// the find
ExplosiveCoderflome/AI-Novel-Writing-Assistant
面向长篇小说创作的 AI Native 开源系统，用 Agent、世界观、写法引擎、RAG 和整本生产工作流，帮助新手从一句灵感走到完整小说。AI-native engine for end-to-end novel creation — from idea to full chapters, with structured planning, worldbuilding, and agent-driven workflows.
A TypeScript monorepo (React + Express + LangGraph) aimed at end-to-end long-form novel generation — not a chat assistant, but a structured pipeline that takes a one-line idea through worldbuilding, character prep, volume planning, chapter execution, and quality repair. The target user is someone who wants to write a 100k-word novel without prior craft knowledge, with AI handling structure and scheduling decisions. Primarily aimed at the Chinese web-novel market.
LangGraph is a reasonable choice for the multi-stage agentic workflow — checkpoint recovery, approval nodes, and state re-injection are all things LangGraph handles well rather than requiring custom orchestration. The 'fact ledger' for extracting hard facts from generated chapters (character names, transaction amounts, dates) and feeding them into the next chapter's context is a genuinely useful mechanism for fighting the consistency problem in long-form generation. Separating concerns into distinct pipeline phases (worldbuilding → volume strategy → chapter planning → execution → quality repair) means each stage can fail independently and be retried, which is the right architecture for a system where LLM calls are expensive and unreliable. The style engine — extracting writing features from reference text and compiling them into reusable style assets — is more principled than just copy-pasting a style description into a system prompt.
SQLite as the primary datastore is a significant bet: the moment two users run concurrent novel production pipelines, you hit SQLite's write serialization limits, and there's no obvious migration path to Postgres without touching every Prisma query. The README describes an impressive feature surface but the actual pipeline reliability for completing a full novel end-to-end is unclear — the roadmap explicitly admits P0 is 'reduce false stops and abnormal token consumption', which suggests the happy path frequently fails in practice. Almost all documentation and UI is in Chinese with some English README translation, which means non-Chinese contributors face a steep barrier and the issue tracker will be effectively inaccessible to most of the GitHub audience. The AGPL-3.0 + CLA dual licensing with a commercial carve-out for SaaS use is a classic open-core trap: the CLA gives the maintainer full relicensing rights while contributors keep AGPL restrictions, which will deter serious open-source contributors.