XN-289/dsh-NINGLET-novel-Agent

NINGLET is a native novel-writing plugin built specifically for the DeepSeek Harness ecosystem, and its single stated purpose is to let AI write fiction that does not read as machine-generated. Rather than acting as a black-box prompt that turns an outline into prose, it exposes a controllable production line composed of book creation, chapter writing (planning, orchestration, drafting, anti-AI-flavor audit, revision, settlement), disk persistence, and a chapter panel for review. Integration happens through four shipped artefacts: a preset/agent.cordis.yml preset, five SKILL.md skill files, a Host domain tool (plugins/host-novel.js), and a Client chapter panel (plugins/client-novel-ui.js), all activated inside DSH via cordis_define / cordis_run on the live ning-1/pkg-4 dynamic-plugin channel. Authoritative state is written under the session workspace at novels/<bookId>/ as both a validated JSON file and human-readable Markdown projections.

The intended workflow is conversational and agent-driven: a user simply asks the assistant to "create an urban-cultivation novel called 吞天魔帝", which seeds a book, generates an 8-to-12 chapter outline, and applies the 25 universal craft rules plus the matching genre-specific rule set (xuanhuan / urban / mystery / romance / sci-fi / history). Asking the assistant to "write the next chapter, focusing on the master-disciple conflict" launches the chapter pipeline; if no intent has been recorded, a Socratic three-question gate (core advancement, protagonist state, closing hook) fires first. During drafting, the 12-dimension anti-AI engine continuously scans for forbidden words, template phrases, AI-style transitions, 的/地 density, sentence-length variance, uniform paragraph length, three-beat parallelisms, sentimental paragraph endings, repeated dialogue tags, cliché density, formulaic pivots, and list-shaped structure, and any single hit automatically triggers the full four-stage rewrite (point-removal, structural repair, stylistic rewrite, human-voice injection). A separate novel-qa skill performs 10-dimension consistency checks and outputs a 0-to-100 AI-flavor score. The audience is long-form web-fiction writers and studios who want continuity across hundreds of chapters without manually polishing every awkward sentence.

Dependencies and operational caveats: NINGLET requires Node.js to execute the 99 pure-function unit tests via node --test, which are all green, plus a DeepSeek Harness instance capable of loading dynamic plugins through cordis_define / cordis_run. The license is MIT, so the code can be freely reused and modified, and the harness-packages/ directory already contains source for a future standard DSH distribution. First-run notes: story/state/state.json is the single source of truth and a schema validator refuses to persist malformed state, so any migration or reset must deliberately preserve that file; the control-plane documents (author_intent.md, current_focus.md, book_rules.md) are auto-generated when a book is created and refreshed after each chapter, but authors are free to edit them to override the default genre rule sets.