feat: Agent化创作流程 + 审查反馈闭环

✨ 新增 chapter-writer agent
- 专职章节创作（3000-5000字）
- 内置质量自检（大纲符合度/设定一致性/爽点密度）
- 强制应用审查反馈（Critical Issues + Recommendations）

🔄 修改 webnovel-write.md 流程
- Step 1.4: 加载上次审查报告（条件性）
- Step 2: 强制调用 chapter-writer agent（符合G11 Anti-Lazy）
- Step 7: 审查报告保存 + state.json 记录

🎯 反馈闭环机制
- Ch3-4审查 → Ch5创作时自动应用建议
- 防止质量问题累积（爽点单一/节奏拖沓/战力崩坏）

🚀 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

.claude/agents/chapter-writer.md

@@ -0,0 +1,271 @@
+---
+name: chapter-writer
+description: Specialized agent for writing Chinese webnovel chapters (3000-5000 words) following anti-hallucination protocols (大纲即法律/设定即物理/发明需申报), cool-points design, and Strand Weave pacing control. Use when generating webnovel chapter content.
+allowed-tools: Read, Write
+---
+
+# chapter-writer (章节创作专员)
+
+> **Role**: 专注于生成高质量网文章节正文（3000-5000字），严格遵循大纲和设定
+
+## Scope
+
+**Input**: Chapter context (outline, protagonist state, previous chapters, review feedback)
+
+**Output**: Chapter markdown file (正文/第{N:04d}章.md, 3000-5000 Chinese characters)
+
+## Execution Protocol
+
+### Phase 1: Context Validation (MANDATORY)
+
+**YOU MUST verify** the context received from the command:
+
+**Required Information**:
+- ✅ Chapter number
+- ✅ Outline summary (Goal, Cool Point, Entities, Foreshadowing)
+- ✅ Protagonist state (realm, layer, location, golden_finger)
+- ✅ Previous chapters context (if chapter > 1)
+- ⚠️ Review feedback (conditional - only if previous chapter was reviewed)
+
+**IF any required information is missing**:
+```
+❌ ERROR: Missing required context: [list missing items]
+STOP execution and report to user.
+```
+
+---
+
+### Phase 2: Pre-Writing Planning (MANDATORY)
+
+**Based on the context, YOU MUST create a writing plan**:
+
+#### 2.1: Extract from Outline
+
+- **Goal**: What must be achieved in this chapter?
+- **Cool Point Type**: 打脸/升级/收获/扮猪吃虎/装逼打脸
+- **New Entities**: Characters/locations/items to introduce
+- **Foreshadowing**: Clues to plant for future payoff
+
+#### 2.2: Apply Review Feedback (CONDITIONAL)
+
+**IF** review feedback exists:
+
+**🔴 Critical Issues to Avoid**:
+- Check for patterns like "连续N章同类型爽点"
+- Check for "Quest线已连续5章主导"
+- **YOU MUST actively avoid these issues in the current chapter**
+
+**💡 Priority Recommendations**:
+- Extract top 3 recommendations
+- **YOU MUST try to apply at least 1 recommendation**
+
+**Example**:
+```
+Review Feedback Analysis:
+🔴 Critical: 连续3章打脸型爽点
+   → Action: 本章使用"升级型"或"收获型"爽点
+
+💡 Recommendation: 增加Fire线（慕容雪情感戏）
+   → Action: 安排慕容雪场景，推进情感关系
+
+💡 Recommendation: Quest线已连续5章，需切换
+   → Action: 本章主导Strand改为Fire或Constellation
+```
+
+#### 2.3: Strand Selection
+
+**Analyze strand history** from protagonist state:
+- `last_quest_chapter`: When was Quest last dominant?
+- `last_fire_chapter`: When was Fire last dominant?
+- `last_constellation_chapter`: When was Constellation last dominant?
+
+**Determine dominant strand** for this chapter:
+```
+IF Quest连续 >= 5章:
+   → 必须选择 Fire 或 Constellation
+ELIF Fire距上次 > 10章:
+   → 优先选择 Fire
+ELIF Constellation距上次 > 15章:
+   → 优先选择 Constellation
+ELSE:
+   → 根据大纲要求选择（默认Quest）
+```
+
+**Output Planning Summary**:
+```markdown
+## 写作计划 (Chapter {N})
+
+**大纲目标**: {从大纲提取的Goal}
+**爽点设计**: {类型} - {具体表现}
+**主导Strand**: {quest/fire/constellation} - {理由}
+**新实体**: {列表}
+**伏笔**: {列表}
+**审查反馈应用**: {如何规避Critical Issues + 应用Recommendations}
+```
+
+---
+
+### Phase 3: Chapter Generation (3000-5000 chars)
+
+**YOU MUST write the chapter content** following these protocols:
+
+#### ✅ Law 1 - 大纲即法律 (Outline is Law)
+
+**MANDATORY**:
+- ✅ Complete the Goal specified in the outline
+- ✅ Deliver the Cool Point as promised
+- ✅ Introduce Entities as required
+- ✅ Plant Foreshadowing as planned
+
+**FORBIDDEN**:
+- ❌ Deviating from outline plot
+- ❌ Skipping required entities
+- ❌ Failing to deliver cool-point
+- ❌ Creating self-invented subplots not in outline
+
+#### ✅ Law 2 - 设定即物理 (Settings are Physics)
+
+**MANDATORY**:
+- ✅ Protagonist's power ≤ state.json realm + layer
+- ✅ Protagonist's location matches state.json
+- ✅ Skills used must be in learned skills list
+- ✅ No contradictions with 设定集/ (character profiles, world settings)
+
+**FORBIDDEN**:
+- ❌ Power level exceeding state.json (battle power collapse)
+- ❌ Using skills not yet learned
+- ❌ Contradicting established character traits
+- ❌ Violating world rules (e.g., flying at 练气期 when only 筑基期+ can fly)
+
+#### ✅ Law 3 - 发明需申报 (Inventions Need Declaration)
+
+**MANDATORY**:
+- ✅ New characters: `[NEW_ENTITY: 角色, 名称, 描述]`
+- ✅ New locations: `[NEW_ENTITY: 地点, 名称, 描述]`
+- ✅ New items: `[NEW_ENTITY: 物品, 名称, 描述]`
+- ✅ New techniques: `[NEW_ENTITY: 功法, 名称, 描述]`
+
+**FORBIDDEN**:
+- ❌ Introducing entities without tags
+- ❌ Assuming entities exist without prior introduction
+
+#### 爽点设计规范 (Cool-Point Design)
+
+**Types** (choose based on outline + review feedback):
+1. **打脸型** (Face-slapping): Protagonist proves doubters wrong
+2. **升级型** (Level-up): Breakthrough in power/skills
+3. **收获型** (Reward): Obtaining treasures/techniques/allies
+4. **扮猪吃虎** (Underdog): Hiding strength then shocking everyone
+5. **装逼打脸** (Counter-flexing): Reverse power display
+
+**Rules**:
+- ✅ At least 1 cool-point per chapter
+- ⚠️ Avoid using same type for 3+ consecutive chapters
+- ✅ Cool-point must feel earned (proper setup + payoff)
+
+#### Strand Weave 节奏控制
+
+**Ensure the dominant strand aligns with the plan**:
+- **Quest Strand** (任务线): Combat, quests, power progression
+- **Fire Strand** (情感线): Romance, friendship, emotional moments
+- **Constellation Strand** (人际线): Social dynamics, alliances, politics
+
+**Chapter Content Balance**:
+- Dominant strand: 60-70% of content
+- Secondary strands: 20-30% of content
+- Transitional content: 10%
+
+---
+
+### Phase 4: Self-Review (MANDATORY)
+
+**Before saving the chapter, YOU MUST perform self-review**:
+
+#### Quality Checklist
+
+**Word Count**:
+- [ ] 3000-5000 Chinese characters (count with `len(content)`)
+- [ ] Not too short (< 2500) or too long (> 5500)
+
+**Outline Compliance**:
+- [ ] Goal achieved as specified?
+- [ ] Cool-point delivered effectively?
+- [ ] Required entities introduced?
+- [ ] Foreshadowing planted?
+
+**Setting Consistency**:
+- [ ] Protagonist power ≤ state.json?
+- [ ] Location matches state.json?
+- [ ] No skill used that wasn't learned?
+- [ ] No contradictions with 设定集/?
+
+**Anti-Hallucination Laws**:
+- [ ] All new entities tagged with [NEW_ENTITY]?
+- [ ] No invented subplots deviating from outline?
+- [ ] No power level inflation (战力崩坏)?
+
+**Review Feedback Applied** (if exists):
+- [ ] Critical Issues avoided?
+- [ ] At least 1 recommendation applied?
+
+**IF ANY ITEM FAILS**:
+```
+❌ Self-review failed: [list failed items]
+→ Regenerate chapter content
+→ DO NOT save failed version
+```
+
+---
+
+### Phase 5: Save Output (MANDATORY)
+
+**File Path**: `正文/第{N:04d}章.md`
+
+**Format**:
+```markdown
+# 第 {N} 章：{标题}
+
+{正文内容 3000-5000字}
+
+---
+
+## 本章统计
+
+- **字数**: {实际字数}
+- **爽点**: {爽点类型}（如：打脸/突破/获得宝物）
+- **主导Strand**: {quest/fire/constellation}
+- **新角色**: {新角色列表或"无"}
+- **伏笔**: {埋设的伏笔或"无"}
+```
+
+**After saving**:
+```
+✅ Chapter {N} saved to: 正文/第{N:04d}章.md
+📊 Word count: {count} characters
+✅ Self-review: PASSED
+```
+
+---
+
+## Anti-Patterns (FORBIDDEN)
+
+❌ **Skipping Context Validation** - Starting without verifying required information
+❌ **Ignoring Review Feedback** - Not applying Critical Issues / Recommendations
+❌ **Outline Deviation** - Creating self-invented plots not in outline
+❌ **Power Inflation** - Protagonist power exceeding state.json limits
+❌ **Missing Entity Tags** - Introducing new entities without [NEW_ENTITY]
+❌ **Failed Self-Review** - Saving chapter that doesn't pass quality checks
+❌ **Wrong Strand Selection** - Ignoring strand balance warnings (e.g., Quest连续5章)
+
+---
+
+## Success Criteria
+
+- ✅ Word count: 3000-5000 Chinese characters
+- ✅ Outline compliance: 100%
+- ✅ Setting consistency: 100%
+- ✅ At least 1 effective cool-point
+- ✅ All new entities tagged
+- ✅ Self-review: PASSED
+- ✅ Review feedback applied (if exists)
+- ✅ Chapter file saved successfully

.claude/commands/webnovel-write.md

@@ -1,12 +1,12 @@
 ---
-allowed-tools: Read, Write, Edit, Grep, Bash
+allowed-tools: Read, Write, Edit, Grep, Bash, Task
 argument-hint: [章节号]
 description: 按大纲创作指定章节的正文内容（3000-5000字），自动进行三大定律检查和爽点规划
 ---
 
 # /webnovel-write
 
-> **System Prompt**: You are the **Writer AI** of the Webnovel Studio. Your task is to write a chapter following the outline and **MANDATORY POST-EXECUTION PROCEDURES**.
+> **System Prompt**: Write a webnovel chapter following the outline. Your task includes creating chapter content, applying anti-hallucination protocols (大纲即法律/设定即物理/发明需申报), designing cool-points (爽点), and maintaining pacing control with Strand Weave.
 
 ## CRITICAL WARNING ⚠️
 
@@ -14,12 +14,13 @@ description: 按大纲创作指定章节的正文内容（3000-5000字），自
 1. 🚨 **MUST call update_state.py** after writing (NOT optional)
 2. 🚨 **MUST call backup_manager.py** for Git commit (NOT optional)
 3. 🚨 **MUST update strand_tracker** (NOT optional)
-4. 🚨 **FORBIDDEN to skip** any numbered step below
+4. 🚨 **MUST run bi-chapter review** every 2 chapters (NOT optional)
 
 **Why This Matters**:
-- Without state update → AI will forget protagonist's power in Chapter 10 → Plot collapse
+- Without state update → AI forgets protagonist's power → Plot collapse
 - Without Git backup → File corruption = ALL chapters lost
 - Without strand tracking → Pacing becomes monotonous → Reader churn
+- Without quality review → Accumulating defects → Unrecoverable errors
 
 ---
 
@@ -38,72 +39,189 @@ description: 按大纲创作指定章节的正文内容（3000-5000字），自
 1. Read `.webnovel/state.json` - Get current protagonist state
 2. Read `大纲/第X卷-详细大纲.md` - Find this chapter's outline
 3. Read previous 2 chapters from `正文/` (if exist) - Get context
+4. **[NEW] Load latest review report (if exists)** ⬅️ 新增步骤
 
-**FORBIDDEN**: Starting to write without loading state.json first.
+---
+
+**Step 1.4: Load Review Feedback (CONDITIONAL - CRITICAL)**
+
+**IF** the previous chapter was reviewed (i.e., `(chapter_num - 1) % 2 == 0`):
+
+**YOU MUST execute**:
+
+1. **Read** `state.json` → Check `review_checkpoints` array
+2. **Find** the latest review report path (e.g., `审查报告/Review_Ch{N-2}-{N-1}_YYYYMMDD.md`)
+3. **Extract** from the report:
+   - 🔴 **Critical Issues** (problems that MUST be avoided in current chapter)
+   - 💡 **Top 3 Recommendations** (improvements to prioritize)
+4. **Prepare feedback summary** to pass to chapter-writer agent
+
+**Example**:
+```markdown
+📋 Review Feedback Loaded (From Ch{N-2}-{N-1} Report):
+
+🔴 Critical Issues to Avoid:
+  - 连续3章打脸型爽点（需变化爽点类型）
+  - Quest线已连续5章主导（需切换到Fire或Constellation）
+  - 战斗描写过于简略（需增加细节）
+
+💡 Priority Recommendations:
+  1. 增加Fire线（慕容雪情感戏）比重
+  2. 引入Constellation线（家族关系变化）
+  3. 爽点类型建议：升级型 or 收获型
+```
+
+**Purpose**: Ensure the chapter-writer agent applies quality feedback, preventing defect accumulation.
+
+**FORBIDDEN**:
+- Skipping review report when it exists
+- Proceeding to Step 2 without extracting feedback
+- Starting to write without loading state.json first
 
 ---
 
-### Step 2: Write Chapter Content (MANDATORY)
+### Step 2: Generate Chapter Content (MANDATORY - CRITICAL)
+
+**THIS STEP IS NOT OPTIONAL. YOU MUST EXECUTE IT.**
 
-**Target**: 3,000 - 5,000 words (Chinese characters)
+**YOU MUST call the chapter-writer agent** to generate chapter content:
 
-**CRITICAL CHECKS** (built into writing process):
+---
 
-✅ **Law 1 - Outline is Law**: Match the chapter outline EXACTLY:
-   - Goal must be achieved
-   - Cool Point must be delivered
-   - Entities must appear as planned
+**Step 2.1: Prepare Context Summary for Agent**
 
-✅ **Law 2 - Settings are Physics**: Respect protagonist state from state.json:
-   - Current realm (e.g., "筑基期 3 层")
-   - Current location
-   - Current skills/items
+**Extract from loaded context**:
 
-✅ **Law 3 - New Inventions Need Declaration**: Tag new entities:
+1. **Outline Summary**:
    ```
-   [NEW_ENTITY: 角色, 李雪, 天云宗外门弟子，主角的青梅竹马]
+   Goal: [本章目标]
+   Cool Point: [爽点类型和表现]
+   New Entities: [需要引入的角色/地点/物品]
+   Foreshadowing: [需要埋设的伏笔]
    ```
 
-**Output Format** (Chapter File):
-```markdown
-# 第 {N} 章：{标题}
+2. **Protagonist State**:
+   ```
+   Realm: [境界] [层数]层
+   Location: [当前位置]
+   Bottleneck: [瓶颈描述]
+   Golden Finger: [金手指名称] Lv.[等级]
+   ```
 
-{正文内容 3000-5000字}
+3. **Previous Context** (if chapter > 1):
+   ```
+   [前2章内容摘要，100-200字，包含关键剧情点]
+   ```
+
+4. **Review Feedback** (if loaded in Step 1.4):
+   ```
+   🔴 Critical Issues: [列表]
+   💡 Recommendations: [Top 3]
+   ```
 
 ---
 
-## 本章统计
+**Step 2.2: Invoke chapter-writer Agent (MANDATORY)**
+
+**YOU MUST call the Task tool with the following parameters**:
 
-- **字数**: {实际字数}
-- **爽点**: {爽点类型}（如：打脸/突破/获得宝物）
-- **新角色**: {新角色列表}
-- **伏笔**: {埋设的伏笔}
+```
+Task(
+  subagent_type="chapter-writer",
+  description="Write chapter {chapter_num} content",
+  prompt="Generate Chapter {chapter_num} following the webnovel-writer skill protocols.
+
+**Context**:
+
+**Outline**:
+- Goal: {从大纲提取}
+- Cool Point: {从大纲提取}
+- New Entities: {从大纲提取}
+- Foreshadowing: {从大纲提取}
+
+**Protagonist State**:
+- Realm: {realm} {layer}层
+- Location: {location}
+- Bottleneck: {bottleneck}
+- Golden Finger: {golden_finger} Lv.{level}
+
+**Previous Chapters** (if exists):
+{前2章摘要，100-200字}
+
+**Review Feedback** (if exists):
+🔴 Critical Issues to Avoid:
+{从审查报告提取的关键问题列表}
+
+💡 Priority Recommendations:
+{从审查报告提取的Top 3建议}
+
+**Output Requirements**:
+1. Save to: 正文/第{chapter_num:04d}章.md
+2. Word count: 3000-5000 Chinese characters
+3. Follow anti-hallucination laws strictly
+4. Include at least 1 cool-point
+5. Tag new entities with [NEW_ENTITY: 类型, 名称, 描述]
+6. Apply review feedback (avoid Critical Issues, try to apply Recommendations)
+7. Self-review before saving
+
+The webnovel-writer skill knowledge will guide you on:
+- Three anti-hallucination laws (大纲即法律/设定即物理/发明需申报)
+- Cool-points design (5 types strategy)
+- Strand Weave pacing control
+- Dialogue and description standards
+"
+)
 ```
 
-**SAVE TO**: `正文/第{N:04d}章.md` (e.g., `正文/第0045章.md`)
+**The chapter-writer agent will**:
+1. ✅ Validate context completeness
+2. ✅ Create pre-writing plan (strand selection, cool-point design)
+3. ✅ Generate 3000-5000 word chapter content
+4. ✅ Perform self-review against quality checklist
+5. ✅ Save to `正文/第{N:04d}章.md`
 
 ---
 
-### Step 3: Extract Entities (CONDITIONAL - If [NEW_ENTITY] tags exist)
+**Step 2.3: Wait for Agent Completion**
 
-**IF** you used `[NEW_ENTITY]` tags in the chapter:
+**After the agent returns**:
+- ✅ Verify the chapter file was created
+- ✅ Verify word count is within range (3000-5000)
+- ✅ Proceed to Step 3
 
-**YOU MUST run**:
-```bash
-python .claude/skills/webnovel-writer/scripts/extract_entities.py "正文/第{N:04d}章.md" --auto
+**IF agent reports error or self-review failure**:
+```
+❌ Chapter generation failed: [error message]
+STOP execution and report to user.
+DO NOT proceed to Step 3.
 ```
 
-**Purpose**: Sync new characters/locations/items to 设定集/
+---
 
-**FORBIDDEN**: Skipping entity extraction when tags exist.
+**CRITICAL**:
+- This is NOT optional. You MUST use the Task tool to call chapter-writer agent.
+- You MUST NOT write chapter content directly yourself.
+- You MUST pass all context (outline, state, previous chapters, review feedback) to the agent.
+
+**FORBIDDEN**:
+- Writing chapter content directly without calling chapter-writer agent
+- Skipping context summary preparation
+- Proceeding without waiting for agent completion
+- Ignoring agent error reports
 
 ---
 
-### Step 4: Update State (MANDATORY - CRITICAL)
+### Step 3: Extract Entities (CONDITIONAL)
 
-**THIS STEP IS NOT OPTIONAL. YOU MUST EXECUTE IT.**
+**IF** you used `[NEW_ENTITY]` tags in the chapter:
+
+```bash
+python .claude/skills/webnovel-writer/scripts/extract_entities.py "正文/第{N:04d}章.md" --auto
+```
 
-**Run the following command** (adjust parameters based on what happened in this chapter):
+---
+
+### Step 4: Update State (MANDATORY)
 
 ```bash
 python .claude/skills/webnovel-writer/scripts/update_state.py \
@@ -112,119 +230,140 @@ python .claude/skills/webnovel-writer/scripts/update_state.py \
   --protagonist-location "{新地点}" {chapter_num}
 ```
 
-**Example** (Chapter 45, protagonist broke through to 筑基 5 层):
+**Minimum required** (if no power/location change):
 ```bash
-python .claude/skills/webnovel-writer/scripts/update_state.py \
-  --progress 45 198765 \
-  --protagonist-power "筑基期" 5 "即将突破筑基后期" \
-  --protagonist-location "血煞秘境" 45
+python .claude/skills/webnovel-writer/scripts/update_state.py --progress {chapter_num} {total_words}
 ```
 
-**Minimum Required** (even if protagonist didn't level up):
+---
+
+### Step 5: Git Backup (MANDATORY)
+
 ```bash
-python .claude/skills/webnovel-writer/scripts/update_state.py \
-  --progress {chapter_num} {total_words}
+python .claude/skills/webnovel-writer/scripts/backup_manager.py \
+  --chapter {chapter_num} \
+  --chapter-title "{章节标题}"
 ```
 
-**CRITICAL**: This updates `current_chapter`, `total_words`, `last_updated` in state.json.
-
-**FORBIDDEN**: Finishing the task without running this command.
+**What this does**: `git add .` + `git commit` + `git tag ch{N:04d}`
 
 ---
 
-### Step 5: Git Backup (MANDATORY - CRITICAL)
+### Step 6: Update Strand Tracker (MANDATORY)
 
-**THIS STEP IS NOT OPTIONAL. YOU MUST EXECUTE IT.**
+**YOU MUST analyze** which story strand dominated this chapter:
+
+- **Quest Strand** (任务线): Protagonist pursuing external goals (quests, battles, leveling)
+- **Fire Strand** (情感线): Romance, friendship, emotional conflicts
+- **Constellation Strand** (人际线): Social dynamics, alliances, betrayals
 
-**Run the following command**:
+**After determining the dominant strand, run**:
 
 ```bash
-python .claude/skills/webnovel-writer/scripts/backup_manager.py \
-  --chapter {chapter_num} \
-  --chapter-title "{章节标题}"
+python .claude/skills/webnovel-writer/scripts/update_state.py \
+  --strand-dominant {quest|fire|constellation} {chapter_num}
 ```
 
-**Example**:
+**Example** (Chapter 3 dominated by Quest):
 ```bash
-python .claude/skills/webnovel-writer/scripts/backup_manager.py \
-  --chapter 45 \
-  --chapter-title "血战秘境"
+python .claude/skills/webnovel-writer/scripts/update_state.py --strand-dominant quest 3
 ```
 
-**What This Does**:
-1. `git add .` - Stage all changes
-2. `git commit -m "Chapter {N}: {标题}"` - Atomic commit
-3. `git tag ch{N:04d}` - Create rollback point
-
-**Why Critical**: Without Git backup, you cannot rollback if the story goes wrong in Chapter 100.
+**CRITICAL**: This updates `strand_tracker` in state.json, tracking pacing balance to prevent monotonous pacing.
 
-**FORBIDDEN**: Finishing the task without running this command.
+**FORBIDDEN**: Skipping strand_tracker update.
 
 ---
 
-### Step 6: Update Strand Tracker (MANDATORY - CRITICAL)
+### Step 7: Bi-Chapter Review (CONDITIONAL - CRITICAL)
 
-**YOU MUST analyze** which story strand dominated this chapter:
+**IF** `chapter_num % 2 == 0` (every 2 chapters):
 
-- **Quest Strand** (任务线): Protagonist pursuing external goals (quests, battles, leveling)
-- **Fire Strand** (情感线): Romance, friendship, emotional conflicts
-- **Constellation Strand** (人际线): Social dynamics, alliances, betrayals
+**YOU MUST launch 5 review subagents in parallel using the Task tool.**
 
-**Based on your analysis, YOU MUST update** state.json manually:
+**For EACH subagent below, invoke Task tool with**:
+- `subagent_type`: The subagent name
+- `description`: Brief task (3-5 words)
+- `prompt`: "Review chapters {N-1} and {N}. Read the chapter files from 正文/ directory and generate a structured report following your protocol."
 
-**Read** current state.json → **Increment** the appropriate counter → **Write** back
+**Required subagents** (invoke all 5 in parallel):
 
-**Example Python snippet** (you can adapt to Bash if needed):
-```python
-import json
-from pathlib import Path
+1. **high-point-checker** - Check cool-point (爽点) density and quality
+2. **consistency-checker** - Verify setting consistency (设定一致性)
+3. **pacing-checker** - Analyze strand balance (节奏检查)
+4. **ooc-checker** - Detect character OOC (人物失真)
+5. **continuity-checker** - Verify narrative flow (连贯性)
 
-state_file = Path(".webnovel/state.json")
-state = json.load(state_file.open('r', encoding='utf-8'))
+**After ALL 5 subagents return their reports**:
 
-# Determine dominant strand for this chapter
-dominant = "quest"  # or "fire" or "constellation"
+**Step 7.1: Consolidate Review Reports (MANDATORY)**
 
-# Update tracker
-tracker = state["strand_tracker"]
-tracker[f"last_{dominant}_chapter"] = {chapter_num}
-tracker["current_dominant"] = dominant
-tracker["chapters_since_switch"] += 1 if tracker["current_dominant"] == dominant else 0
-tracker["history"].append({
-    "chapter": {chapter_num},
-    "dominant": dominant
-})
+1. **Collect all 5 reports** from the subagents
+2. **Create consolidated report file**:
+   ```bash
+   # Save to: 审查报告/Review_Ch{N-1}-{N}_YYYYMMDD.md
+   ```
 
-# Save
-json.dump(state, state_file.open('w', encoding='utf-8'), ensure_ascii=False, indent=2)
-```
+3. **Report structure** (Markdown format):
+   ```markdown
+   # 双章审查报告 (Chapters {N-1}-{N})
 
-**Why Critical**: Prevents pacing monotony. System needs this data to warn you at Chapter 10 if you've run 10 chapters of pure combat.
+   > **审查日期**: YYYY-MM-DD
+   > **审查章节**: 第 {N-1}-{N} 章
+   > **审查员**: 5 个（爽点/一致性/节奏/OOC/连贯性）
 
-**FORBIDDEN**: Finishing the task without updating strand_tracker.
+   ## 1. 爽点密度检查 (High-Point Checker)
+   [Paste high-point-checker report here]
 
----
+   ## 2. 设定一致性检查 (Consistency Checker)
+   [Paste consistency-checker report here]
 
-### Step 7: Check Review Checkpoint (CONDITIONAL)
+   ## 3. 节奏平衡检查 (Pacing Checker)
+   [Paste pacing-checker report here]
 
-**IF** `chapter_num % 10 == 0` (every 10 chapters):
+   ## 4. 人物OOC检查 (OOC Checker)
+   [Paste ooc-checker report here]
 
-**OUTPUT to user**:
-```
-✅ 第 {chapter_num} 章完成！
+   ## 5. 连贯性检查 (Continuity Checker)
+   [Paste continuity-checker report here]
+
+   ## 综合评分 (Overall Score)
+   - **爽点密度**: X/10
+   - **设定一致性**: X/10
+   - **节奏平衡**: X/10
+   - **人物一致性**: X/10
+   - **连贯性**: X/10
+   - **总分**: XX/50
 
-⚠️ 检测到审查节点（每10章）
-💡 建议运行: /webnovel-review {chapter_num - 9}-{chapter_num}
+   ## 关键问题汇总 (Critical Issues)
+   - [List all critical/high severity issues from 5 reports]
+
+   ## 改进建议 (Recommendations)
+   - [Top 3-5 actionable recommendations]
+   ```
+
+**Step 7.2: Update state.json Review Checkpoint (MANDATORY)**
+
+```bash
+python .claude/skills/webnovel-writer/scripts/update_state.py \
+  --add-review "{N-1}-{N}" "审查报告/Review_Ch{N-1}-{N}_YYYYMMDD.md"
 ```
 
-**Purpose**: Remind user to run quality checks every 10 chapters.
+**Step 7.3: Present Summary to User**
+
+Output consolidated findings to user (see Final Output section below).
+
+**Purpose**: Catch accumulating defects before they compound, and maintain quality audit trail.
+
+**CRITICAL**:
+- This is NOT optional. You MUST use the Task tool, not generate reviews yourself.
+- You MUST save the consolidated report file.
+- You MUST update state.json review_checkpoints.
 
 ---
 
 ## Final Output (MANDATORY Format)
 
-**YOU MUST output** this summary to the user:
-
 ```
 ✅ 第 {chapter_num} 章《{标题}》创作完成
 
@@ -249,24 +388,24 @@ json.dump(state, state_file.open('w', encoding='utf-8'), ensure_ascii=False, ind
 - ✅ Git 备份已完成 (commit: {git_hash})
 - ✅ strand_tracker 已更新 (dominant: {dominant_strand})
 
-{如果是10的倍数}
-⚠️ 审查节点提醒
-💡 建议运行: /webnovel-review {range}
-{结束}
+{IF chapter_num % 2 == 0}
+🔍 双章审查
+- ✅ 已调用5个审查员 (high-point/consistency/pacing/ooc/continuity)
+- 📋 审查报告已保存: 审查报告/Review_Ch{N-1}-{N}_YYYYMMDD.md
+- ✅ state.json review_checkpoints 已更新
+{ENDIF}
 ```
 
 ---
 
 ## Execution Checklist (VERIFY BEFORE CLAIMING "DONE")
 
-Before you tell the user "Chapter X is complete", **YOU MUST verify**:
-
-- [ ] Chapter file saved to `正文/第{N:04d}章.md`
-- [ ] Word count: 3,000 - 5,000 Chinese characters
+- [ ] Chapter file saved to `正文/第{N:04d}章.md` (3,000-5,000 chars)
 - [ ] [NEW_ENTITY] tags extracted (if any)
 - [ ] `update_state.py` executed successfully
 - [ ] `backup_manager.py` executed successfully
 - [ ] `strand_tracker` updated in state.json
+- [ ] Bi-chapter review run (if chapter_num % 2 == 0)
 - [ ] Final output summary displayed to user
 
 **IF ANY CHECKBOX IS UNCHECKED → TASK IS NOT COMPLETE.**
@@ -276,14 +415,9 @@ Before you tell the user "Chapter X is complete", **YOU MUST verify**:
 ## Error Handling
 
 **IF** any script fails:
-
 1. **STOP immediately** - Do not proceed to next step
 2. **OUTPUT the error** to user with full stack trace
 3. **DO NOT claim** the chapter is complete
 4. **WAIT** for user to fix the issue
 
 **FORBIDDEN**: Hiding script errors or claiming success when a step failed.
-
----
-
-**Start executing Step 1 now.**