long-doc-agent（长文档 Agent）

Ultra-Long Feasibility Report Multi-Agent Collaborative Writing v3.3

Changelog (v3.3)

• ✅ Table parsing fully fixed (_flush_table not calling _parse_md_table caused character-by-character splitting)
• ✅ Colorful chapter headings officially launched (H1 deep navy full-row band / H2 medium blue block / H3 light blue left bar)
• ✅ Styled tables launched (deep blue header + white text + alternating row colors)
• ✅ Key callout boxes launched (【关键】【注意】【优势】【风险】【数据】 color-coded cards)
• ✅ Cover style enhanced (tech-digital style deep ocean blue full-screen background + white text)
• ✅ Cover/Contents/Executive Summary title bars all colored
• ✅ Fixed cover_style integer vs string comparison preventing cover from applying
• ✅ Fixed RGBColor.from_string() instead of eval to avoid type errors

Core Capabilities

• Multi-Agent Parallel: Up to 5 sub-agents writing concurrently, doubled efficiency
• Incremental Updates: Chapters with unchanged content are skipped, faster processing
• Beautiful Formatting: Auto-generated cover, table-style TOC, colorful chapter headings, key callout boxes, styled tables
• Feishu RAG: Auto-search Feishu knowledge base to supplement reference materials
• 6 Cover Styles: Switch freely to suit different scenarios

File Structure

skill_dir/
├── SKILL.md                    # This file
├── integrate_report.py          # Core engine: parse/ integrate/CLI
├── parallel_tracker.py          # Parallel progress tracking
└── references/                 # Sub-process reference documents
    ├── phase0_guide.md         # Phase 0 requirements confirmation flow
    ├── phase1_guide.md         # Planner prompt template
    ├── phase2_guide.md         # Sub-Agent prompt template
    ├── table_format_guide.md   # Markdown table format specification
    └── bug_fix_guide.md        # Bug troubleshooting & forced rebuild

First-time setup: Ensure the F:/agent/chapters/ directory exists.

Pipeline Routing

User Task
├─ First writing request ("I want to write xxx"/"help me write a feasibility report")
│   → Phase 0 Requirements → Phase 1 Planner
│
├─ Outline exists, request to start writing
│   → Phase 2 Parallel Sub-Agents
│
├─ A chapter needs modification
│   → Small change: directly edit F:/agent/chapters/0X-xxx.txt
│   → Large change: regenerate that chapter
│
├─ All chapters done, request docx generation
│   → Phase 2.5 Review → Phase 3 Integrator
│
├─ Independent small proposal (2~5 chapters, no existing chapters dependency)
│   → Write Markdown directly → make_docx.py to generate styled docx
│   → See: references/bug_fix_guide.md "make_docx.py Mode"
│
└─ Just need to check progress/glossary/reference materials
    → Direct CLI commands

Phase 0: Requirements Confirmation

Confirm 4 items in order; all confirmed → Phase 1:

1. Writing Topic: document type / audience / style / special constraints
2. Background Information: project background / construction goals / industry context
3. Reference Materials (most important):

- A. Local file path or paste directly - B. Feishu document (RAG search) - C. Paste content directly - D. Not provided for now

1. Outline Confirmation: After planner outputs outline, user chooses A.Start / B.Adjust / C.Cancel

More reference materials → more business-aligned content. See references/phase0_guide.md for details.

Phase 1: Planner

Input: Phase 0 topic / background / reference materials

Execute:

python integrate_report.py glossary

Auto-generates plan.json + plan_outline_snapshot.md

Full prompt template in references/phase1_guide.md

After completion, send WeChat notification (using message tool, channel=openclaw-weixin):

📋 Report Outline Generated

📌 《[Report Topic]》
📊 Chapters: [X] chapters
🔍 Industry: [Industry Field]

✅ Reply "start writing" once the outline is confirmed — the system will launch parallel creation!

Phase 2: Parallel Sub-Agents

Execution flow (fully automatic, no manual confirmation):

1. Display outline / current batch status (display only, no waiting)
2. python parallel_tracker.py clear to clear previous batch state
3. Start up to 5 concurrent sub-agents (sessions_spawn), automatically execute all batches
4. python parallel_tracker.py wait to monitor in background until this batch is complete
5. After completion, automatically run python integrate_report.py convert-batch

Sub-Agent prompt template: see references/phase2_guide.md

After each batch completes, send WeChat notification (using message tool, channel=openclaw-weixin):

✅ Batch [X] Chapter Writing Complete!

📖 Completed: [Done]/[Total] chapters
📝 This batch:
   • [Chapter 1 Title]
   • [Chapter 2 Title]
   • [Chapter 3 Title] (if any)

⏳ Next batch: [Next batch chapter list]
(Automatically proceeds to next batch, no manual confirmation needed)

• Small change: directly edit F:/agent/chapters/0X-xxx.txt, save and regenerate
• Large change: re-trigger sub-agent to rewrite, replacing the original file

Phase 2.5: Cross-Chapter Consistency Review

python integrate_report.py check

Review numerical indicator consistency and terminology uniformity (对照 glossary.json)

After review completes, send WeChat notification (using message tool, channel=openclaw-weixin):

🔍 Consistency Review Complete

✅ Terminology uniformity: OK
✅ Numerical indicators: consistent
✅ Cross-chapter references: no conflicts

📄 Proceeding to final integration phase...

Phase 3: Integrator Summary

python integrate_report.py

Auto-completes: parse chapters (error isolation) → update glossary → consistency review → generate styled docx

After final completion, send WeChat notification (using message tool, channel=openclaw-weixin):

🎉🎉🎉 Report Writing Complete! 🎉🎉🎉

📄 《[Report Topic]》
📊 Scale: [X] chapters / ~[Y] thousand characters
🎨 Cover Style: [Style Name]

✅ Styled report generated!
📁 File location: F:/agent/chapters/output/

Wenxin, full text ready for your review~

Document Beautification Features (Auto-Applied)

Generated reports automatically include the following formatting effects (selected via cover_style field in plan.json):

1. 6 Cover Styles — Edit plan.json → cover_style field (integer 1~6)
2. Executive Summary — Deep blue title bar (#1F4E79) background + white text + body indent
3. Table-Style TOC — Deep blue title bar + three-column entries (number/chapter/page)
4. Colorful Chapter Headings:

- H1: Full-row deep navy background #1F4E79 + white text Microsoft YaHei - H2: Medium blue background #2E75B6 + white text - H3: Light blue background #D6E4F0 + dark blue text + ▌ left bar

1. Key Callout Boxes — Auto-detect 【关键】【注意】【优势】【风险】【数据】 tags, render as color cards (background/white text/border)
2. Styled Tables — Header deep navy background #1F4E79 + white text + alternating row colors (#DEEAF6 / #FFFFFF)

Cover Styles (6 Types)

Cover style specified via cover_style field in plan.json (integer, 1~6):

Note: cover_style value is integer (e.g., 4), code automatically converts to string for comparison.

CLI Command Reference

Switching cover style: Edit cover_style field (integer 1~6) in F:/agent/chapters/plan.json, then regenerate. After modifying code: delete .pyc files under __pycache__ + content_hashes.json to force rebuild.

State Files

Critical Rules

Markdown Table Format (Sub-Agents Must Follow)

See references/table_format_guide.md for full spec

Key points:

• Separator row must be |---|---|---| (leading/trailing | required)
• All rows must have same column count as header
• Cell content should avoid containing | (use ～ or - for ranges)

Force Rebuild (Must Do Both Steps After Code Changes)

After modifying integrate_report.py, must delete both files for new code to take effect:

# 1. Delete .pyc cache (required after code changes)
del "C:\Users\Administrator\AppData\Roaming\LobsterAI\SKILLs\long-doc-agent\__pycache__\integrate_report.cpython-311.pyc"

# 2. Delete incremental hash (or incremental mode skips everything)
del F:\agent\chapters\content_hashes.json

# 3. Regenerate
python integrate_report.py

Known Bugs Fixed (For Reference)

See references/bug_fix_guide.md, including:

• _flush_table not calling _parse_md_table causing character-by-character table splitting
• cover_style integer vs string comparison preventing cover from applying
• eval RGB color assignment type error
• .pyc cache causing new code to not take effect
• RGBColor using index access rgb[0]/rgb[1]/rgb[2] instead of .red/.green/.blue
• add_cover() setting section.margin=0 causing body text to have no margins
• PermissionError when docx file is open in WPS → auto-add _v2 suffix
• write tool has 50KB line limit → large scripts must be written in chunks

References