A deep context-aware agent-based novel creation system featuring Writer/Editor/Archivist collaboration, BM25 fact tracking, and token budget management, built as a local-first desktop writing workstation for 100k-word long-form and fan fiction.
WenShape (formerly NOVIX Writing) is a local-first desktop application for long-form novel creation, built on a multi-agent collaboration architecture. The system drives a five-stage writing pipeline through three specialized agents: Writer (draft generation), Editor (revision review), and Archivist (scene preparation and fact management): scene preparation → context construction → draft generation → revision loop → wrap-up analysis.
For long-form consistency, WenShape implements a comprehensive context engine: facts are stored in canon/facts.jsonl and selectively injected using BM25 + keyword overlap hybrid scoring + entity enhancement + chapter binding + logarithmic chapter distance decay. Precise token budget management allocates explicit quotas to rules, cards, facts, summaries, and drafts, effectively reducing long-form hallucination.
Setting management relies on three core card types—character cards, world-building cards, and style cards—all maintained in YAML format for searchability and reusability. Structured volume-chapter management persists chapter ordering via order_index. The fan fiction workflow supports searching and page scraping from sources like Moegirl Encyclopedia, Wikipedia, and Fandom, with a proposal mechanism for quickly importing character and world-building settings.
The backend is built on FastAPI (Python) with a Vite + React (JSX) frontend. Data is stored in YAML/Markdown/JSONL plain text formats, naturally compatible with Git version control. The llm_gateway abstraction layer supports OpenAI, Anthropic, DeepSeek, Gemini, Qwen, Wenxin, AI Studio, and custom OpenAI-compatible endpoints, with an LLM Profile mechanism that allows assigning different models to different agents. Installation options include a Windows one-click package and cross-platform source deployment. The system works fully offline, with internet connectivity used only for enhancement features.
Installation & Quick Start
Prerequisites: Python 3.10+, Node.js 18+
One-click start (recommended):
cd WenShape-main
python start.py
Automatically checks the environment, launches frontend and backend services, and initializes base configuration on first run. A Windows one-click package is also available from the Releases page.
Default addresses: Backend API http://localhost:8000, Swagger docs http://localhost:8000/docs, Frontend UI http://localhost:3000.
Core Architecture
WenShape-main/
├── start.py / start.bat / start.sh
├── frontend/src/
│ ├── pages/ # WritingSession.jsx as main workspace
│ ├── components/ hooks/ context/
│ ├── lib/ utils/
│ └── i18n/
└── backend/app/
├── routers/ # HTTP / WebSocket routes
├── orchestrator/ # Multi-agent orchestration
├── agents/ # Writer / Editor / Archivist
├── context_engine/ # Context selection & budget sorting
├── llm_gateway/ # Model provider adaptation
├── services/ # Summary, evidence, scraping services
├── storage/ # YAML / Markdown / JSONL storage
└── prompt_templates/
Project data resides in data/{project_id}/, containing project.yaml, cards/, drafts/, summaries/, canon/ (facts.jsonl), traces/, etc.
Unconfirmed Information: Repository owner is unitagain, but no explicit team or individual background is listed; only Windows one-click package is currently available; no associated academic papers or HuggingFace model pages found; no standardized writing quality benchmarks provided.