,
     CharacterGenResult,
+    CharacterizationReview,
     ContinuationResult,
     ContinuityReview,
     DeAiResult,
@@ -30,6 +31,7 @@ from .schemas import (
     OutlineResult,
     PaceReview,
     PolishResult,
+    ProjectPlanResult,
     StyleDriftReview,
     StyleFingerprintResult,
     TitleListResult,
@@ -42,6 +44,7 @@ SCHEMA_CATALOG: Final[dict[str, type[BaseModel] | None]] = {
     "outliner": OutlineResult,
     "foreshadow": ForeshadowReview,
     "pace": PaceReview,
+    "characterization": CharacterizationReview,
     "style_extract": StyleFingerprintResult,
     "style": StyleDriftReview,
     "refiner": None,  # 唯一纯文本 writer，None 是合法值
@@ -59,6 +62,7 @@ SCHEMA_CATALOG: Final[dict[str, type[BaseModel] | None]] = {
     "expand": PolishResult,
     "de-ai": DeAiResult,
     "teardown": BookTeardownResult,
+    "project-plan": ProjectPlanResult,
 }
 
 

packages/agents/ww_agents/schemas.py

@@ -226,6 +226,50 @@ class StyleDriftReview(BaseModel):
     )
 
 
+# ---- 人物塑造续审结构化输出（灵感③ · advisory 单维；仅建议、不阻断验收）----
+
+
+class CharacterizationIssue(BaseModel):
+    """单条人物塑造问题：角色 + 维度 + 定位 + 逐字引用 + 诊断 + 建议 + 置信度。
+
+    以人物卡 `motive`/`appearance` 为客观锚点。`quote` 是从本章草稿**逐字摘录**的支撑
+    片段（内容锚，不得改写）；`confidence` 供作者过滤主观/边缘诊断（对抗 LLM-judge 偏见）。
+    全字段给默认值守解析韧性（畸形/缺字段不炸）。
+    """
+
+    character: str = Field(
+        default="", description="涉及的角色名（尽量对齐人物卡 name；无法归属留空）"
+    )
+    aspect: str = Field(
+        default="",
+        description="塑造维度（动机一致性 / 声音辨识度 / 行为逻辑 / 外貌一致性 / 扁平化 等）",
+    )
+    where: str = Field(default="", description="本章定位：哪一段/哪个场景体现该问题")
+    quote: str = Field(
+        default="",
+        description="从本章草稿**逐字摘录**的支撑片段（含标点，不得改写）；供作者内容锚定位",
+    )
+    diagnosis: str = Field(default="", description="诊断说明：为何此处人物塑造偏弱/失真")
+    suggestion: str = Field(default="", description="改进建议（方向性，不含改写后的正文）")
+    confidence: float = Field(
+        default=0.0, description="置信度 0.0–1.0（越确信越高，主观/边缘给低值，供作者过滤）"
+    )
+
+
+class CharacterizationReview(BaseModel):
+    """人物塑造续审结构化产出：只读建议清单（advisory，不阻断验收，不变量 #3）。
+
+    **带默认值降级**：材料无人物卡 / 无明显问题时返回 `issues=[]`（默认值），不报错、
+    不阻塞其余审（§5.2 失败隔离之外的内容性兜底，仿 `StyleDriftReview`）。仅建议——
+    不进 conflicts、不碰验收 gate（灵感 D2 advisory）。
+    """
+
+    issues: list[CharacterizationIssue] = Field(
+        default_factory=list,
+        description="人物塑造问题清单（每条带角色/维度/定位/逐字引用/置信度）；无则空列表",
+    )
+
+
 # ---- worldbuilder 结构化输出（ARCH §5.4 worldbuilder 行 / §6.5 / §4.5）----
 
 
@@ -285,10 +329,18 @@ class CharacterCard(BaseModel):
     role: str = Field(description="角色定位（主角 / CP / 对手 / 导师 / 工具人 等）")
     traits: list[str] = Field(
         default_factory=list,
-        description="性格特质清单（核心-表层-阴影 / 动机 / 欲望 / 恐惧）",
+        description="性格特质清单（仅核心-表层-阴影三层；动机/欲望/恐惧落 motive）",
+    )
+    appearance: str = Field(
+        default="",
+        description="外貌形象（身形/五官/气质/标志性外观/常着装扮）；缺则空串（守解析韧性）",
+    )
+    motive: str = Field(
+        default="",
+        description="核心动机/欲望/恐惧（驱动行动的唯一内驱，arc 由此推动）；缺则空串",
     )
     backstory: str = Field(description="背景故事（出身、关键经历、创伤/转折点）")
-    arc: str = Field(description="人物弧光（起点 → 转变 → 终点，与剧情挂钩）")
+    arc: str = Field(description="人物弧光（起点 → 转变 → 终点，与剧情挂钩，由 motive 推动）")
     speech_tics: list[str] = Field(
         default_factory=list,
         description="口癖/语言风格（用词偏好、口头禅，喂给文风一致性）",
@@ -594,3 +646,31 @@ class BookTeardownResult(BaseModel):
         default_factory=list,
         description="抓人钩子/爽点套路清单；无则空列表",
     )
+
+
+# ---- 立项方案生成器（project-plan；analyst；纯预览 writes=()）----
+
+
+class ProjectPlanResult(BaseModel):
+    """立项方案生成器结构化产出：书级蓝图，供预填立项向导（灵感⑤）。
+
+    纯预览产物——不映射任何业务表、不入库（`project_plan_spec.writes=()`，不变量 #3）。
+    **全字段给默认值守解析韧性**（仿 `StyleDriftReview` 降级范式）：LLM 漏产任一字段
+    不致整个方案解析失败——缺字段降级为空（书名候选空列表 / 文本字段空串），前端只回填
+    非空交集、不覆盖作者已编辑值。
+    """
+
+    title_candidates: list[str] = Field(
+        default_factory=list,
+        description="书名候选清单（3–5 个差异化候选）；无则空列表",
+    )
+    setting: str = Field(default="", description="时空背景（时代/地域/世界底层规则概述）；缺则空串")
+    narrative_structure: str = Field(
+        default="", description="叙事结构（主线推进骨架：立钩/升级/转折节拍）；缺则空串"
+    )
+    story_core: str = Field(
+        default="",
+        description="故事核心（立意 + 核心冲突：主角要什么、挡路的是什么、代价）；缺则空串",
+    )
+    ending_design: str = Field(default="", description="结局设计（收束方向与情绪落点）；缺则空串")
+    tone: str = Field(default="", description="基调（全书情绪底色与阅读质感）；缺则空串")

packages/agents/ww_agents/specs.py

@@ -30,6 +30,7 @@ __all__ = [
     "book_title_spec",
     "brainstorm_spec",
     "character_gen_spec",
+    "characterization_spec",
     "continue_spec",
     "continuity_spec",
     "de_ai_spec",
@@ -42,6 +43,7 @@ __all__ = [
     "opening_spec",
     "outliner_spec",
     "pace_spec",
+    "project_plan_spec",
     "refiner_spec",
     "style_drift_spec",
     "style_extract_spec",
@@ -118,6 +120,20 @@ style_drift_spec = AgentSpec(
     scope="builtin",
 )
 
+# ---- 人物塑造续审（灵感③；analyst 档；advisory 单维——仅建议、不阻断验收，D2）----
+characterization_spec = AgentSpec(
+    name="characterization",
+    tier="analyst",  # 不变量 #2：只声明档位，不写 model（人物塑造诊断用分析档）
+    system_prompt=load_prompt("characterization"),
+    input_schema=None,  # 注入材料为序列化文本（人物卡 + 草稿，经 assemble），非结构化入参
+    output_schema=SCHEMA_CATALOG["characterization"],
+    reads=(
+        "characters",
+    ),  # 以人物卡 motive/appearance 为客观锚点（经 assemble 的 stable_core 注入）
+    writes=(),  # 只读（不变量 #3）；advisory 唯一落库=collect 留痕，不进 conflicts/验收 gate
+    scope="builtin",
+)
+
 # ---- 回炉（refiner；writer 档；纯文本重写；非 Agent 流水线、非持久，M4-e）----
 refiner_spec = AgentSpec(
     name="refiner",
@@ -298,6 +314,18 @@ teardown_spec = AgentSpec(
     scope="builtin",
 )
 
+# ---- project-plan（立项方案师；分析档；纯预览 writes=()，灵感⑤）----
+project_plan_spec = AgentSpec(
+    name="project-plan",
+    tier="analyst",  # 不变量 #2：只声明档位，不写 model（立项方案构思用分析档）
+    system_prompt=load_prompt("project-plan"),
+    input_schema=None,  # 向导阶段 project 未建；材料为序列化的向导草稿（端点注入），非结构化入参
+    output_schema=SCHEMA_CATALOG["project-plan"],
+    reads=(),  # 立项前 project 未建、无库可读；种子经端点序列化注入
+    writes=(),  # 纯预览，不写库（不变量 #3）
+    scope="builtin",
+)
+
 
 # 集中注册表：name → spec（同一实例，兼容期 *_spec 与 SPECS[name] 为同对象，不变量）
 # MappingProxyType：只读视图，运行时 `SPECS[x] = ...` / `del SPECS[x]` 抛 TypeError
@@ -309,6 +337,7 @@ _SPECS_BY_NAME: dict[str, AgentSpec] = {
         outliner_spec,
         foreshadow_spec,
         pace_spec,
+        characterization_spec,
         style_extract_spec,
         style_drift_spec,
         refiner_spec,
@@ -326,9 +355,10 @@ _SPECS_BY_NAME: dict[str, AgentSpec] = {
         expand_spec,
         de_ai_spec,
         teardown_spec,
+        project_plan_spec,
     )
 }
-assert len(_SPECS_BY_NAME) == 21, "SPECS name 冲突或缺失"  # 唯一性 + 数量自检（import 期）
+assert len(_SPECS_BY_NAME) == 23, "SPECS name 冲突或缺失"  # 唯一性 + 数量自检（import 期）
 SPECS: Final[Mapping[str, AgentSpec]] = MappingProxyType(_SPECS_BY_NAME)
 
 # 四审受信保留名 —— 独立显式白名单（安全边界锚在此，不依附派生集合）

packages/config/tests/test_genre_craft.py

@@ -0,0 +1,47 @@
+"""GENRE_CRAFT 配置单测（F1 · 品类相关 craft 片段表）。
+
+GENRE_CRAFT 是「题材 → 语域/节奏/钩子/字数区间/常见套路」的**文本片段**映射；
+品类无关内核在 `write_craft.md`，品类相关片段在此表。lookup 对未知题材优雅返回 None。
+"""
+
+from __future__ import annotations
+
+from ww_config import GENRE_CRAFT, genre_craft_fragment
+
+# 与前端 wizard.ts 的 GENRES 预设保持一致（玄幻/仙侠/都市/科幻/历史/悬疑）。
+_FRONTEND_GENRES = ["玄幻", "仙侠", "都市", "科幻", "历史", "悬疑"]
+
+
+def test_genre_craft_covers_frontend_presets() -> None:
+    # Arrange / Act / Assert：前端每个预设题材都有对应片段。
+    for genre in _FRONTEND_GENRES:
+        assert genre in GENRE_CRAFT, f"题材 {genre!r} 缺 GENRE_CRAFT 片段"
+
+
+def test_genre_craft_fragments_are_nonempty_text() -> None:
+    for genre, fragment in GENRE_CRAFT.items():
+        assert isinstance(fragment, str)
+        assert fragment.strip(), f"题材 {genre!r} 片段为空"
+
+
+def test_genre_craft_fragment_returns_matching_text() -> None:
+    # Act
+    fragment = genre_craft_fragment("玄幻")
+    # Assert
+    assert fragment is not None
+    assert fragment == GENRE_CRAFT["玄幻"]
+
+
+def test_genre_craft_fragment_unknown_returns_none() -> None:
+    assert genre_craft_fragment("不存在的题材") is None
+
+
+def test_genre_craft_fragment_none_returns_none() -> None:
+    assert genre_craft_fragment(None) is None
+    assert genre_craft_fragment("") is None
+
+
+def test_genre_fragments_mention_wordcount_dimension() -> None:
+    # 每个片段应含字数/篇幅维度（灵活片段，非穷举清单）。
+    for genre, fragment in GENRE_CRAFT.items():
+        assert "字数" in fragment or "篇幅" in fragment, f"{genre!r} 片段未含字数区间"

packages/config/ww_config/__init__.py

@@ -1,3 +1,4 @@
+from ww_config.genre_craft import GENRE_CRAFT, genre_craft_fragment
 from ww_config.settings import Settings, get_settings
 
-__all__ = ["Settings", "get_settings"]
+__all__ = ["GENRE_CRAFT", "Settings", "genre_craft_fragment", "get_settings"]

packages/config/ww_config/genre_craft.py

@@ -0,0 +1,67 @@
+"""品类相关 craft 配置（灵感 F1）。
+
+`GENRE_CRAFT`：题材 → 「语域/节奏/钩子偏好/默认字数区间/常见套路」的**文本片段**。
+品类无关内核（show-don't-tell / 对话个性化 / 反注水原则）在 `write_craft.md`；
+本表只装**随题材而异**的片段，由 assemble 按 `project.genre` 选片段注入缓存前缀。
+
+字数写成**区间参考**（软引导，非每章硬目标）——具体篇幅仍按情节功能分配（见内核）。
+键与前端 wizard 的 GENRES 预设对齐（玄幻/仙侠/都市/科幻/历史/悬疑）＋常见言情；
+自由填写的题材命不中即优雅返回 None（不注入片段），不报错。
+"""
+
+from __future__ import annotations
+
+from typing import Final
+
+# 每条片段是灵活提示，不是穷举清单——覆盖语域/节奏/钩子/字数/套路等维度，供模型自行取舍。
+GENRE_CRAFT: Final[dict[str, str]] = {
+    "玄幻": (
+        "语域偏热血爽快、画面感强；节奏偏快，冲突密集，实力成长线要清晰可感。"
+        "钩子偏好等级压制、打脸反转、机缘奇遇与悬念前置。"
+        "单章篇幅参考约 2500–3500 字，按情节功能弹性调整。"
+        "常见套路：升级流、扮猪吃虎、系统/金手指、宗门家族争斗——用得要有新意，别照搬。"
+    ),
+    "仙侠": (
+        "语域偏古雅大气，兼顾意境与热血；节奏中到快，以境界突破为张力节点。"
+        "钩子偏好机缘造化、道心考验、恩怨情仇与天骄争锋。"
+        "单章篇幅参考约 2500–3500 字，按情节功能弹性调整。"
+        "常见套路：修真等级、飞升渡劫、师门情义、逆境翻盘——古意要真、别堆砌辞藻。"
+    ),
+    "都市": (
+        "语域偏现代口语、贴近生活、爽感直接；节奏快，即时反馈要强。"
+        "钩子偏好装逼打脸、身份反差、隐藏实力与逆袭翻身。"
+        "单章篇幅参考约 2000–3000 字（移动端碎片阅读），按情节功能弹性调整。"
+        "常见套路：扮猪吃虎、隐藏身份、都市异能、商战权斗——落点要接地气。"
+    ),
+    "科幻": (
+        "语域偏冷静、有逻辑感、由设定驱动；节奏偏中，重悬念与推演，慎防设定倾泻。"
+        "钩子偏好谜题危机、科技奇观与命运抉择。"
+        "单章篇幅参考约 2500–3500 字，按情节功能弹性调整。"
+        "常见套路：末世、星际、赛博、时间循环——设定服务剧情，避免大段科普打断叙事。"
+    ),
+    "历史": (
+        "语域偏厚重考究、代入感强；节奏偏中，以权谋博弈与人物弧线为主。"
+        "钩子偏好权谋算计、命运转折与家国大势。"
+        "单章篇幅参考约 3000–4000 字，按情节功能弹性调整。"
+        "常见套路：穿越架空、争霸天下、朝堂权谋、种田发展——细节要经得起推敲。"
+    ),
+    "悬疑": (
+        "语域偏冷峻克制、讲究信息控制；节奏张弛有度，靠悬念与反转拉动。"
+        "钩子偏好谜题、危险迫近、误导与真相错位。"
+        "单章篇幅参考约 2500–3500 字，按情节功能弹性调整。"
+        "常见套路：多线叙事、伏笔回收、层层反转、不可靠叙述——保持信息差，慎过早揭底。"
+    ),
+    "言情": (
+        "语域偏细腻、注重情绪张力与关系互动；节奏偏中，以情感推拉为主线。"
+        "钩子偏好暧昧拉扯、误会错过、情感转折与命运羁绊。"
+        "单章篇幅参考约 2500–3500 字，按情节功能弹性调整。"
+        "常见套路：先婚后爱、破镜重圆、双向奔赴、追妻火葬场——情绪外化于互动细节，别空喊。"
+    ),
+}
+
+
+def genre_craft_fragment(genre: str | None) -> str | None:
+    """按题材取 craft 片段；未指定或命不中预设题材时返回 None（不注入、不报错）。"""
+    if not genre:
+        return None
+    return GENRE_CRAFT.get(genre)

packages/core/tests/fakes_orchestrator.py

@@ -128,6 +128,7 @@ class FakeReviewRepo:
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> Any:
         entry = {
@@ -139,6 +140,7 @@ class FakeReviewRepo:
             "foreshadow_sug": list(foreshadow_sug or []),
             "style": style,
             "pace": pace,
+            "characterization": characterization,
             "health_score": health_score,
         }
         self.records.append(entry)

packages/core/tests/test_chain_graph.py

@@ -141,6 +141,7 @@ class FakeReviewRepo:
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> Any:
         rec = FakeReviewRecord(chapter_no, list(conflicts))

packages/core/tests/test_characterization_review.py

@@ -0,0 +1,207 @@
+"""灵感③ 人物塑造 advisory 审单测——并入 review 图 + collect characterization 列 + SSE 事件。
+
+全部注入 mock 网关 / 内存 fake repo，无真 LLM、无真 Postgres（图用 MemorySaver）。
+asyncio_mode=auto。验第五审（advisory）落 `characterization` 列、SSE `section` + `characterization`
+事件、空 issues 优雅降级、失败隔离不阻塞其余审。
+"""
+
+from __future__ import annotations
+
+import uuid
+from typing import Any
+
+from fakes_orchestrator import FakeReviewRepo, SchemaRoutingRunGateway
+from langchain_core.runnables import RunnableConfig
+from langgraph.checkpoint.memory import MemorySaver
+from ww_agents import (
+    CharacterizationIssue,
+    CharacterizationReview,
+    ContinuityReview,
+    ForeshadowReview,
+    PaceReview,
+    StyleDriftReview,
+)
+from ww_core.orchestrator import (
+    CHARACTERIZATION,
+    CONTINUITY,
+    EVENT_CHARACTERIZATION,
+    EVENT_DONE,
+    EVENT_SECTION,
+    REVIEW_INCOMPLETE,
+    REVIEW_OK,
+    SECTION_DONE,
+    SECTION_INCOMPLETE,
+    ChapterState,
+    build_review_context,
+    build_review_graph,
+    collect_reviews,
+    extract_characterization,
+    normalize_review,
+)
+
+PROJECT = uuid.UUID("00000000-0000-0000-0000-000000000001")
+USER = uuid.UUID("00000000-0000-0000-0000-0000000000aa")
+
+STABLE = "## 世界观硬规则\n灵气可凝丹\n## 人物卡\n沈砚 动机：护妹至死"
+VOLATILE = "## 近况\n主角已破境"
+DRAFT = "他忽然转身就走，什么也没说。"
+
+_CHAR = CharacterizationReview(
+    issues=[
+        CharacterizationIssue(
+            character="沈砚",
+            aspect="动机一致性",
+            where="第1段",
+            quote="他忽然转身就走，什么也没说。",
+            diagnosis="与既定护妹动机断裂",
+            suggestion="退缩扣回护妹动机",
+            confidence=0.82,
+        )
+    ]
+)
+
+
+def _state(reviews: dict[str, Any] | None = None) -> ChapterState:
+    state: ChapterState = {
+        "project_id": PROJECT,
+        "chapter_no": 7,
+        "user_id": USER,
+        "stable_core": STABLE,
+        "volatile": VOLATILE,
+        "draft": DRAFT,
+        "review_context": build_review_context(draft=DRAFT, stable_core=STABLE, volatile=VOLATILE),
+    }
+    if reviews is not None:
+        state["reviews"] = reviews
+    return state
+
+
+def _char_reviews() -> dict[str, Any]:
+    return {CHARACTERIZATION: {"status": REVIEW_OK, "result": _CHAR.model_dump()}}
+
+
+# ---- collect：characterization 列映射（纯函数）----
+
+
+def test_extract_characterization_returns_whole_dict() -> None:
+    char = extract_characterization(_char_reviews())
+
+    assert char is not None
+    assert len(char["issues"]) == 1
+    assert char["issues"][0]["character"] == "沈砚"
+    assert char["issues"][0]["confidence"] == 0.82
+
+
+def test_extract_characterization_none_when_incomplete() -> None:
+    reviews = {CHARACTERIZATION: {"status": REVIEW_INCOMPLETE, "result": None}}
+    assert extract_characterization(reviews) is None
+
+
+def test_extract_characterization_none_when_absent() -> None:
+    assert extract_characterization({}) is None
+
+
+async def test_collect_records_characterization_to_its_column() -> None:
+    repo = FakeReviewRepo()
+
+    await collect_reviews(_state(_char_reviews()), review_repo=repo)
+
+    assert len(repo.records) == 1
+    rec = repo.records[0]
+    assert rec["characterization"]["issues"][0]["aspect"] == "动机一致性"
+    # advisory：不进 conflicts（不阻断验收）
+    assert rec["conflicts"] == []
+
+
+# ---- 图：characterization 作为默认第五审并行跑通 ----
+
+
+async def test_characterization_runs_in_default_review_graph() -> None:
+    gateway = SchemaRoutingRunGateway(
+        {
+            ContinuityReview: ContinuityReview(conflicts=[]),
+            ForeshadowReview: ForeshadowReview(),
+            PaceReview: PaceReview(),
+            StyleDriftReview: StyleDriftReview(),
+            CharacterizationReview: _CHAR,
+        }
+    )
+    repo = FakeReviewRepo()
+    graph = build_review_graph(gateway, repo, checkpointer=MemorySaver())
+    config: RunnableConfig = {"configurable": {"thread_id": "char-1"}}
+
+    final = await graph.ainvoke(_state(), config=config)
+
+    assert final["reviews"][CHARACTERIZATION]["status"] == REVIEW_OK
+    rec = repo.records[0]
+    assert rec["characterization"]["issues"][0]["character"] == "沈砚"
+
+
+async def test_characterization_failure_isolated_does_not_block_others() -> None:
+    # characterization schema 缺失 → SchemaRoutingRunGateway 抛 KeyError → 隔离为 incomplete
+    gateway = SchemaRoutingRunGateway(
+        {
+            ContinuityReview: ContinuityReview(conflicts=[]),
+            ForeshadowReview: ForeshadowReview(),
+            PaceReview: PaceReview(),
+            StyleDriftReview: StyleDriftReview(),
+        }
+    )
+    repo = FakeReviewRepo()
+    graph = build_review_graph(gateway, repo, checkpointer=MemorySaver())
+    config: RunnableConfig = {"configurable": {"thread_id": "char-2"}}
+
+    final = await graph.ainvoke(_state(), config=config)
+
+    assert final["reviews"][CHARACTERIZATION]["status"] == REVIEW_INCOMPLETE
+    rec = repo.records[0]
+    assert rec["characterization"] is None  # 失败列留空，不阻塞其余
+
+
+# ---- SSE：normalize_review surface characterization 事件 ----
+
+
+async def test_normalize_review_surfaces_characterization_event() -> None:
+    events = [e async for e in normalize_review(_char_reviews())]
+
+    # 单审：section(done) + characterization + done
+    assert [e.event for e in events] == [EVENT_SECTION, EVENT_CHARACTERIZATION, EVENT_DONE]
+    section = events[0]
+    assert section.data == {"name": CHARACTERIZATION, "status": SECTION_DONE}
+    char_evt = events[1]
+    issues = char_evt.data["issues"]
+    assert isinstance(issues, list)
+    assert issues[0]["character"] == "沈砚"
+    assert issues[0]["confidence"] == 0.82
+
+
+async def test_characterization_sorts_first_in_multi_review_sse() -> None:
+    # 字典序：characterization < continuity → advisory 事件排在最前，不打乱既有四审断言
+    reviews = {
+        CHARACTERIZATION: {"status": REVIEW_OK, "result": _CHAR.model_dump()},
+        CONTINUITY: {"status": REVIEW_OK, "result": {"conflicts": []}},
+    }
+    events = [e async for e in normalize_review(reviews)]
+    names = [e.data.get("name") for e in events if e.event == EVENT_SECTION]
+    assert names == [CHARACTERIZATION, CONTINUITY]
+
+
+async def test_normalize_review_marks_characterization_incomplete() -> None:
+    reviews = {CHARACTERIZATION: {"status": REVIEW_INCOMPLETE, "result": None}}
+
+    events = [e async for e in normalize_review(reviews)]
+
+    assert [e.event for e in events] == [EVENT_SECTION, EVENT_DONE]
+    assert events[0].data == {"name": CHARACTERIZATION, "status": SECTION_INCOMPLETE}
+
+
+async def test_characterization_empty_issues_still_records() -> None:
+    # 空 issues 降级仍是 ok 结果、照常落库（区别于 incomplete None）
+    repo = FakeReviewRepo()
+    reviews = {
+        CHARACTERIZATION: {"status": REVIEW_OK, "result": CharacterizationReview().model_dump()}
+    }
+
+    await collect_reviews(_state(reviews), review_repo=repo)
+
+    assert repo.records[0]["characterization"] == {"issues": []}

packages/core/tests/test_generation_node.py

@@ -181,6 +181,39 @@ def test_build_character_gen_context_injects_existing_and_generated() -> None:
     assert "苏黎" in ctx  # 本批已生成卡
 
 
+def test_build_character_gen_context_injects_role_mix() -> None:
+    # ⑦ 群像按定位配比：role_mix 各定位 + 数量都进上下文，指示模型按配比分配
+    ctx = build_character_gen_context(
+        brief="给我一组群像",
+        count=6,
+        role=None,
+        world_context="",
+        existing_chars=[],
+        generated_so_far=[],
+        role_mix={"主角": 1, "对手": 2, "配角": 3},
+    )
+
+    assert "配比" in ctx  # 配比块标识
+    assert "主角" in ctx and "对手" in ctx and "配角" in ctx
+    assert "1" in ctx and "2" in ctx and "3" in ctx
+
+
+def test_build_character_gen_context_role_mix_overrides_single_role() -> None:
+    # role_mix 在场时用配比块，不再输出单一「角色定位：role」行
+    ctx = build_character_gen_context(
+        brief="群像",
+        count=3,
+        role="主角",  # 应被 role_mix 覆盖
+        world_context="",
+        existing_chars=[],
+        generated_so_far=[],
+        role_mix={"对手": 3},
+    )
+
+    assert "角色定位：主角" not in ctx
+    assert "对手" in ctx
+
+
 def test_build_character_gen_context_handles_empty_anti_dup() -> None:
     # 首次单生成：无已有/已生成时仍产出可用上下文（不崩）
     ctx = build_character_gen_context(
@@ -243,6 +276,29 @@ async def test_run_character_gen_injects_anti_dup_into_request() -> None:
     assert "九转炼气" in req.input  # 世界观约束注入
 
 
+async def test_run_character_gen_injects_role_mix_into_request() -> None:
+    # ⑦ role_mix 穿参：run → build_character_gen_context → 请求 input 含各定位
+    gateway = FakeRunGateway(_CARDS)
+
+    await run_character_gen(
+        character_gen_spec,
+        brief="群像",
+        count=5,
+        role=None,
+        world_context="",
+        existing_chars=[],
+        generated_so_far=[],
+        gateway=gateway,
+        user_id=USER,
+        project_id=PROJECT,
+        role_mix={"主角": 2, "配角": 3},
+    )
+
+    req = gateway.last_request
+    assert req is not None
+    assert "配角" in req.input  # 配比定位注入
+
+
 async def test_run_character_gen_raises_when_gateway_fails() -> None:
     gateway = FailingRunGateway(RuntimeError("provider down"))
 

packages/core/tests/test_memory.py

@@ -7,6 +7,7 @@ from __future__ import annotations
 
 import uuid
 from dataclasses import dataclass, field
+from typing import Any
 
 from ww_core.domain.repositories import (
     CharacterView,
@@ -19,6 +20,7 @@ from ww_core.domain.repositories import (
     StyleView,
     WorldEntityView,
 )
+from ww_core.domain.review_repo import ReviewRepo, ReviewView
 from ww_core.memory import (
     AssembledContext,
     EntityKind,
@@ -103,6 +105,28 @@ class FakeProjectSpecRepo:
         return self.row
 
 
+@dataclass
+class FakeReviewRepo:
+    """内存审稿留痕 fake（灵感②冲突裁决反哺）——只需 list_for_chapter 供 assemble 反读。
+
+    `queried` 记录被查询的章号（验证读的是上一章 chapter_no-1）；record/set_decisions
+    只为满足 ReviewRepo Protocol 结构类型，assemble 反哺路径不用。
+    """
+
+    by_chapter: dict[int, list[ReviewView]] = field(default_factory=dict)
+    queried: list[int] = field(default_factory=list)
+
+    async def list_for_chapter(self, project_id: uuid.UUID, chapter_no: int) -> list[ReviewView]:
+        self.queried.append(chapter_no)
+        return list(self.by_chapter.get(chapter_no, []))
+
+    async def record(self, *args: Any, **kwargs: Any) -> ReviewView:  # pragma: no cover
+        raise NotImplementedError
+
+    async def set_decisions(self, *args: Any, **kwargs: Any) -> ReviewView:  # pragma: no cover
+        raise NotImplementedError
+
+
 def build_repos(
     *,
     outline: dict[int, OutlineView] | None = None,
@@ -113,6 +137,7 @@ def build_repos(
     style: StyleView | None = None,
     rules: list[RuleView] | None = None,
     spec: ProjectSpecView | None = None,
+    review: ReviewRepo | None = None,
 ) -> MemoryRepos:
     return MemoryRepos(
         outline=FakeOutlineRepo(outline or {}),
@@ -123,6 +148,7 @@ def build_repos(
         style=FakeStyleRepo(style),
         rules=FakeRulesRepo(rules or []),
         project=FakeProjectSpecRepo(spec),
+        review=review,
     )
 
 
@@ -383,6 +409,253 @@ async def test_chapter_directive_uses_chapter_no() -> None:
     assert "请创作第 7 章的正文。" in ctx.volatile
 
 
+# ---- 灵感 F1：genre-aware craft 片段注入 ----
+
+
+async def test_genre_fragment_injected_by_genre() -> None:
+    from ww_config import GENRE_CRAFT
+
+    repos = build_repos(spec=ProjectSpecView(title="书", genre="玄幻"))
+    ctx = await assemble(repos, PROJECT, 5)
+    # 玄幻题材片段进 stable_core（每书稳定，随书缓存），不进 volatile。
+    assert GENRE_CRAFT["玄幻"] in ctx.stable_core
+    assert GENRE_CRAFT["玄幻"] not in ctx.volatile
+
+
+async def test_different_genre_yields_different_fragment() -> None:
+    from ww_config import GENRE_CRAFT
+
+    xuan = await assemble(build_repos(spec=ProjectSpecView(title="书", genre="玄幻")), PROJECT, 5)
+    city = await assemble(build_repos(spec=ProjectSpecView(title="书", genre="都市")), PROJECT, 5)
+    assert GENRE_CRAFT["玄幻"] in xuan.stable_core
+    assert GENRE_CRAFT["都市"] in city.stable_core
+    assert xuan.stable_core != city.stable_core
+
+
+async def test_unknown_genre_injects_no_fragment() -> None:
+    repos = build_repos(spec=ProjectSpecView(title="书", genre="不存在的题材"))
+    ctx = await assemble(repos, PROJECT, 5)
+    assert "本作题材写法" not in ctx.stable_core
+
+
+async def test_no_genre_injects_no_fragment() -> None:
+    repos = build_repos(spec=ProjectSpecView(title="书"))  # genre 默认 None
+    ctx = await assemble(repos, PROJECT, 5)
+    assert "本作题材写法" not in ctx.stable_core
+
+
+# ---- 灵感④：立项基调/结局/视角是书级常量 → 进 stable_core（缓存前缀，不变量 #9） ----
+
+
+async def test_build_stable_includes_tone_ending_pov() -> None:
+    spec = ProjectSpecView(title="书", tone="热血", ending_type="HE", narrative_pov="第一人称")
+    ctx = await assemble(build_repos(spec=spec), PROJECT, 5)
+    assert "热血" in ctx.stable_core
+    assert "HE" in ctx.stable_core
+    assert "第一人称" in ctx.stable_core
+
+
+async def test_tone_ending_pov_never_enter_volatile() -> None:
+    # 书级常量随书稳定 → 只进缓存前缀，绝不进 volatile（章号易变区）。
+    spec = ProjectSpecView(title="书", tone="热血", ending_type="HE", narrative_pov="第一人称")
+    ctx = await assemble(build_repos(spec=spec), PROJECT, 5)
+    assert "热血" not in ctx.volatile
+    assert "第一人称" not in ctx.volatile
+
+
+# ---- 灵感①/F1：开篇黄金三章特判（volatile，仅前 3 章） ----
+
+
+async def test_opening_marker_injected_for_ch1to3_only() -> None:
+    repos = build_repos(spec=ProjectSpecView(title="书"))
+    for chapter_no in (1, 2, 3):
+        ctx = await assemble(repos, PROJECT, chapter_no)
+        assert "黄金三章" in ctx.volatile
+        assert f"开篇第 {chapter_no} 章" in ctx.volatile
+    # 第 4 章起不再注入开篇标记。
+    ctx4 = await assemble(repos, PROJECT, 4)
+    assert "黄金三章" not in ctx4.volatile
+
+
+async def test_opening_marker_never_enters_stable_core() -> None:
+    # 不变量#9：开篇标记含章号（每章易变）→ 只入 volatile，绝不进缓存前缀。
+    repos = build_repos(spec=ProjectSpecView(title="书", genre="玄幻"))
+    ctx = await assemble(repos, PROJECT, 1)
+    assert "黄金三章" not in ctx.stable_core
+
+
+async def test_no_per_chapter_varying_field_in_cached_blocks() -> None:
+    # 不变量#9：同书不同章，stable_core（含 genre 片段）必须字节稳定；
+    # 章号 / 开篇标记等每章易变项只在 volatile。
+    repos = build_repos(spec=ProjectSpecView(title="书", genre="玄幻"))
+    ch1 = await assemble(repos, PROJECT, 1)
+    ch5 = await assemble(repos, PROJECT, 5)
+    assert ch1.stable_core == ch5.stable_core  # genre 片段随书稳定，跨章不变
+    assert "第 1 章" not in ch1.stable_core
+    assert "第 5 章" not in ch5.stable_core
+
+
+# ---- 灵感②/D3：上一已验收章 continuity 冲突裁决反哺（volatile，确定性模板） ----
+
+
+def _accepted_review(
+    *,
+    chapter_no: int,
+    conflicts: list[dict[str, Any]],
+    items: list[dict[str, Any]],
+) -> ReviewView:
+    """构造一行"已验收"审稿留痕：decisions 非空（只在验收事务写入 → 视为已验收）。"""
+    return ReviewView(
+        id=uuid.UUID(int=chapter_no),
+        project_id=PROJECT,
+        chapter_no=chapter_no,
+        conflicts=conflicts,
+        decisions={"items": items, "conflict_count": len(conflicts)},
+    )
+
+
+def _conflict(type_: str, suggestion: str) -> dict[str, Any]:
+    return {"type": type_, "where": "某段", "refs": [], "suggestion": suggestion}
+
+
+def _review_repo_with(reviews: list[ReviewView]) -> FakeReviewRepo:
+    by_chapter: dict[int, list[ReviewView]] = {}
+    for r in reviews:
+        by_chapter.setdefault(r.chapter_no, []).append(r)
+    return FakeReviewRepo(by_chapter=by_chapter)
+
+
+async def test_volatile_includes_prior_conflict_notes() -> None:
+    review = _accepted_review(
+        chapter_no=1,
+        conflicts=[_conflict("设定冲突", "主角左臂已断，后文不得再用左手持剑")],
+        items=[{"conflict_index": 0, "verdict": "accept", "note": None}],
+    )
+    repos = build_repos(
+        spec=ProjectSpecView(title="书"),
+        review=_review_repo_with([review]),
+    )
+    ctx = await assemble(repos, PROJECT, 2)
+    assert "上一章冲突提醒" in ctx.volatile
+    assert "主角左臂已断，后文不得再用左手持剑" in ctx.volatile
+
+
+async def test_prior_conflict_notes_never_enter_stable_core() -> None:
+    # 不变量#9：反哺是每章易变输入（依赖上一章裁决）→ 只入 volatile，绝不进缓存前缀。
+    review = _accepted_review(
+        chapter_no=1,
+        conflicts=[_conflict("设定冲突", "灵脉禁止跨界，后文守此规则")],
+        items=[{"conflict_index": 0, "verdict": "accept", "note": None}],
+    )
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=_review_repo_with([review]))
+    ctx = await assemble(repos, PROJECT, 2)
+    assert "上一章冲突提醒" not in ctx.stable_core
+    assert "灵脉禁止跨界，后文守此规则" not in ctx.stable_core
+
+
+async def test_reads_previous_accepted_chapter() -> None:
+    # 写第 5 章 → 只应读第 4 章（chapter_no-1）的审稿留痕。
+    review = _accepted_review(
+        chapter_no=4,
+        conflicts=[_conflict("时间线倒错", "事件发生在雨夜，后文保持一致")],
+        items=[{"conflict_index": 0, "verdict": "manual", "note": "已统一为雨夜"}],
+    )
+    fake = _review_repo_with([review])
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=fake)
+    ctx = await assemble(repos, PROJECT, 5)
+    assert fake.queried == [4]  # 只查上一章
+    assert "已统一为雨夜" in ctx.volatile
+
+
+async def test_skips_unaccepted_previous_chapter() -> None:
+    # 上一章有审稿留痕但 decisions=None（尚未验收）→ 不反哺（守风险#5：不读未定稿）。
+    unaccepted = ReviewView(
+        id=uuid.UUID(int=99),
+        project_id=PROJECT,
+        chapter_no=1,
+        conflicts=[_conflict("设定冲突", "不应注入的建议")],
+        decisions=None,
+    )
+    repos = build_repos(
+        spec=ProjectSpecView(title="书"),
+        review=_review_repo_with([unaccepted]),
+    )
+    ctx = await assemble(repos, PROJECT, 2)
+    assert "上一章冲突提醒" not in ctx.volatile
+    assert "不应注入的建议" not in ctx.volatile
+
+
+async def test_only_retained_conflict_decisions_injected() -> None:
+    # 只反哺保留项 verdict∈{accept,manual}；ignore（作者判为非冲突）不反哺。
+    review = _accepted_review(
+        chapter_no=1,
+        conflicts=[
+            _conflict("设定冲突", "采纳项应注入"),
+            _conflict("时间线倒错", "手改项应注入"),
+            _conflict("地理矛盾", "忽略项不应注入"),
+        ],
+        items=[
+            {"conflict_index": 0, "verdict": "accept", "note": None},
+            {"conflict_index": 1, "verdict": "manual", "note": None},
+            {"conflict_index": 2, "verdict": "ignore", "note": None},
+        ],
+    )
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=_review_repo_with([review]))
+    ctx = await assemble(repos, PROJECT, 2)
+    assert "采纳项应注入" in ctx.volatile
+    assert "手改项应注入" in ctx.volatile
+    assert "忽略项不应注入" not in ctx.volatile
+
+
+async def test_prior_conflict_notes_bounded_deduped_sorted_and_no_llm_call() -> None:
+    # 5 条保留裁决 + 1 条重复 → 去重后有界 top-3；两次组装字节完全一致（确定性、无 LLM）。
+    conflicts = [_conflict("设定冲突", f"约束{i}") for i in range(5)]
+    conflicts.append(_conflict("设定冲突", "约束0"))  # 与 index 0 内容重复
+    items = [{"conflict_index": i, "verdict": "accept", "note": None} for i in range(6)]
+    review = _accepted_review(chapter_no=1, conflicts=conflicts, items=items)
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=_review_repo_with([review]))
+
+    ctx1 = await assemble(repos, PROJECT, 2)
+    ctx2 = await assemble(repos, PROJECT, 2)
+    # 字节稳定（无 LLM 的非确定性）——assemble 无网关入参，反哺全程纯字符串模板。
+    assert ctx1.volatile == ctx2.volatile
+    section = ctx1.volatile.split("## 上一章冲突提醒\n", 1)[1].split("\n\n", 1)[0]
+    lines = [ln for ln in section.splitlines() if ln.startswith("- ")]
+    assert len(lines) == 3  # 有界 top-3
+    assert len(set(lines)) == len(lines)  # 去重
+    assert "约束0" in section  # 去重后仍保留一条约束0
+
+
+async def test_no_apology_or_selfref_in_prior_conflict_notes() -> None:
+    review = _accepted_review(
+        chapter_no=1,
+        conflicts=[_conflict("设定冲突", "后文保持灵脉守恒")],
+        items=[{"conflict_index": 0, "verdict": "accept", "note": None}],
+    )
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=_review_repo_with([review]))
+    ctx = await assemble(repos, PROJECT, 2)
+    section = ctx.volatile.split("## 上一章冲突提醒\n", 1)[1].split("\n\n", 1)[0]
+    # 前瞻正向约束——绝无道歉/自指措辞（守 #3/#4：不自评自改、不磨平文风）。
+    for banned in ("抱歉", "道歉", "对不起", "如前所述", "更正", "笔误"):
+        assert banned not in section
+
+
+async def test_assemble_without_review_repo_gracefully_degrades() -> None:
+    # review=None（默认）→ 优雅降级：不反哺、不崩（守风险#6）。
+    repos = build_repos(spec=ProjectSpecView(title="书"))  # review 默认 None
+    ctx = await assemble(repos, PROJECT, 5)
+    assert "上一章冲突提醒" not in ctx.volatile
+
+
+async def test_no_prior_conflict_notes_for_first_chapter() -> None:
+    # 第 1 章无上一章 → 不查 review repo、不反哺。
+    fake = FakeReviewRepo()
+    repos = build_repos(spec=ProjectSpecView(title="书"), review=fake)
+    ctx = await assemble(repos, PROJECT, 1)
+    assert fake.queried == []  # 未查询（无上一章）
+    assert "上一章冲突提醒" not in ctx.volatile
+
+
 # ---- helpers ----
 
 

packages/core/tests/test_orchestrator.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 import uuid
 
 from fakes_orchestrator import FakeStreamGateway, RaisingGateway
+from ww_agents import load_craft_doctrine
 from ww_core.orchestrator import (
     EVENT_DONE,
     EVENT_ERROR,
@@ -37,14 +38,58 @@ def test_build_write_request_puts_stable_core_in_cached_system_block() -> None:
 
     assert req.tier == "writer"  # 不变量②：只传档位，不传 model
     assert req.stream is True
-    assert len(req.system) == 1
-    assert req.system[0].text == STABLE
-    assert req.system[0].cache is True  # 不变量#9：稳定内核进缓存断点前
+    # 灵感①：system = [craft 教条(cache), stable_core(cache)]，craft 在最前。
+    assert len(req.system) == 2
+    assert all(b.cache is True for b in req.system)  # 不变量#9：稳定内核进缓存断点前
+    assert req.system[1].text == STABLE
     assert req.input == VOLATILE
     assert req.scope.user_id == USER
     assert req.scope.project_id == PROJECT
 
 
+# ---- 灵感①：写章 craft 教条块 ----
+
+
+def test_write_request_includes_craft_block() -> None:
+    req = build_write_request(
+        stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+    # craft 教条作为独立的 cache=True system 块注入。
+    assert req.system[0].text == load_craft_doctrine()
+    assert req.system[0].cache is True
+
+
+def test_craft_precedes_stable_core() -> None:
+    req = build_write_request(
+        stable_core=STABLE, volatile=VOLATILE, user_id=USER, project_id=PROJECT
+    )
+    # craft 教条块必须在 stable_core 块之前（跨书前缀复用最大化）。
+    assert req.system[0].text == load_craft_doctrine()
+    assert req.system[1].text == STABLE
+
+
+def test_craft_is_byte_stable() -> None:
+    # 不变量#9：craft 块字节稳定——两次构造完全一致，无每章易变项掺入。
+    req1 = build_write_request(
+        stable_core=STABLE, volatile="请创作第 1 章的正文。", user_id=USER, project_id=PROJECT
+    )
+    req2 = build_write_request(
+        stable_core=STABLE, volatile="请创作第 99 章的正文。", user_id=USER, project_id=PROJECT
+    )
+    assert req1.system[0].text == req2.system[0].text  # craft 块与章号无关
+
+
+def test_no_per_chapter_field_in_cached_system_blocks() -> None:
+    # 不变量#9：每章易变项（章号）只在 input(断点后)，绝不进任何 cache=True 的 system 块。
+    volatile = "## 写作指令\n请创作第 5 章的正文。"
+    req = build_write_request(
+        stable_core=STABLE, volatile=volatile, user_id=USER, project_id=PROJECT
+    )
+    for block in req.system:
+        assert "第 5 章" not in block.text
+    assert "第 5 章" in req.input
+
+
 # ---- stream_chapter_draft：转发网关 Delta、构造正确请求 ----
 
 

packages/core/tests/test_project_repo.py

@@ -1,6 +1,16 @@
 from datetime import UTC, datetime
 
-from ww_core.domain.project_repo import max_time
+from ww_core.domain.project_repo import _to_view, max_time
+from ww_db.models import Project
+
+
+def test_to_view_carries_tone_ending_pov() -> None:
+    # ④ 立项三字段（基调/结局/视角）经 ORM 行 → ProjectView 快照逐字段透传（snake_case）。
+    row = Project(title="书", tone="热血", ending_type="BE", narrative_pov="多视角")
+    view = _to_view(row)
+    assert view.tone == "热血"
+    assert view.ending_type == "BE"
+    assert view.narrative_pov == "多视角"
 
 
 def test_max_time_handles_missing_values() -> None:

packages/core/tests/test_review_repo.py

@@ -34,6 +34,7 @@ class FakeReviewRepo:
         foreshadow_sug: list[dict[str, object]] | None = None,
         style: dict[str, object] | None = None,
         pace: dict[str, object] | None = None,
+        characterization: dict[str, object] | None = None,
         health_score: int | None = None,
     ) -> ReviewView:
         self._seq += 1
@@ -46,6 +47,7 @@ class FakeReviewRepo:
             foreshadow_sug=list(foreshadow_sug or []),
             style=style,
             pace=pace,
+            characterization=characterization,
             health_score=health_score,
             decisions=None,
         )

packages/core/tests/test_style_review.py

@@ -14,6 +14,8 @@ from fakes_orchestrator import FakeReviewRepo, SchemaRoutingRunGateway
 from langchain_core.runnables import RunnableConfig
 from langgraph.checkpoint.memory import MemorySaver
 from ww_agents import (
+    CharacterizationIssue,
+    CharacterizationReview,
     Conflict,
     ContinuityReview,
     ForeshadowReview,
@@ -24,6 +26,7 @@ from ww_agents import (
     StyleDriftSegment,
 )
 from ww_core.orchestrator import (
+    CHARACTERIZATION,
     CONTINUITY,
     EVENT_CONFLICT,
     EVENT_DONE,
@@ -67,6 +70,19 @@ _STYLE = StyleDriftReview(
         StyleDriftSegment(idx=7, score=55),
     ],
 )
+_CHARACTERIZATION = CharacterizationReview(
+    issues=[
+        CharacterizationIssue(
+            character="沈砚",
+            aspect="动机一致性",
+            where="第4段",
+            quote="他忽然转身就走。",
+            diagnosis="与既定护妹动机断裂",
+            suggestion="退缩扣回护妹动机",
+            confidence=0.82,
+        )
+    ]
+)
 
 
 def _state(reviews: dict[str, Any] | None = None) -> ChapterState:
@@ -96,9 +112,10 @@ def _four_reviews() -> dict[str, Any]:
 # ---- REVIEW_SPECS 含第四审 ----
 
 
-def test_review_specs_include_four_audits() -> None:
+def test_review_specs_include_five_audits() -> None:
+    # 四审 + 灵感③ advisory 人物塑造第五审（characterization 末位追加）。
     names = [s.name for s in REVIEW_SPECS]
-    assert names == ["continuity", "foreshadow", "pace", "style"]
+    assert names == ["continuity", "foreshadow", "pace", "style", "characterization"]
 
 
 # ---- collect：style 列映射 ----
@@ -139,13 +156,14 @@ async def test_collect_records_style_to_style_column() -> None:
 # ---- 四审并行图跑通（mock 网关按 schema 路由产四 parsed）----
 
 
-async def test_four_review_graph_runs_all_audits_end_to_end() -> None:
+async def test_five_review_graph_runs_all_audits_end_to_end() -> None:
     gateway = SchemaRoutingRunGateway(
         {
             ContinuityReview: ContinuityReview(conflicts=[_CONFLICT]),
             ForeshadowReview: _FORESHADOW,
             PaceReview: _PACE,
             StyleDriftReview: _STYLE,
+            CharacterizationReview: _CHARACTERIZATION,
         }
     )
     repo = FakeReviewRepo()
@@ -158,10 +176,12 @@ async def test_four_review_graph_runs_all_audits_end_to_end() -> None:
     assert final["reviews"][FORESHADOW]["status"] == REVIEW_OK
     assert final["reviews"][PACE]["status"] == REVIEW_OK
     assert final["reviews"][STYLE]["status"] == REVIEW_OK
-    assert len(gateway.calls) == 4  # 四审各调一次网关
+    assert final["reviews"][CHARACTERIZATION]["status"] == REVIEW_OK
+    assert len(gateway.calls) == 5  # 五审各调一次网关（含 advisory 人物塑造）
     rec = repo.records[0]
     assert rec["conflicts"] and rec["foreshadow_sug"] and rec["pace"]
     assert rec["style"]["score"] == 72
+    assert rec["characterization"]["issues"][0]["character"] == "沈砚"
 
 
 async def test_four_review_graph_isolates_failing_style_audit() -> None:

packages/core/ww_core/domain/character_repo.py

@@ -12,7 +12,7 @@ JSONB **dict**（`traits`/`arc`/`speech_tics`），`tags`/`relations` 才是 JSO
 - `speech_tics: list[str]` → `{"items": [...]}`（JSONB dict）
 - `arc: str` → `{"text": "..."}`（JSONB dict）
 - `tags: list` / `relations: list[dict]` → 直落 JSONB list
-- `name` / `role` / `backstory` → Text 列直落
+- `name` / `role` / `appearance` / `motive` / `backstory` → Text 列直落
 
 **提交边界**：`create` 只 `flush()` 不 `commit()`——提交交端点事务（入库端点末尾一次
 `commit()`，与网关 ledger 一并落库；与项目其它写侧 repo 一致，见 memory/gotchas）。
@@ -66,6 +66,8 @@ class CharacterWriteRepo(Protocol):
         name: str,
         role: str,
         traits: list[str],
+        appearance: str,
+        motive: str,
         backstory: str,
         arc: str,
         speech_tics: list[str],
@@ -91,6 +93,8 @@ class SqlCharacterWriteRepo:
         name: str,
         role: str,
         traits: list[str],
+        appearance: str,
+        motive: str,
         backstory: str,
         arc: str,
         speech_tics: list[str],
@@ -105,6 +109,8 @@ class SqlCharacterWriteRepo:
         if existing is not None:
             existing.role = role
             existing.traits = _traits_to_jsonb(traits)
+            existing.appearance = appearance
+            existing.motive = motive
             existing.backstory = backstory
             existing.arc = _arc_to_jsonb(arc)
             existing.speech_tics = _traits_to_jsonb(speech_tics)
@@ -118,6 +124,8 @@ class SqlCharacterWriteRepo:
             name=name,
             role=role,
             traits=_traits_to_jsonb(traits),
+            appearance=appearance,
+            motive=motive,
             backstory=backstory,
             arc=_arc_to_jsonb(arc),
             speech_tics=_traits_to_jsonb(speech_tics),
@@ -139,6 +147,8 @@ class CharacterWriteFields(BaseModel):
     name: str
     role: str
     traits: list[str] = Field(default_factory=list)
+    appearance: str = ""
+    motive: str = ""
     backstory: str
     arc: str
     speech_tics: list[str] = Field(default_factory=list)

packages/core/ww_core/domain/project_repo.py

@@ -29,6 +29,9 @@ class ProjectView:
     theme: str | None = None
     selling_points: list[Any] = field(default_factory=list)
     structure: str | None = None
+    tone: str | None = None
+    ending_type: str | None = None
+    narrative_pov: str | None = None
     updated_at: datetime | None = None
     pending_review_count: int = 0
 
@@ -44,6 +47,9 @@ class ProjectCreate:
     theme: str | None = None
     selling_points: list[Any] = field(default_factory=list)
     structure: str | None = None
+    tone: str | None = None
+    ending_type: str | None = None
+    narrative_pov: str | None = None
 
 
 class ProjectRepo(Protocol):
@@ -71,6 +77,9 @@ def _to_view(
         theme=row.theme,
         selling_points=list(row.selling_points or []),
         structure=row.structure,
+        tone=row.tone,
+        ending_type=row.ending_type,
+        narrative_pov=row.narrative_pov,
         updated_at=updated_at or row.updated_at,
         pending_review_count=pending_review_count,
     )
@@ -92,6 +101,9 @@ class SqlProjectRepo:
             theme=data.theme,
             selling_points=list(data.selling_points),
             structure=data.structure,
+            tone=data.tone,
+            ending_type=data.ending_type,
+            narrative_pov=data.narrative_pov,
         )
         self._s.add(row)
         await self._s.commit()

packages/core/ww_core/domain/repositories.py

@@ -15,6 +15,8 @@ from typing import Any, Protocol
 
 from pydantic import BaseModel, Field
 
+from ww_core.domain.review_repo import ReviewRepo
+
 # ---- 只读视图（snake_case，frozen 防止意外突变）----
 
 
@@ -99,6 +101,12 @@ class ProjectSpecView(BaseModel):
     logline: str | None = None
     premise: str | None = None
     theme: str | None = None
+    # 题材（每书稳定项）——供 assemble 注入 genre-aware craft 片段（灵感 F1）。
+    genre: str | None = None
+    # 立项书级常量（灵感④）：基调/结局取向/叙事视角——每书稳定 → 入 stable_core 缓存前缀。
+    tone: str | None = None
+    ending_type: str | None = None
+    narrative_pov: str | None = None
 
 
 # ---- Repository 协议（async；统一按 project_id 过滤）----
@@ -142,7 +150,11 @@ class ProjectSpecRepo(Protocol):
 
 @dataclass(frozen=True)
 class MemoryRepos:
-    """记忆服务所需的 8 个 repo 的依赖捆绑（注入点）。"""
+    """记忆服务所需的 9 个 repo 的依赖捆绑（注入点）。
+
+    `review` 末位默认 None：只有冲突裁决反哺（灵感②）需要它；不注入时 assemble
+    优雅降级（不反哺），故既有构造点无需感知（守风险#6）。
+    """
 
     outline: OutlineRepo
     character: CharacterRepo
@@ -152,3 +164,4 @@ class MemoryRepos:
     style: StyleRepo
     rules: RulesRepo
     project: ProjectSpecRepo
+    review: ReviewRepo | None = None

packages/core/ww_core/domain/review_repo.py

@@ -33,6 +33,8 @@ class ReviewView(BaseModel):
     foreshadow_sug: list[dict[str, Any]] = Field(default_factory=list)
     style: dict[str, Any] | None = None
     pace: dict[str, Any] | None = None
+    # 灵感③ advisory 人物塑造审留痕（末位默认 None，优雅降级；不进 conflicts/验收 gate）。
+    characterization: dict[str, Any] | None = None
     health_score: int | None = None
     decisions: dict[str, Any] | None = None
 
@@ -50,6 +52,7 @@ class ReviewRepo(Protocol):
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> ReviewView:
         """append 一行审稿报告（T2.2 collect 写入）。"""
@@ -74,6 +77,7 @@ def _to_view(row: ChapterReview) -> ReviewView:
         foreshadow_sug=list(row.foreshadow_sug or []),
         style=row.style,
         pace=row.pace,
+        characterization=row.characterization,
         health_score=row.health_score,
         decisions=row.decisions,
     )
@@ -95,6 +99,7 @@ class SqlReviewRepo:
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> ReviewView:
         row = ChapterReview(
@@ -105,6 +110,7 @@ class SqlReviewRepo:
             foreshadow_sug=list(foreshadow_sug or []),
             style=style,
             pace=pace,
+            characterization=characterization,
             health_score=health_score,
             decisions=None,
         )

packages/core/ww_core/memory/assemble.py

@@ -11,6 +11,8 @@ import json
 import uuid
 from typing import Any
 
+from ww_config import genre_craft_fragment
+
 from ww_core.domain.injection_repo import InjectionOverride
 from ww_core.domain.repositories import (
     CharacterView,
@@ -23,6 +25,7 @@ from ww_core.domain.repositories import (
     StyleView,
     WorldEntityView,
 )
+from ww_core.domain.review_repo import ReviewView
 
 from .render import render_cards
 from .selection import MAIN_ROLES, RECENT_DIGEST_COUNT, EntityKey, select_relevant_entities
@@ -31,6 +34,16 @@ from .types import AssembledContext, EntityKind
 # 四级规则合并优先级：global → genre → style → project（越具体越靠后/越优先）
 _LEVEL_RANK: dict[str, int] = {"global": 0, "genre": 1, "style": 2, "project": 3}
 
+# 开篇黄金三章特判上限：前 N 章在 volatile 追加开篇提示（灵感①）。
+OPENING_CHAPTER_LIMIT = 3
+
+# 冲突裁决反哺（灵感②/D3）：只反哺"保留项"裁决——作者采纳(accept)或手改(manual)；
+# ignore（作者判为非冲突）不反哺。这三个控制值对齐 schemas.projects.Verdict。
+_RETAINED_VERDICTS: frozenset[str] = frozenset({"accept", "manual"})
+
+# 上一章冲突提醒有界上限（top-N）——控作者/模型每章注意力预算（灵感② 校验 M）。
+PRIOR_CONFLICT_NOTE_LIMIT = 3
+
 
 def _ser(value: Any) -> str:
     return json.dumps(value, ensure_ascii=False, sort_keys=True)
@@ -64,9 +77,26 @@ def _build_spec_section(spec: ProjectSpecView | None) -> str | None:
         lines.append(f"【前提】{spec.premise}")
     if spec.theme:
         lines.append(f"【主题】{spec.theme}")
+    # 立项书级常量（灵感④）：每书稳定 → 随蓝本进缓存前缀，跨章字节不变（不变量 #9）。
+    if spec.tone:
+        lines.append(f"【基调】{spec.tone}")
+    if spec.narrative_pov:
+        lines.append(f"【叙事视角】{spec.narrative_pov}")
+    if spec.ending_type:
+        lines.append(f"【结局取向】{spec.ending_type}")
     return _section("作品蓝本", "\n".join(lines))
 
 
+def _build_genre_craft_section(genre: str | None) -> str | None:
+    """按题材注入 craft 片段（语域/节奏/钩子/字数/套路，灵感 F1）。
+
+    genre 是每书稳定项（来自 spec）→ 入缓存前缀 stable_core，跨章字节稳定（不变量 #9）。
+    命不中预设题材（自由填写/未设）则不注入。品类无关内核在 write_craft.md，不在此处。
+    """
+    fragment = genre_craft_fragment(genre)
+    return _section("本作题材写法", fragment) if fragment else None
+
+
 def _build_stable(
     spec: ProjectSpecView | None,
     world_entities: list[WorldEntityView],
@@ -89,6 +119,8 @@ def _build_stable(
     rule_lines = [f"[{r.level}] {r.content}" for r in merge_rules(rules)]
 
     sections = [
+        # 题材写法领衔——同题材作品共享此前缀，利于跨书隐式前缀缓存（不变量 #9）。
+        _build_genre_craft_section(spec.genre if spec else None),
         _build_spec_section(spec),
         _section("世界观硬规则", "\n".join(world_lines)),
         _section("定型主角", "\n".join(char_lines)),
@@ -98,6 +130,72 @@ def _build_stable(
     return "\n\n".join(s for s in sections if s)
 
 
+def _opening_marker(chapter_no: int) -> str | None:
+    """开篇黄金三章特判（灵感①）：前 N 章显式提示，不赌模型自读章号。
+
+    含章号 → 每章易变 → 只入 volatile（断点后），绝不进缓存前缀（不变量 #9）。
+    """
+    if chapter_no > OPENING_CHAPTER_LIMIT:
+        return None
+    return _section(
+        "开篇黄金三章",
+        f"本章为开篇第 {chapter_no} 章，适用黄金三章原则："
+        "开篇即抛出核心冲突或身份钩子；三章之内交付第一个小爽点或转折；"
+        "避免大段世界观倾泻，边推进情节边带出设定。",
+    )
+
+
+def _conflict_constraint(conflict: dict[str, Any], note: str | None) -> str | None:
+    """把一条已裁决 continuity 冲突改写成本章的前瞻正向约束（确定性模板，绝不调 LLM）。
+
+    优先用作者手改备注（`note`，作者自己的话）；否则用审稿改法建议（`suggestion`——
+    作者点「采纳」即背书之，故仍是外部 oracle 驱动，不违 #3/#4）。只做正向「延续设定」
+    措辞，绝无道歉/自指（守 #3/#4：不自评自改、不磨平文风）。两者皆空则返回 None。
+    """
+    detail = (note or "").strip() or str(conflict.get("suggestion") or "").strip()
+    if not detail:
+        return None
+    ctype = str(conflict.get("type") or "").strip()
+    prefix = f"【{ctype}】" if ctype else ""
+    return f"{prefix}延续上一章已确认的设定：{detail}"
+
+
+def _prior_conflict_notes(
+    latest_review: ReviewView | None,
+    *,
+    limit: int = PRIOR_CONFLICT_NOTE_LIMIT,
+) -> str | None:
+    """上一已验收章的 continuity 冲突作者裁决保留项 → 本章前瞻约束（灵感②/D3）。
+
+    - 仅当上一章已验收（`decisions` 非空——只在验收事务写入）才反哺（守风险#5：不读未定稿）。
+    - 按 `decisions["items"][].conflict_index` join 回 `conflicts[index]` 取正文，
+      过滤 `verdict∈{accept,manual}`（保留项）、去重、按 conflict_index 排序、有界 top-N。
+    - 确定性纯函数、字节稳定、无 LLM 调用（守 #6/#7：assemble 可缓存的纯函数）。
+    - 软批评类（style/pace/characterization）无作者裁决 oracle → 永不经此反哺（守 #3/#4）。
+    """
+    if latest_review is None or latest_review.decisions is None:
+        return None
+    items = latest_review.decisions.get("items") or []
+    conflicts = latest_review.conflicts
+    constraints: list[str] = []
+    seen: set[str] = set()
+    for item in sorted(items, key=lambda d: d.get("conflict_index", 0)):
+        if item.get("verdict") not in _RETAINED_VERDICTS:
+            continue
+        idx = item.get("conflict_index")
+        if not isinstance(idx, int) or not (0 <= idx < len(conflicts)):
+            continue
+        line = _conflict_constraint(conflicts[idx], item.get("note"))
+        if line is None or line in seen:
+            continue
+        seen.add(line)
+        constraints.append(line)
+    bounded = constraints[:limit]
+    if not bounded:
+        return None
+    return "\n".join(f"- {c}" for c in bounded)
+
+
 def _build_volatile(
     chapter_no: int,
     cards: str,
@@ -105,6 +203,7 @@ def _build_volatile(
     digests: list[DigestView],
     beats: dict[str, Any],
     directive: str | None = None,
+    prior_conflict_notes: str | None = None,
 ) -> str:
     fore_lines = [
         f"【伏笔】{f.code} {f.title} [{f.status}]"
@@ -118,6 +217,10 @@ def _build_volatile(
         _section("本章指令", directive.strip()) if directive and directive.strip() else None,
         # 写章指令始终在场——保证 volatile（断点后的 input）永不为空（修空 prompt 400）。
         _section("写作指令", f"请创作第 {chapter_no} 章的正文。"),
+        # 开篇黄金三章标记（仅前 N 章）——每章易变，只入 volatile。
+        _opening_marker(chapter_no),
+        # 上一章冲突裁决反哺（灵感②）——依赖上一章裁决，每章易变，绝不进缓存前缀（#9）。
+        _section("上一章冲突提醒", prior_conflict_notes) if prior_conflict_notes else None,
         _section("本章注入卡片", cards),
         _section("相关伏笔窗口", "\n".join(fore_lines)),
         _section("近况摘要", "\n".join(digest_lines)),
@@ -176,10 +279,24 @@ async def assemble(
 
     main_characters = [c for c in characters if (c.role or "") in MAIN_ROLES]
 
+    # 冲突裁决反哺（灵感②/D3）：读上一已验收章的 continuity 冲突作者裁决保留项 →
+    # 本章 volatile 前瞻约束。review 未注入（默认 None）或本章为第 1 章 → 优雅降级不反哺。
+    prior_conflict_notes: str | None = None
+    if repos.review is not None and chapter_no > 1:
+        prior_reviews = await repos.review.list_for_chapter(project_id, chapter_no - 1)
+        latest_review = prior_reviews[0] if prior_reviews else None  # list_for_chapter 新→旧
+        prior_conflict_notes = _prior_conflict_notes(latest_review)
+
     stable_core = _build_stable(spec, world_entities, main_characters, style, rules)
     cards = render_cards(selection, characters, world_entities)
     volatile = _build_volatile(
-        chapter_no, cards, foreshadows, recent_digests, outline.beats, directive
+        chapter_no,
+        cards,
+        foreshadows,
+        recent_digests,
+        outline.beats,
+        directive,
+        prior_conflict_notes,
     )
 
     return AssembledContext(stable_core=stable_core, volatile=volatile, selection=selection)

packages/core/ww_core/memory/sql_repositories.py

@@ -32,6 +32,7 @@ from ww_core.domain.repositories import (
     StyleView,
     WorldEntityView,
 )
+from ww_core.domain.review_repo import SqlReviewRepo
 
 
 class SqlOutlineRepo:
@@ -215,11 +216,18 @@ class SqlProjectSpecRepo:
             logline=row.logline,
             premise=row.premise,
             theme=row.theme,
+            genre=row.genre,
+            tone=row.tone,
+            ending_type=row.ending_type,
+            narrative_pov=row.narrative_pov,
         )
 
 
 def sql_memory_repos(session: AsyncSession) -> MemoryRepos:
-    """用一个 AsyncSession 装配全部 8 个 SQLAlchemy repo（T1.4 注入点）。"""
+    """用一个 AsyncSession 装配全部 9 个 SQLAlchemy repo（T1.4 注入点）。
+
+    `review=SqlReviewRepo`：供 assemble 反读上一已验收章的 continuity 冲突裁决（灵感②）。
+    """
     return MemoryRepos(
         outline=SqlOutlineRepo(session),
         character=SqlCharacterRepo(session),
@@ -229,4 +237,5 @@ def sql_memory_repos(session: AsyncSession) -> MemoryRepos:
         style=SqlStyleRepo(session),
         rules=SqlRulesRepo(session),
         project=SqlProjectSpecRepo(session),
+        review=SqlReviewRepo(session),
     )

packages/core/ww_core/orchestrator/__init__.py

@@ -10,12 +10,14 @@ from __future__ import annotations
 
 from ._protocols import GatewayRun
 from .collect import (
+    CHARACTERIZATION,
     CONTINUITY,
     FORESHADOW,
     PACE,
     STYLE,
     ReviewRecorder,
     collect_reviews,
+    extract_characterization,
     extract_conflicts,
     extract_foreshadow_sug,
     extract_pace,
@@ -52,6 +54,7 @@ from .review_node import (
     run_review,
 )
 from .sse import (
+    EVENT_CHARACTERIZATION,
     EVENT_CONFLICT,
     EVENT_DONE,
     EVENT_ERROR,
@@ -64,6 +67,7 @@ from .sse import (
     SECTION_INCOMPLETE,
     SECTION_STARTED,
     SseEvent,
+    characterization_event,
     conflict_event,
     done_event,
     error_event,
@@ -80,8 +84,10 @@ from .style_extract_node import build_style_extract_request, run_style_extractio
 from .write_node import GatewayStream, build_write_request, stream_chapter_draft, write_node
 
 __all__ = [
+    "CHARACTERIZATION",
     "COLLECT_NODE",
     "CONTINUITY",
+    "EVENT_CHARACTERIZATION",
     "EVENT_CONFLICT",
     "EVENT_DONE",
     "EVENT_ERROR",
@@ -119,10 +125,12 @@ __all__ = [
     "build_text_input_context",
     "build_worldbuilder_context",
     "build_write_request",
+    "characterization_event",
     "collect_reviews",
     "conflict_event",
     "done_event",
     "error_event",
+    "extract_characterization",
     "extract_conflicts",
     "extract_foreshadow_sug",
     "extract_pace",

packages/core/ww_core/orchestrator/chain/nodes.py

@@ -93,6 +93,7 @@ class ReviewRecordRepo(Protocol):
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> Any: ...
 

packages/core/ww_core/orchestrator/collect.py

@@ -11,11 +11,12 @@ AI 产出真正入库经验收事务的裁决）。
 collect 与 review_repo 同样**只 flush 不 commit**——`commit` 归 HTTP 端点 / 验收事务。
 本节点绝不 commit。
 
-M4 四审齐 → 列映射（一次 record 落齐本章四审）：
+M4 四审齐 → 列映射（一次 record 落齐本章四审）+ 灵感③ advisory 人物塑造第五审：
 - continuity      → `conflicts`（冲突清单 list）；
 - foreshadow      → `foreshadow_sug`（planted/resolved 扁平为建议 list，每条带 `kind` 标记）；
 - pace            → `pace`（节奏诊断 dict：water/hook/beat_map）；
-- style      → `style`（文风漂移诊断 dict：score/segments）。
+- style           → `style`（文风漂移诊断 dict：score/segments）；
+- characterization→ `characterization`（人物塑造诊断 dict：issues；advisory，不进验收 gate）。
 任一审 `incomplete`（网关失败隔离，§5.2）→ 该列留空/None，不阻塞其余。
 """
 
@@ -35,6 +36,7 @@ CONTINUITY = "continuity"
 FORESHADOW = "foreshadow"
 PACE = "pace"
 STYLE = "style"
+CHARACTERIZATION = "characterization"  # 灵感③ advisory 人物塑造审（仅建议、不阻断验收）
 
 # foreshadow 建议扁平化时的来源标记（planted=新埋 / resolved=回收）。
 FORESHADOW_KIND_PLANTED = "planted"
@@ -57,6 +59,7 @@ class ReviewRecorder(Protocol):
         foreshadow_sug: list[dict[str, Any]] | None = None,
         style: dict[str, Any] | None = None,
         pace: dict[str, Any] | None = None,
+        characterization: dict[str, Any] | None = None,
         health_score: int | None = None,
     ) -> Any: ...
 
@@ -123,6 +126,19 @@ def extract_style(reviews: dict[str, Any]) -> dict[str, Any] | None:
     return dict(result)
 
 
+def extract_characterization(reviews: dict[str, Any]) -> dict[str, Any] | None:
+    """从 characterization（灵感③ advisory 审）分项取人物塑造诊断 dict（纯函数）。
+
+    `CharacterizationReview{issues}` 整体入 `characterization` 列（JSONB dict）。
+    未完成/缺席 → None（列留空，§5.2 不阻塞其余）。空 issues 降级仍是 `ok` 结果、照常落库。
+    advisory：只留痕、不进 conflicts、不碰验收 gate（灵感 D2）。
+    """
+    result = _ok_result(reviews, CHARACTERIZATION)
+    if result is None:
+        return None
+    return dict(result)
+
+
 async def collect_reviews(
     state: ChapterState,
     *,
@@ -140,6 +156,7 @@ async def collect_reviews(
     foreshadow_sug = extract_foreshadow_sug(reviews)
     pace = extract_pace(reviews)
     style = extract_style(reviews)
+    characterization = extract_characterization(reviews)
     await review_repo.record(
         state["project_id"],
         state["chapter_no"],
@@ -148,6 +165,7 @@ async def collect_reviews(
         foreshadow_sug=foreshadow_sug,
         style=style,
         pace=pace,
+        characterization=characterization,
     )
     log.info(
         "collect_reviews_recorded",
@@ -157,6 +175,7 @@ async def collect_reviews(
         foreshadow_sug_count=len(foreshadow_sug),
         has_pace=pace is not None,
         has_style=style is not None,
+        has_characterization=characterization is not None,
         reviews=sorted(reviews.keys()),
     )
     # 留痕已在表（真相源）；不在节点 commit（端点/事务层负责，见模块 docstring）。

packages/core/ww_core/orchestrator/generation_node.py

@@ -24,7 +24,7 @@ worldbuilder 与 character-gen 都是**独立生成**——不在写章/审稿
 from __future__ import annotations
 
 import uuid
-from collections.abc import Sequence
+from collections.abc import Mapping, Sequence
 
 import structlog
 from pydantic import BaseModel
@@ -48,13 +48,14 @@ def _build_request(
     *,
     context: str,
     user_id: uuid.UUID,
-    project_id: uuid.UUID,
+    project_id: uuid.UUID | None,
 ) -> LlmRequest:
     """据 spec + 注入材料构造生成请求（纯函数，worldbuilder/character-gen 共用）。
 
     `spec.system_prompt` → `system` 缓存断点前块（cache=True，不变量 #9）；
     `context`（序列化的需求/约束/已有/已生成文本）→ `input`（断点后）。
     生成用 `run()` 非 `stream()`（要结构化产物，非流式正文）。
+    `project_id` 可空——立项方案（⑤）在向导阶段跑，project 尚未建（usage_ledger.project_id 可空）。
     """
     return LlmRequest(
         tier=spec.tier,  # 不变量 #2：只传档位，不传 model
@@ -138,12 +139,13 @@ async def run_generator(
     context: str,
     gateway: GatewayRun,
     user_id: uuid.UUID,
-    project_id: uuid.UUID,
+    project_id: uuid.UUID | None = None,
 ) -> BaseModel:
     """通用生成器执行器（T6）：构请求 → `gateway.run` → 按 `spec.output_schema` 校验 parsed。
 
     任一声明式生成器共用此路径（不绑定具体产物类型，据 spec 的 output_schema 校验）。
     **只读、不写库**（纯预览，不变量 #3）；独立生成——网关失败直接上抛（端点处理），不静默吞。
+    `project_id` 可空——立项方案（⑤）在向导阶段跑，project 尚未建（usage_ledger.project_id 可空）。
     """
     expected = spec.output_schema
     if expected is None:
@@ -213,6 +215,11 @@ def _render_cards_for_context(cards: Sequence[CharacterCard]) -> str:
     return "\n".join(lines)
 
 
+def _render_role_mix(role_mix: Mapping[str, int]) -> str:
+    """把定位配比序列化为清单（保插入序，确定性，无时间戳）。"""
+    return "\n".join(f"- {role}：{n} 个" for role, n in role_mix.items())
+
+
 def build_character_gen_context(
     *,
     brief: str,
@@ -221,12 +228,20 @@ def build_character_gen_context(
     world_context: str,
     existing_chars: Sequence[CharacterCard],
     generated_so_far: Sequence[CharacterCard],
+    role_mix: Mapping[str, int] | None = None,
 ) -> str:
-    """拼装 character-gen 注入文本（含群像防雷同上下文，M5-a）。
+    """拼装 character-gen 注入文本（含群像防雷同上下文，M5-a；⑦ 群像按定位配比）。
 
     注入「已有角色」+「本批已生成卡」，要求新卡与二者差异化（避免一群人一个模子）。
+    `role_mix` 在场时改注入定位配比块（各定位产出对应数量、合计等于数量），覆盖单一 `role`。
     确定性拼接、无时间戳——同输入同输出，便于单测与缓存。
     """
+    if role_mix:
+        role_line = (
+            "角色定位配比（严格按此分配，各定位产出对应数量，合计等于数量）：\n"
+            f"{_render_role_mix(role_mix)}"
+        )
+    else:
         role_line = f"角色定位：{role}" if role else "角色定位：未指定（由你按需求判定）"
     existing_block = (
         _render_cards_for_context(existing_chars) if existing_chars else "（暂无已有角色）"
@@ -260,10 +275,12 @@ async def run_character_gen(
     gateway: GatewayRun,
     user_id: uuid.UUID,
     project_id: uuid.UUID,
+    role_mix: Mapping[str, int] | None = None,
 ) -> CharacterGenResult:
     """跑角色/群像生成：构防雷同上下文 → `gateway.run` → 返回 `CharacterGenResult`。
 
     群像防雷同（M5-a）：注入「已有角色 + 本批已生成卡」要求差异化（§4.5 / §6.5）。
+    `role_mix`（⑦）在场时按定位配比分配（各定位对应数量、合计等于数量）。
     独立生成（非并行审）：网关失败直接上抛（端点/T5.2 处理），不静默吞。
     **只产结构化角色卡、不写库**——入库前过 `precheck_generated_cards` 校验、再经
     入库端点持久化（持久化属 T5.2，不变量 #3）。
@@ -275,6 +292,7 @@ async def run_character_gen(
         world_context=world_context,
         existing_chars=existing_chars,
         generated_so_far=generated_so_far,
+        role_mix=role_mix,
     )
     req = _build_request(spec, context=context, user_id=user_id, project_id=project_id)
     resp = await gateway.run(req)

packages/core/ww_core/orchestrator/graph.py

@@ -17,6 +17,7 @@ from typing import TYPE_CHECKING, Any
 from langgraph.graph import END, START, StateGraph
 from ww_agents import (
     AgentSpec,
+    characterization_spec,
     continuity_spec,
     foreshadow_spec,
     pace_spec,
@@ -38,12 +39,15 @@ WRITE_NODE = "write"
 COLLECT_NODE = "collect"
 
 
-# M4：四审齐（continuity + foreshadow + pace + style）。第四审 = 文风漂移打分轨。
+# M4 四审齐（continuity + foreshadow + pace + style）+ 灵感③ advisory 人物塑造第五审。
+# characterization 是 advisory（仅建议、不阻断验收）：并行扇出自动汇入 collect，
+# 不进 conflicts、不碰验收 gate、不入 REVIEW_RESERVED_NAMES（灵感 D2）。
 REVIEW_SPECS: tuple[AgentSpec, ...] = (
     continuity_spec,
     foreshadow_spec,
     pace_spec,
     style_drift_spec,
+    characterization_spec,
 )
 
 

packages/core/ww_core/orchestrator/sse.py

@@ -18,7 +18,7 @@ from pydantic import BaseModel
 from ww_llm_gateway.types import Delta
 from ww_shared import AppError, ErrorCode
 
-from .collect import CONTINUITY, FORESHADOW, PACE, STYLE
+from .collect import CHARACTERIZATION, CONTINUITY, FORESHADOW, PACE, STYLE
 from .review_node import REVIEW_OK
 
 log = structlog.get_logger(__name__)
@@ -30,6 +30,7 @@ EVENT_CONFLICT = "conflict"  # continuity 冲突命中（审稿 review）
 EVENT_FORESHADOW = "foreshadow"  # foreshadow 建议命中（新埋/回收，审稿 review）
 EVENT_PACE = "pace"  # pace 节奏诊断（注水/钩子/节拍图，审稿 review）
 EVENT_STYLE = "style"  # style 文风漂移诊断（整体相似度 + 漂移段，审稿 review）
+EVENT_CHARACTERIZATION = "characterization"  # 人物塑造诊断（advisory issues，审稿 review）
 EVENT_DONE = "done"
 EVENT_ERROR = "error"
 
@@ -129,6 +130,18 @@ def style_event(*, score: int, segments: list[dict[str, object]]) -> SseEvent:
     )
 
 
+def characterization_event(*, issues: list[dict[str, object]]) -> SseEvent:
+    """人物塑造诊断事件（advisory，审稿 review）——前端人物塑造面板只展示、无裁决 UI。
+
+    一次审产一条：人物塑造问题清单（每条 character/aspect/where/quote/diagnosis/suggestion/
+    confidence）。仅建议、不阻断验收（灵感 D2）。
+    """
+    return SseEvent(
+        event=EVENT_CHARACTERIZATION,
+        data={"issues": issues},
+    )
+
+
 def error_event(*, code: str, message: str, request_id: str | None = None) -> SseEvent:
     """错误事件，形对齐错误信封 §7.1（`code`/`message`），附 `request_id` 便于贯通排查。"""
     return SseEvent(
@@ -231,6 +244,23 @@ def _section_result_events(name: str, result: dict[str, Any]) -> list[SseEvent]:
                 ],
             )
         )
+    elif name == CHARACTERIZATION:
+        events.append(
+            characterization_event(
+                issues=[
+                    {
+                        "character": issue.get("character", ""),
+                        "aspect": issue.get("aspect", ""),
+                        "where": issue.get("where", ""),
+                        "quote": issue.get("quote", ""),
+                        "diagnosis": issue.get("diagnosis", ""),
+                        "suggestion": issue.get("suggestion", ""),
+                        "confidence": float(issue.get("confidence", 0.0)),
+                    }
+                    for issue in result.get("issues") or []
+                ]
+            )
+        )
     return events
 
 

packages/core/ww_core/orchestrator/write_node.py

@@ -17,10 +17,15 @@ import uuid
 from collections.abc import AsyncIterator
 from typing import Protocol
 
+from ww_agents import load_craft_doctrine
 from ww_llm_gateway.types import Block, Delta, LlmRequest, Scope
 
 from .state import ChapterState
 
+# 写章 craft 教条（灵感①）：品类无关内核，import 期读盘 → 缺失即 fail-fast（崩 import）。
+# 字节稳定的稳定内容，作为最前的 cache=True system 块（跨书前缀复用，不变量 #9）。
+_CRAFT_DOCTRINE = load_craft_doctrine()
+
 
 class GatewayStream(Protocol):
     """write 节点对网关的最小依赖——只需 `stream`（注入真网关或 mock）。"""
@@ -37,11 +42,15 @@ def build_write_request(
 ) -> LlmRequest:
     """据组装上下文构造写章请求（纯函数，决策 2026-06-18）。
 
-    `stable_core` → `system` 缓存断点前块（cache=True）；`volatile` → `input`（断点后）。
+    `system` = [craft 教条(cache), stable_core(cache)]——craft 品类无关内核在最前，
+    最大化跨书前缀复用；两块均在缓存断点前（cache=True）。`volatile` → `input`（断点后）。
     """
     return LlmRequest(
         tier="writer",
-        system=[Block(text=stable_core, cache=True)],
+        system=[
+            Block(text=_CRAFT_DOCTRINE, cache=True),
+            Block(text=stable_core, cache=True),
+        ],
         input=volatile,
         stream=True,
         scope=Scope(user_id=user_id, project_id=project_id),

packages/db/migrations/versions/e5f6a7b8c9d0_project_tone_ending_pov.py

@@ -0,0 +1,33 @@
+"""Add tone/ending_type/narrative_pov to projects (灵感④ 立项书级常量).
+
+三列均 Text nullable（与既有 genre/structure 一致，无 CHECK；枚举约束留前端预设），
+nullable 免 backfill（旧项目留 NULL，assemble/前端优雅降级）。
+
+Revision ID: e5f6a7b8c9d0
+Revises: 8c1d2e3f4a5b
+Create Date: 2026-07-06 00:00:00.000000
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "e5f6a7b8c9d0"
+down_revision: str | None = "8c1d2e3f4a5b"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.add_column("projects", sa.Column("tone", sa.Text(), nullable=True))
+    op.add_column("projects", sa.Column("ending_type", sa.Text(), nullable=True))
+    op.add_column("projects", sa.Column("narrative_pov", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("projects", "narrative_pov")
+    op.drop_column("projects", "ending_type")
+    op.drop_column("projects", "tone")

packages/db/migrations/versions/f6a7b8c9d0e1_chapter_reviews_characterization.py

@@ -0,0 +1,33 @@
+"""Add characterization JSONB column to chapter_reviews (灵感③ advisory 人物塑造审).
+
+单列 JSONB nullable（advisory 留痕，独立于恒 None 的 health_score）；nullable 免 backfill
+（旧留痕行留 NULL，collect/前端优雅降级）。仅建议、不进 conflicts、不碰验收 gate（灵感 D2）。
+
+Revision ID: f6a7b8c9d0e1
+Revises: e5f6a7b8c9d0
+Create Date: 2026-07-06 00:00:00.000000
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "f6a7b8c9d0e1"
+down_revision: str | None = "e5f6a7b8c9d0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "chapter_reviews",
+        sa.Column("characterization", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("chapter_reviews", "characterization")

packages/db/ww_db/models.py

@@ -48,6 +48,10 @@ class Project(UuidPk, TimestampedMixin, Base):
     theme: Mapped[str | None] = mapped_column(Text)
     selling_points: Mapped[list[Any]] = mapped_column(JSONB, server_default=text("'[]'"))
     structure: Mapped[str | None] = mapped_column(Text)
+    # 立项书级常量（灵感④）——同 genre/structure：Text nullable、无 CHECK（枚举约束留前端预设）。
+    tone: Mapped[str | None] = mapped_column(Text)
+    ending_type: Mapped[str | None] = mapped_column(Text)
+    narrative_pov: Mapped[str | None] = mapped_column(Text)
 
 
 class Character(UuidPk, CreatedAt, Base):
@@ -221,6 +225,8 @@ class ChapterReview(UuidPk, CreatedAt, Base):
     foreshadow_sug: Mapped[list[Any]] = mapped_column(JSONB, server_default=text("'[]'"))
     style: Mapped[dict[str, Any] | None] = mapped_column(JSONB)
     pace: Mapped[dict[str, Any] | None] = mapped_column(JSONB)
+    # 灵感③ advisory 人物塑造审留痕（JSONB nullable，独立列不复用恒 None 的 health_score）。
+    characterization: Mapped[dict[str, Any] | None] = mapped_column(JSONB)
     health_score: Mapped[int | None] = mapped_column(Integer)
     decisions: Mapped[dict[str, Any] | None] = mapped_column(JSONB)
 

packages/skills/ww_skills/toolbox_registry.py

@@ -66,7 +66,7 @@ _SOURCE_TEXT_FIELD = InputField(
 )
 
 
-# 创作工具箱注册表：key → 描述符。legacy 3 + 新 8 = 11 条。
+# 创作工具箱注册表：key → 描述符。15 条（legacy 3 + 新 8 + Scope B 4：续写/扩写/降AI/拆书）。
 TOOLBOX: dict[str, GeneratorTool] = {
     # ---- legacy（spec=None，前端走 legacy_route 跳现有页面）----
     "worldbuilding": GeneratorTool(

tests/craft_eval/README.md

@@ -0,0 +1,59 @@
+# craft A/B 成稿质量前后对照评估回路（灵感 F2）
+
+把「加 craft 杠杆」与「验证 craft 有效」绑成一件事的评估回路。每引入或改动一个 craft
+杠杆（① 写章教条 / F1 genre 片段 / 开篇黄金三章 / 后续 ② 反哺、③ 人物塑造审查等），
+就跑一次这里，做 craft-on vs craft-off 的前后对照——**没有它，craft 改动都是盲飞，
+很可能把文风做得更 AI 公式化而不自知**（见计划 §5 风险 #13）。
+
+## 它是什么
+
+- 固定一组**金标准题材输入**（`fixtures.py`：仙侠 / 都市 / 悬疑，各含最小 spec+world+outline，
+  且都是开篇章以覆盖黄金三章 marker）。
+- 对每份 fixture 走**真实生产组装路径**（`assemble()` + `build_write_request()`）产出两版
+  写章请求：
+  - **craft-on** = 生产路径：craft 教条置顶 system 块 + genre 片段进 stable_core 缓存前缀 +
+    开篇 marker 进 volatile。
+  - **craft-off** = 同一 fixture 抹掉 craft 层的基线（剥掉 genre/开篇两节、不加教条块）。
+    两版**唯一差异即 craft 贡献**，便于并排 diff 与 LLM 裁判对比。
+
+## 两种模式
+
+### dry（默认，零成本、免真实 LLM）
+
+```bash
+export PATH="$HOME/.local/bin:$PATH"
+uv run python -m tests.craft_eval            # 全部题材
+uv run python -m tests.craft_eval --genre 仙侠
+```
+
+只做**确定性校验 + prompt 并排 dump**：断言 craft 教条块、genre 片段、开篇 marker 确实按
+fixture 注入了正确内容（且 craft-off 基线不含这三者），并把两版**将发送给模型的 prompt**
+并排写进报告供人眼看。**不碰 LLM**，任一断言失败退出码非零——这些断言即
+`tests/test_craft_eval.py`，随常规 `pytest` / CI 跑，是 craft 注入路径的回归门禁。
+
+### live（真实付费生成 + LLM 五维裁判）
+
+```bash
+uv run python -m tests.craft_eval --live
+```
+
+**成本警告**：`--live` **真实调用付费 LLM**——每个题材跑 writer 生成 craft-on/craft-off 两章
+＋ analyst 五维裁判 1 次（3 题材 ≈ 6 次写章 + 3 次裁判）。凭据从 DB 读、路由走 `tier_routing`，
+与生产写章同一条网关路径（需 pg 在跑且已配好 provider 凭据；`--live` 前会先跑 dry 校验，
+未全绿则拒绝进入付费生成）。**绝不进 CI**：live 路径在 `live.py`，pytest 不导入它。
+
+报告默认写系统临时目录（`--out PATH` 自定义）；含每题材的 dry 校验清单、prompt 并排、
+以及 live 模式下的两版成稿与五维裁判结论。
+
+## 加了新 craft 杠杆后怎么用它把关
+
+1. 改 craft（教条 `write_craft.md` / `GENRE_CRAFT` / 新审查维度……）。
+2. 跑 `pytest tests/test_craft_eval.py` 确认注入路径没被改坏（dry 门禁）。
+3. 跑 `--live` 生成前后对照，人眼 + LLM 五维裁判判断**是更对味还是更公式化**——
+   据此决定该杠杆保留 / 回退（计划 §5 通用回滚：每杠杆可独立回退）。
+
+## 加题材 / 调 fixture
+
+在 `fixtures.py` 追加一个 `FixtureCase`（`genre` 必须命中 `ww_config.GENRE_CRAFT` 才有
+片段可注入）。`harness.py` 的 craft-off 靠 `assemble` 的节标题（`本作题材写法` / `开篇黄金三章`）
+剥离——上游若改节名，剥离失配会被 dry 断言当场炸出（craft-off 里仍含片段 → 断言失败）。

tests/craft_eval/__init__.py

@@ -0,0 +1,38 @@
+"""craft A/B 成稿质量前后对照评估回路（灵感 F2）。
+
+固定金标准题材输入 → 组装 craft-on / craft-off 两版写章请求 → dry 确定性校验（零成本、
+可进 CI）或 `--live` 真实生成 + LLM 五维裁判。用法见 `README.md` / `python -m tests.craft_eval -h`。
+"""
+
+from __future__ import annotations
+
+from .fixtures import FIXTURES, FixtureCase
+from .harness import (
+    OPENING_CHAPTER_LIMIT,
+    CaseReport,
+    Check,
+    PromptPair,
+    all_passed,
+    build_prompt_pair,
+    input_text,
+    render_report,
+    system_text,
+    verify_case,
+    write_report,
+)
+
+__all__ = [
+    "FIXTURES",
+    "OPENING_CHAPTER_LIMIT",
+    "CaseReport",
+    "Check",
+    "FixtureCase",
+    "PromptPair",
+    "all_passed",
+    "build_prompt_pair",
+    "input_text",
+    "render_report",
+    "system_text",
+    "verify_case",
+    "write_report",
+]

tests/craft_eval/__main__.py

@@ -0,0 +1,104 @@
+"""craft A/B 评估脚本入口（灵感 F2）。
+
+    # dry（默认，零成本、免真实 LLM）——组装+校验+并排 dump prompt
+    uv run python -m tests.craft_eval
+
+    # live（真实付费生成 + LLM 五维裁判）
+    uv run python -m tests.craft_eval --live
+
+选项：`--genre 仙侠` 只跑某题材；`--out PATH` 指定报告路径；`--quiet` 少打印。
+退出码：dry 任一断言失败 → 非零（可当回归门禁）。详见 README.md。
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import sys
+import tempfile
+from pathlib import Path
+
+from .fixtures import FIXTURES, FixtureCase
+from .harness import (
+    CaseReport,
+    all_passed,
+    build_prompt_pair,
+    render_report,
+    verify_case,
+    write_report,
+)
+
+
+def _select(genre: str | None) -> list[FixtureCase]:
+    if genre is None:
+        return list(FIXTURES)
+    picked = [c for c in FIXTURES if c.genre == genre]
+    if not picked:
+        raise SystemExit(f"没有题材为 {genre!r} 的 fixture；可选：{[c.genre for c in FIXTURES]}")
+    return picked
+
+
+def _default_out(live: bool) -> Path:
+    name = "craft_ab_live.md" if live else "craft_ab_dry.md"
+    return Path(tempfile.gettempdir()) / name
+
+
+async def _run(args: argparse.Namespace) -> int:
+    cases = _select(args.genre)
+
+    reports: list[CaseReport] = []
+    for case in cases:
+        pair = await build_prompt_pair(case)
+        reports.append(CaseReport(case=case, pair=pair, checks=verify_case(case, pair)))
+
+    failed = [(r.case.slug, c) for r in reports for c in r.checks if not c.passed]
+    if not args.quiet:
+        for r in reports:
+            status = "OK" if all_passed(r.checks) else "FAIL"
+            print(f"[dry] {r.case.slug:10s} {r.case.genre}  第{r.pair.chapter_no}章  {status}")
+    for slug, chk in failed:
+        print(f"[dry][FAIL] {slug} · {chk.name}: {chk.detail}", file=sys.stderr)
+
+    if args.live:
+        if failed:
+            print("[live] dry 校验未全绿，拒绝进入付费生成——先修正注入。", file=sys.stderr)
+            return 1
+        from .live import run_live, summarize
+
+        print("[live] 开始真实生成（付费）……", flush=True)
+        reports = await run_live(reports)
+        print("\n[live] 摘要：\n" + summarize(reports))
+
+    out = Path(args.out) if args.out else _default_out(args.live)
+    write_report(render_report(reports, live=args.live), out)
+    print(f"\n[out] 报告写至 {out}")
+
+    return 1 if failed else 0
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        prog="python -m tests.craft_eval",
+        description="craft A/B 成稿质量前后对照评估（灵感 F2）",
+    )
+    parser.add_argument(
+        "--live",
+        action="store_true",
+        help="真实调用付费 LLM 生成两版正文 + 五维裁判（默认 dry，免真实 LLM）",
+    )
+    parser.add_argument("--genre", default=None, help="只跑指定题材（如 仙侠/都市/悬疑）")
+    parser.add_argument("--out", default=None, help="报告输出路径（默认写系统临时目录）")
+    parser.add_argument("--quiet", action="store_true", help="减少 stdout 打印")
+    args = parser.parse_args()
+
+    if args.live:
+        # get_settings() 按 cwd 读 .env；确保 cwd = 仓库根（本文件在 tests/craft_eval/ 下三层）。
+        import os
+
+        os.chdir(Path(__file__).resolve().parents[2])
+
+    raise SystemExit(asyncio.run(_run(args)))
+
+
+if __name__ == "__main__":
+    main()

tests/craft_eval/fixtures.py

@@ -0,0 +1,158 @@
+"""金标准输入 fixtures（灵感 F2）——3 个不同题材的最小 spec+world+outline。
+
+每个 fixture 是一份**自造、不依赖 DB** 的最小设定：书级蓝本（含 genre）+ 世界观硬规则 +
+定型主角 + 本章大纲节拍。评估脚本据此组装 craft-on / craft-off 两版写章请求。
+新增题材只需追加一个 `FixtureCase`（genre 必须命中 `ww_config.GENRE_CRAFT` 才有片段可注）。
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+from ww_core.domain.repositories import CharacterView, ProjectSpecView, WorldEntityView
+
+
+@dataclass(frozen=True)
+class FixtureCase:
+    """一份金标准写章输入（题材 / 章号 / 书级蓝本 / 世界观 / 主角 / 本章节拍）。"""
+
+    slug: str
+    genre: str
+    chapter_no: int
+    spec: ProjectSpecView
+    world: tuple[WorldEntityView, ...]
+    characters: tuple[CharacterView, ...]
+    beats: dict[str, Any] = field(default_factory=dict)
+
+
+_XIANXIA = FixtureCase(
+    slug="xianxia",
+    genre="仙侠",
+    chapter_no=3,  # 开篇章（≤3）——覆盖黄金三章 marker 注入
+    spec=ProjectSpecView(
+        title="逆脉证道",
+        genre="仙侠",
+        logline="废脉少年叶尘为救绝症妹妹寻回春灵药，滴血觉醒上古逆天功法，逆修九境搅动仙门风云。",
+        premise="天地灵气稀薄，修行看灵脉品阶；废脉者纳灵极慢，被视作修行绝路。",
+        theme="卑微者的不屈与救赎——修的是道，护的是人。",
+        tone="热血隐忍",
+        narrative_pov="第三人称限知",
+        ending_type="正剧",
+    ),
+    world=(
+        WorldEntityView(
+            type="设定",
+            name="九境修为",
+            rules={
+                "境界": ["淬体", "开脉", "灵动", "纳灵", "凝魂", "化虚", "通玄", "渡劫", "证道"],
+                "每境": "分前中后三重",
+            },
+        ),
+        WorldEntityView(
+            type="设定",
+            name="灵脉",
+            rules={"作用": "决定纳灵速度", "废脉": "纳灵极慢，被视作修行绝路"},
+        ),
+    ),
+    characters=(
+        CharacterView(
+            name="叶尘",
+            role="主角",
+            motive="为身患寒髓之症、命不久矣的妹妹叶暖寻回春灵药",
+            backstory="青云宗外门弟子，天生废脉、常年受人轻慢，性子沉静隐忍。",
+            speech_tics={"风格": "话少、点到为止，逼到绝境时语气反而更冷更稳"},
+        ),
+    ),
+    beats={
+        "summary": "叶尘被同门构陷逐出青云宗，独行荒山避雨于古洞，"
+        "滴血觉醒《逆脉真解》，一招震退追杀者。",
+        "entities": ["叶尘"],
+        "text": "叶尘 古洞 滴血 觉醒 逆脉真解 震退 追杀者",
+    },
+)
+
+_URBAN = FixtureCase(
+    slug="urban",
+    genre="都市",
+    chapter_no=1,  # 开篇第一章
+    spec=ProjectSpecView(
+        title="重启我的黄金时代",
+        genre="都市",
+        logline="落魄程序员林川一夜重生回到二十年前，凭前世记忆在时代风口逆天改命。",
+        premise="重生回到世纪初互联网草莽年代，机会遍地、格局未定。",
+        theme="不甘平庸者抓住的第二次机会。",
+        tone="爽快励志",
+        narrative_pov="第一人称",
+        ending_type="HE",
+    ),
+    world=(
+        WorldEntityView(
+            type="设定",
+            name="重生规则",
+            rules={
+                "回溯": "重生回到 2003 年的大学时代",
+                "记忆": "完整保留前世记忆与二十年行业趋势",
+            },
+        ),
+    ),
+    characters=(
+        CharacterView(
+            name="林川",
+            role="主角",
+            motive="改写前世潦倒猝死的结局，护住家人、抓住时代机遇翻身",
+            backstory="前世三十五岁猝死于出租屋的落魄外包程序员。",
+            speech_tics={"风格": "表面随和、心里门儿清，说话带点自嘲"},
+        ),
+    ),
+    beats={
+        "summary": "林川在出租屋猝死后睁眼回到 2003 年的大学宿舍，"
+        "认出这是重来一次的机会，立下第一个目标。",
+        "entities": ["林川"],
+        "text": "林川 重生 2003 大学宿舍 认清机会",
+    },
+)
+
+_SUSPENSE = FixtureCase(
+    slug="suspense",
+    genre="悬疑",
+    chapter_no=2,  # 开篇第二章
+    spec=ProjectSpecView(
+        title="雾隐镇失踪案",
+        genre="悬疑",
+        logline="刑警周砚回到多雾的故乡小镇调查连环失踪案，线索却一步步指向自己尘封的童年。",
+        premise="常年浓雾笼罩的封闭小镇，二十年前也发生过同样的失踪，无人被起诉。",
+        theme="逼近真相与自我和解的代价。",
+        tone="冷峻压抑",
+        narrative_pov="第三人称限知",
+        ending_type="开放",
+    ),
+    world=(
+        WorldEntityView(
+            type="设定",
+            name="雾隐镇",
+            rules={
+                "环境": "常年浓雾、信号时断时续",
+                "旧案": "二十年前发生过同类失踪，卷宗离奇缺页",
+            },
+        ),
+    ),
+    characters=(
+        CharacterView(
+            name="周砚",
+            role="主角",
+            motive="查清连环失踪真相，也为解开自己童年缺失的那段记忆",
+            backstory="市局刑警，幼年从雾隐镇被送走，对故乡的记忆断片。",
+            speech_tics={"风格": "话短、爱用反问，审讯时不动声色"},
+        ),
+    ),
+    beats={
+        "summary": "周砚勘查第二名失踪者的住所，"
+        "在墙缝里发现一张与二十年前旧案笔迹相同的纸条，心头一凛。",
+        "entities": ["周砚"],
+        "text": "周砚 现场勘查 墙缝 纸条 二十年前 旧案",
+    },
+)
+
+# 金标准题材集：仙侠 / 都市 / 悬疑（各含开篇章，覆盖黄金三章 marker）。
+FIXTURES: tuple[FixtureCase, ...] = (_XIANXIA, _URBAN, _SUSPENSE)

tests/craft_eval/harness.py

@@ -0,0 +1,363 @@
+"""craft A/B 评估回路核心（灵感 F2）——组装请求 + dry 确定性校验 + 报告渲染。
+
+**零 LLM 依赖**：本模块只走确定性组装（真 `assemble()` + 真 `build_write_request()`）与
+纯字符串校验，可安全进 CI。真实生成/裁判在 `live.py`（惰性导入，不进本模块）。
+
+craft-on = 生产路径（`assemble` 注入 genre 片段进 stable_core + 开篇 marker 进 volatile，
+`build_write_request` 把 craft 教条置为最前 system 块）。
+craft-off = 同一 fixture 抹掉 craft 层的基线：剥掉 genre/开篇两节 + 不加教条块——
+两版**唯一差异即 craft 贡献**，便于人眼并排 diff 与 LLM 裁判对比。
+"""
+
+from __future__ import annotations
+
+import uuid
+from dataclasses import dataclass
+from typing import Any
+
+from ww_core.domain.repositories import (
+    CharacterView,
+    DigestView,
+    ForeshadowView,
+    MemoryRepos,
+    OutlineView,
+    ProjectSpecView,
+    RuleView,
+    StyleView,
+    WorldEntityView,
+)
+from ww_core.memory.assemble import OPENING_CHAPTER_LIMIT, assemble
+from ww_core.orchestrator.write_node import build_write_request
+from ww_llm_gateway.types import Block, LlmRequest, Scope
+
+from .fixtures import FixtureCase
+
+# 单用户原型 stub（dry 路径不落库，取值无关紧要，仅为满足 Scope）。
+PROJECT_ID = uuid.UUID("00000000-0000-0000-0000-0000000c7a17")
+USER_ID = uuid.UUID(int=1)
+
+# assemble 内联的 craft 节标题（assemble.py:97 `本作题材写法` / :141 `开篇黄金三章`）。
+# craft-off 据此剥离；若上游改名，本表失配 → craft-off 仍含片段 → 校验断言会当场炸出漂移。
+_GENRE_SECTION_TITLE = "本作题材写法"
+_OPENING_SECTION_TITLE = "开篇黄金三章"
+
+__all__ = [
+    "OPENING_CHAPTER_LIMIT",
+    "PROJECT_ID",
+    "USER_ID",
+    "CaseReport",
+    "Check",
+    "PromptPair",
+    "all_passed",
+    "build_prompt_pair",
+    "input_text",
+    "render_report",
+    "system_text",
+    "verify_case",
+    "write_report",
+]
+
+
+# ---- 内存 fake repositories（只喂 assemble 所需，其余返回空）----
+
+
+@dataclass
+class _FakeOutlineRepo:
+    """按查询章号合成 OutlineView（复用同一份 beats），支持任意 chapter_no 的边界测试。"""
+
+    volume: int
+    beats: dict[str, Any]
+
+    async def get(self, project_id: uuid.UUID, chapter_no: int) -> OutlineView | None:
+        return OutlineView(volume=self.volume, chapter_no=chapter_no, beats=dict(self.beats))
+
+    async def list_for_project(
+        self, project_id: uuid.UUID
+    ) -> list[OutlineView]:  # pragma: no cover
+        return []
+
+
+@dataclass
+class _FakeCharacterRepo:
+    rows: tuple[CharacterView, ...]
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[CharacterView]:
+        return list(self.rows)
+
+
+@dataclass
+class _FakeWorldEntityRepo:
+    rows: tuple[WorldEntityView, ...]
+
+    async def list_for_project(self, project_id: uuid.UUID) -> list[WorldEntityView]:
+        return list(self.rows)
+
+
+@dataclass
+class _FakeDigestRepo:
+    async def recent(self, project_id: uuid.UUID, k: int) -> list[DigestView]:
+        return []
+
+
+@dataclass
+class _FakeForeshadowRepo:
+    async def list_for_codes(self, project_id: uuid.UUID, codes: list[str]) -> list[ForeshadowView]:
+        return []
+
+
+@dataclass
+class _FakeStyleRepo:
+    async def latest(self, project_id: uuid.UUID) -> StyleView | None:
+        return None
+
+
+@dataclass
+class _FakeRulesRepo:
+    async def all_for_project(self, project_id: uuid.UUID) -> list[RuleView]:
+        return []
+
+
+@dataclass
+class _FakeProjectSpecRepo:
+    row: ProjectSpecView
+
+    async def spec(self, project_id: uuid.UUID) -> ProjectSpecView | None:
+        return self.row
+
+
+def _build_repos(case: FixtureCase) -> MemoryRepos:
+    volume = 1
+    return MemoryRepos(
+        outline=_FakeOutlineRepo(volume=volume, beats=case.beats),
+        character=_FakeCharacterRepo(case.characters),
+        world_entity=_FakeWorldEntityRepo(case.world),
+        digest=_FakeDigestRepo(),
+        foreshadow=_FakeForeshadowRepo(),
+        style=_FakeStyleRepo(),
+        rules=_FakeRulesRepo(),
+        project=_FakeProjectSpecRepo(case.spec),
+        review=None,  # 冲突裁决反哺不在 craft A/B 范围；优雅降级不反哺
+    )
+
+
+# ---- craft-on / craft-off 请求组装 ----
+
+
+@dataclass(frozen=True)
+class PromptPair:
+    craft_on: LlmRequest
+    craft_off: LlmRequest
+    chapter_no: int
+
+
+def _strip_section(text: str, title: str) -> str:
+    """按 `## {title}` 剥掉一节（节以 "\\n\\n" 分隔，节体无空行——见 assemble._section）。"""
+    header = f"## {title}"
+    kept = [chunk for chunk in text.split("\n\n") if not chunk.startswith(header)]
+    return "\n\n".join(kept)
+
+
+async def build_prompt_pair(case: FixtureCase, *, chapter_no: int | None = None) -> PromptPair:
+    """据 fixture 组装 craft-on（生产路径）与 craft-off（抹除 craft 层的基线）两版请求。"""
+    ch = chapter_no if chapter_no is not None else case.chapter_no
+    ctx = await assemble(_build_repos(case), PROJECT_ID, ch)
+
+    craft_on = build_write_request(
+        stable_core=ctx.stable_core,
+        volatile=ctx.volatile,
+        user_id=USER_ID,
+        project_id=PROJECT_ID,
+    )
+
+    off_stable = _strip_section(ctx.stable_core, _GENRE_SECTION_TITLE)
+    off_volatile = _strip_section(ctx.volatile, _OPENING_SECTION_TITLE)
+    craft_off = LlmRequest(
+        tier="writer",
+        system=[Block(text=off_stable, cache=True)],
+        input=off_volatile,
+        stream=True,
+        scope=Scope(user_id=USER_ID, project_id=PROJECT_ID),
+    )
+    return PromptPair(craft_on=craft_on, craft_off=craft_off, chapter_no=ch)
+
+
+# ---- prompt 文本抽取（供校验与报告）----
+
+
+def system_text(req: LlmRequest) -> str:
+    return "\n\n".join(b.text for b in req.system)
+
+
+def input_text(req: LlmRequest) -> str:
+    if isinstance(req.input, str):
+        return req.input
+    return "\n\n".join(b.text for b in req.input)
+
+
+# ---- dry 确定性校验（无 LLM，可进 CI）----
+
+
+@dataclass(frozen=True)
+class Check:
+    name: str
+    passed: bool
+    detail: str
+
+
+def _load_doctrine() -> str:
+    # 惰性导入避免 fixtures 模块与本模块的导入顺序耦合。
+    from ww_agents import load_craft_doctrine
+
+    return load_craft_doctrine()
+
+
+def _genre_fragment(genre: str) -> str:
+    from ww_config import genre_craft_fragment
+
+    fragment = genre_craft_fragment(genre)
+    if fragment is None:  # pragma: no cover - fixtures 已保证命中
+        raise ValueError(f"题材 {genre!r} 无 GENRE_CRAFT 片段")
+    return fragment
+
+
+def verify_case(case: FixtureCase, pair: PromptPair) -> list[Check]:
+    """对一份 fixture 的 craft-on/off 请求做确定性断言（craft 改动的回归门禁）。"""
+    on, off = pair.craft_on, pair.craft_off
+    doctrine = _load_doctrine()
+    fragment = _genre_fragment(case.genre)
+    on_sys, on_in = system_text(on), input_text(on)
+    off_sys, off_in = system_text(off), input_text(off)
+    is_opening = pair.chapter_no <= OPENING_CHAPTER_LIMIT
+
+    checks: list[Check] = [
+        Check(
+            "craft_on_doctrine_first_block",
+            len(on.system) >= 2 and on.system[0].text == doctrine and on.system[0].cache,
+            "教条须为首个 system 块、字节等于 load_craft_doctrine()、cache=True（#9）",
+        ),
+        Check(
+            "craft_on_genre_in_stable_not_doctrine_not_input",
+            fragment in on.system[1].text
+            and fragment not in on.system[0].text
+            and fragment not in on_in,
+            f"genre 片段须进 stable_core 缓存块、不进教条块/volatile（{case.genre}）",
+        ),
+        Check(
+            "craft_on_opening_marker",
+            (_OPENING_SECTION_TITLE in on_in and _OPENING_SECTION_TITLE not in on_sys)
+            if is_opening
+            else (_OPENING_SECTION_TITLE not in on_in),
+            f"开篇 marker：第 {pair.chapter_no} 章"
+            + ("须进 volatile、不进缓存 system" if is_opening else "非开篇章须不注入"),
+        ),
+        Check(
+            "craft_on_tier_writer",
+            on.tier == "writer",
+            "不变量 #2：只声明 tier=writer，不写具体 model",
+        ),
+        Check(
+            "craft_off_no_doctrine",
+            doctrine not in off_sys and doctrine not in off_in,
+            "craft-off 基线不含教条块",
+        ),
+        Check(
+            "craft_off_no_genre_fragment",
+            fragment not in off_sys and fragment not in off_in,
+            "craft-off 基线不含 genre 片段",
+        ),
+        Check(
+            "craft_off_no_opening_marker",
+            _OPENING_SECTION_TITLE not in off_in,
+            "craft-off 基线不含开篇 marker",
+        ),
+        Check(
+            "craft_off_retains_body",
+            case.spec.title in off_sys and "请创作第" in off_in,
+            "剥离未误伤本体：书名与写作指令仍在",
+        ),
+    ]
+    return checks
+
+
+def all_passed(checks: list[Check]) -> bool:
+    return all(c.passed for c in checks)
+
+
+# ---- 报告渲染 ----
+
+
+@dataclass(frozen=True)
+class CaseReport:
+    case: FixtureCase
+    pair: PromptPair
+    checks: list[Check]
+    live: Any = None  # live.LiveResult | None（惰性，避免 dry 路径导入 live）
+
+
+def _fence(label: str, body: str) -> str:
+    return f"**{label}**\n\n```\n{body}\n```"
+
+
+def _render_pair(pair: PromptPair) -> str:
+    on, off = pair.craft_on, pair.craft_off
+    return "\n\n".join(
+        [
+            "#### craft-on（发送给模型的 prompt）",
+            _fence("system（缓存前缀）", system_text(on)),
+            _fence("input（volatile）", input_text(on)),
+            "#### craft-off（基线 prompt）",
+            _fence("system（缓存前缀）", system_text(off)),
+            _fence("input（volatile）", input_text(off)),
+        ]
+    )
+
+
+def _render_live(live: Any) -> str:
+    return "\n\n".join(
+        [
+            "#### 生成结果（--live）",
+            f"served_by: `{live.served_by}`",
+            _fence(f"craft-on 成稿（{live.chars_on} 字）", live.chapter_on),
+            _fence(f"craft-off 成稿（{live.chars_off} 字）", live.chapter_off),
+            "#### LLM 五维裁判",
+            live.judge_text,
+        ]
+    )
+
+
+def render_report(reports: list[CaseReport], *, live: bool) -> str:
+    total_checks = sum(len(r.checks) for r in reports)
+    failed = sum(1 for r in reports for c in r.checks if not c.passed)
+    lines: list[str] = [
+        "# craft A/B 评估报告（灵感 F2）",
+        "",
+        f"模式：{'live（真实生成+裁判）' if live else 'dry（零成本确定性校验）'}",
+        f"题材：{len(reports)} 个 · 断言：{total_checks - failed}/{total_checks} 通过",
+        "",
+    ]
+    for r in reports:
+        c = r.case
+        lines.append(f"## {c.slug} · {c.genre} · 《{c.spec.title}》 · 第 {r.pair.chapter_no} 章")
+        lines.append("")
+        lines.append("### dry 校验")
+        for chk in r.checks:
+            mark = "✅" if chk.passed else "❌"
+            lines.append(f"- {mark} `{chk.name}` — {chk.detail}")
+        lines.append("")
+        lines.append("### prompt 并排")
+        lines.append("")
+        lines.append(_render_pair(r.pair))
+        lines.append("")
+        if r.live is not None:
+            lines.append(_render_live(r.live))
+            lines.append("")
+    return "\n".join(lines)
+
+
+def write_report(text: str, path: Any) -> Any:
+    from pathlib import Path
+
+    out = Path(path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(text, encoding="utf-8")
+    return out

tests/craft_eval/live.py

@@ -0,0 +1,127 @@
+"""`--live` 真实生成 + LLM 五维裁判（灵感 F2）。
+
+**成本警告**：本模块**真实调用付费 LLM**（每题材：writer 生成 2 章 + analyst 裁判 1 次）。
+仅供人工手动跑，**绝不进 pytest / CI**（故与 dry 校验分文件，pytest 不导入本模块）。
+凭据从 DB 读（`SqlCredentialStore`），路由从 `tier_routing` 取——与生产写章同一条网关路径。
+
+用法见 `README.md`：`uv run python -m tests.craft_eval --live`。
+"""
+
+from __future__ import annotations
+
+import dataclasses
+from dataclasses import dataclass
+
+from ww_api.services.credentials import STUB_OWNER_ID, SqlCredentialStore
+from ww_api.services.project_deps import build_gateway_for_tier
+from ww_db import get_sessionmaker
+from ww_llm_gateway.types import Block, LlmRequest, Scope
+
+from .harness import CaseReport, PromptPair
+
+# 五维对比裁判（对齐 craft_ab 实验的口径）——只就文本本身逐维对比，不复述原文。
+JUDGE_SYSTEM = """你是资深中文网络小说主编，阅稿无数，眼光毒辣、直言不讳。
+现在给你同一章大纲下、由两套不同写作指令产出的两版正文（A=craft-on，B=craft-off）。
+请仅就文本本身，逐维度对比谁更好，每个维度给出「A更好 / B更好 / 难分」并附一句依据：
+① show-don't-tell（情绪是否靠动作细节外化，而非旁白直说）
+② 对话是否有角色声音/口癖
+③ 章末钩子是否成立（是否让读者想追下一章）
+④ 是否注水/AI 腔（空泛排比、翻译腔、无信息量的辞藻堆砌）
+⑤ 开篇代入感（前几段是否立住冲突/身份钩子、抓人）
+最后给出总判：「A 更好 / B 更好 / 难分」，并用 1-2 句话说明理由。
+保持简洁、务实，不要复述原文。"""
+
+# 裁判每稿截断字数（省 token）。
+_JUDGE_EXCERPT = 1500
+
+
+@dataclass(frozen=True)
+class LiveResult:
+    chapter_on: str
+    chapter_off: str
+    chars_on: int
+    chars_off: int
+    judge_text: str
+    served_by: str
+
+
+def char_count(text: str) -> int:
+    """中文成稿近似字数：去空白后的字符数。"""
+    return len("".join(text.split()))
+
+
+async def _stream_text(gateway: object, req: LlmRequest) -> str:
+    parts: list[str] = []
+    async for delta in gateway.stream(req):  # type: ignore[attr-defined]
+        parts.append(delta.text)
+    return "".join(parts)
+
+
+def _judge_request(pair: PromptPair, text_on: str, text_off: str) -> LlmRequest:
+    body = (
+        "以下是两版正文（各截取前若干字以省 token）。\n\n"
+        f"===== A 稿 / craft-on（前 {_JUDGE_EXCERPT} 字）=====\n"
+        + text_on[:_JUDGE_EXCERPT]
+        + f"\n\n===== B 稿 / craft-off（前 {_JUDGE_EXCERPT} 字）=====\n"
+        + text_off[:_JUDGE_EXCERPT]
+    )
+    return LlmRequest(
+        tier="analyst",
+        system=[Block(text=JUDGE_SYSTEM, cache=True)],
+        input=body,
+        stream=False,
+        scope=Scope(user_id=STUB_OWNER_ID, project_id=None),
+    )
+
+
+async def run_live(reports: list[CaseReport]) -> list[CaseReport]:
+    """对每个 report 的 craft-on/off 请求真实生成一章 + 五维裁判，返回附 live 的新 report 列表。
+
+    不可变更新：用 `dataclasses.replace` 产出新 `CaseReport`，不原地改冻结实例。
+    """
+    sm = get_sessionmaker()
+    out: list[CaseReport] = []
+    async with sm() as session:
+        store = SqlCredentialStore(session)
+        # writer / analyst 各建一次网关，跨题材复用（省装配开销）。
+        writer_gw = await build_gateway_for_tier(session, store, "writer")
+        analyst_gw = await build_gateway_for_tier(session, store, "analyst")
+
+        for report in reports:
+            pair = report.pair
+            print(f"[live] {report.case.slug}：生成 craft-on……", flush=True)
+            text_on = await _stream_text(writer_gw, _as_scoped(pair.craft_on))
+            print(f"[live] {report.case.slug}：生成 craft-off……", flush=True)
+            text_off = await _stream_text(writer_gw, _as_scoped(pair.craft_off))
+            print(f"[live] {report.case.slug}：analyst 五维裁判……", flush=True)
+            resp = await analyst_gw.run(_judge_request(pair, text_on, text_off))
+
+            result = LiveResult(
+                chapter_on=text_on,
+                chapter_off=text_off,
+                chars_on=char_count(text_on),
+                chars_off=char_count(text_off),
+                judge_text=resp.text,
+                served_by=f"{resp.served_by.provider}:{resp.served_by.model}"
+                + (" (fell_back)" if resp.served_by.fell_back else ""),
+            )
+            out.append(dataclasses.replace(report, live=result))
+    return out
+
+
+def _as_scoped(req: LlmRequest) -> LlmRequest:
+    """把请求 scope 换成 stub owner（usage_ledger.owner_id FK 须指向已 seed 的用户行）。"""
+    scoped = Scope(user_id=STUB_OWNER_ID, project_id=None)
+    return req.model_copy(update={"scope": scoped})
+
+
+def summarize(reports: list[CaseReport]) -> str:
+    """live 结果的一行式摘要（stdout 用）。"""
+    lines: list[str] = []
+    for r in reports:
+        if r.live is None:  # pragma: no cover - live 路径必有结果
+            continue
+        lines.append(
+            f"{r.case.slug}: craft-on {r.live.chars_on}字 / craft-off {r.live.chars_off}字"
+        )
+    return "\n".join(lines)

tests/test_craft_eval.py

@@ -0,0 +1,121 @@
+"""craft A/B 评估回路的 dry 模式确定性断言（灵感 F2）。
+
+这些断言 = 评估脚本 dry 模式的回归门禁：验证 craft 教条块 / genre 片段 / 开篇 marker
+**确实按 fixture 注入了正确内容**，且 craft-off 基线**不含**这三者。dry 路径不碰 LLM，
+故可进常规 pytest / CI（`--live` 真实生成路径不在此、不进 CI，见 tests/craft_eval/README.md）。
+"""
+
+from __future__ import annotations
+
+import pytest
+from ww_agents import load_craft_doctrine
+from ww_config import genre_craft_fragment
+
+from tests.craft_eval import (
+    FIXTURES,
+    FixtureCase,
+    all_passed,
+    build_prompt_pair,
+    input_text,
+    render_report,
+    system_text,
+    verify_case,
+    write_report,
+)
+from tests.craft_eval.harness import OPENING_CHAPTER_LIMIT, CaseReport
+
+_OPENING_HEADER = "开篇黄金三章"
+
+
+def test_fixtures_cover_three_distinct_genres() -> None:
+    genres = {c.genre for c in FIXTURES}
+    assert {"仙侠", "都市", "悬疑"} <= genres
+    assert len(FIXTURES) == len(genres), "每个 fixture 应是不同题材"
+    for genre in genres:
+        assert genre_craft_fragment(genre), f"题材 {genre!r} 缺 GENRE_CRAFT 片段"
+
+
+@pytest.mark.parametrize("case", FIXTURES, ids=lambda c: c.slug)
+async def test_dry_verification_all_pass(case: FixtureCase) -> None:
+    """每个金标准 fixture 的 dry 校验全绿——这是 craft 改动的回归门禁本体。"""
+    pair = await build_prompt_pair(case)
+    checks = verify_case(case, pair)
+    failed = [c for c in checks if not c.passed]
+    assert not failed, "dry 校验失败：\n" + "\n".join(f"  {c.name}: {c.detail}" for c in failed)
+    assert all_passed(checks)
+
+
+@pytest.mark.parametrize("case", FIXTURES, ids=lambda c: c.slug)
+async def test_craft_on_injects_doctrine_genre_and_marker(case: FixtureCase) -> None:
+    pair = await build_prompt_pair(case)
+    on = pair.craft_on
+
+    # 教条块：第一个 system 块、字节等于 load_craft_doctrine()、cache=True（不变量 #9）。
+    assert on.system[0].text == load_craft_doctrine()
+    assert on.system[0].cache is True
+    assert len(on.system) >= 2, "craft-on 应有 [教条, stable_core] 两块"
+
+    # genre 片段：进 stable_core 缓存块（book-stable → 缓存前缀），不进易变 input（#9）。
+    fragment = genre_craft_fragment(case.genre)
+    assert fragment is not None
+    assert fragment in on.system[1].text
+    assert fragment not in on.system[0].text, "genre 片段不应污染教条块"
+    assert fragment not in input_text(on), "genre 片段是 book-stable，不应进 volatile input"
+
+    # 开篇 marker：开篇章进 volatile input、不进缓存 system（含章号→易变，#9）。
+    if case.chapter_no <= OPENING_CHAPTER_LIMIT:
+        assert _OPENING_HEADER in input_text(on)
+        assert _OPENING_HEADER not in system_text(on)
+
+    # 不变量 #2：只声明 tier，不写 model。
+    assert on.tier == "writer"
+
+
+@pytest.mark.parametrize("case", FIXTURES, ids=lambda c: c.slug)
+async def test_craft_off_baseline_excludes_craft(case: FixtureCase) -> None:
+    pair = await build_prompt_pair(case)
+    off = pair.craft_off
+
+    assert load_craft_doctrine() not in system_text(off)
+    fragment = genre_craft_fragment(case.genre)
+    assert fragment is not None
+    assert fragment not in system_text(off)
+    assert fragment not in input_text(off)
+    assert _OPENING_HEADER not in input_text(off)
+    # 基线仍是可用写章请求：保留书名与写作指令（stripping 未误伤本体）。
+    assert case.spec.title in system_text(off)
+    assert "请创作第" in input_text(off)
+
+
+async def test_opening_marker_absent_beyond_opening_limit() -> None:
+    """章号 > 开篇上限时 craft-on 不注入开篇 marker（黄金三章边界）。"""
+    case = FIXTURES[0]
+    pair = await build_prompt_pair(case, chapter_no=OPENING_CHAPTER_LIMIT + 1)
+    assert _OPENING_HEADER not in input_text(pair.craft_on)
+    assert _OPENING_HEADER not in system_text(pair.craft_on)
+
+
+async def test_render_and_write_report(tmp_path) -> None:  # type: ignore[no-untyped-def]
+    reports: list[CaseReport] = []
+    for case in FIXTURES:
+        pair = await build_prompt_pair(case)
+        reports.append(CaseReport(case=case, pair=pair, checks=verify_case(case, pair)))
+    text = render_report(reports, live=False)
+    assert "craft-on" in text and "craft-off" in text
+    for case in FIXTURES:
+        assert case.slug in text
+    out = write_report(text, tmp_path / "report.md")
+    assert out.exists()
+    assert out.read_text("utf-8") == text
+
+
+def test_build_prompt_pair_is_deterministic_across_runs() -> None:
+    """同一 fixture 两次组装字节一致（缓存前缀稳定 / 无时间戳-UUID，#9）。"""
+    import asyncio
+
+    case = FIXTURES[0]
+    a = asyncio.run(build_prompt_pair(case))
+    b = asyncio.run(build_prompt_pair(case))
+    assert system_text(a.craft_on) == system_text(b.craft_on)
+    assert input_text(a.craft_on) == input_text(b.craft_on)
+    assert system_text(a.craft_off) == system_text(b.craft_off)

tests/test_m5_e2e.py

@@ -101,6 +101,8 @@ _CHARACTER_CARDS = CharacterGenResult(
             name="叶无尘",
             role="主角",
             traits=["隐忍", "腹黑"],
+            appearance="剑眉星目，玄袍加身，眉心一点朱砂",
+            motive="核心动机：查明灭门真相并复仇；深处渴望守护余下之人",
             backstory="灭门遗孤，背负血仇。",
             arc="从复仇者走向守护者。",
             speech_tics=["呵，有意思", "不过如此"],
@@ -393,6 +395,9 @@ async def test_m5_generate_characters_gate_acknowledge_and_persist(
             assert ye.relations == []
             assert ye.role == "主角"
             assert ye.backstory == "灭门遗孤，背负血仇。"
+            # ⑧ 穿线：appearance/motive Text 列真落 pg（此前一直 NULL）。
+            assert ye.appearance == "剑眉星目，玄袍加身，眉心一点朱砂"
+            assert ye.motive == "核心动机：查明灭门真相并复仇；深处渴望守护余下之人"
 
             # 记账真落 pg：world generate(writer) + characters generate(writer) +
             # precheck(analyst, 409 路径) + precheck(analyst, 入库路径) = 4 条。
@@ -412,6 +417,72 @@ async def test_m5_generate_characters_gate_acknowledge_and_persist(
             await _cleanup(e2e_sm, project_uuid)
 
 
+async def test_m5_character_upsert_backfills_appearance_motive(
+    e2e_sm: async_sessionmaker[AsyncSession],
+) -> None:
+    """⑧ 穿线回归：老角色（appearance/motive 为空）再次 upsert 应回填两列（update 分支）。
+
+    直连 `SqlCharacterWriteRepo`（不经端点）钉住 update 分支对新列赋值——否则老角色
+    永远补不上 appearance/motive（风险 8）。
+    """
+    from ww_api.services.credentials import STUB_OWNER_ID
+    from ww_api.services.project_deps import seed_stub_user
+    from ww_core.domain.character_repo import SqlCharacterWriteRepo
+    from ww_core.domain.project_repo import ProjectCreate, SqlProjectRepo
+
+    def _card_kwargs(appearance: str, motive: str) -> dict[str, object]:
+        return {
+            "name": "旧角色",
+            "role": "配角",
+            "traits": ["沉默"],
+            "appearance": appearance,
+            "motive": motive,
+            "backstory": "早期入库",
+            "arc": "无",
+            "speech_tics": [],
+            "tags": [],
+            "relations": [],
+        }
+
+    project_uuid: uuid.UUID | None = None
+    try:
+        # Arrange：seed stub user + 建项目 + 先入一张 appearance/motive 为空的老卡。
+        async with e2e_sm() as s:
+            await seed_stub_user(s)
+            project = await SqlProjectRepo(s).create(
+                STUB_OWNER_ID, ProjectCreate(title="⑧ 回填验证")
+            )
+            project_uuid = uuid.UUID(str(project.id))
+            await SqlCharacterWriteRepo(s).create(project_uuid, **_card_kwargs("", ""))  # type: ignore[arg-type]
+            await s.commit()
+
+        # Act：同名再次 upsert，这次带 appearance/motive（模拟重生成补全）。
+        async with e2e_sm() as s:
+            await SqlCharacterWriteRepo(s).create(
+                project_uuid,
+                **_card_kwargs("满脸风霜，独臂", "替旧主守住一个秘密"),  # type: ignore[arg-type]
+            )
+            await s.commit()
+
+        # Assert：仍一行（按 (project_id,name) upsert），两列被回填。
+        async with e2e_sm() as verify:
+            rows = (
+                (
+                    await verify.execute(
+                        select(Character).where(Character.project_id == project_uuid)
+                    )
+                )
+                .scalars()
+                .all()
+            )
+            assert len(rows) == 1
+            assert rows[0].appearance == "满脸风霜，独臂"
+            assert rows[0].motive == "替旧主守住一个秘密"
+    finally:
+        if project_uuid is not None:
+            await _cleanup(e2e_sm, project_uuid)
+
+
 async def test_m5_provider_fallback_serves_when_primary_fails(
     e2e_sm: async_sessionmaker[AsyncSession],
 ) -> None: