Introduction

DeepSeek Harness (DSH) novel-writing plugin. The design principle is singular: code enforcement > prompt self-discipline. Any constraint that relies on “model self-discipline” will fail—consistency, gating, and auditing must be hard constraints at the tool/data layer. Through 20 novel_* tools and the right sidebar “Forge” panel, the plugin translates issues in long-form web novel writing such as setting collapse, context loss, filler, and AI tone into code-level hard constraints.

Core Features

The plugin converts common long-form writing problems into concrete engineering constraints:

  • Setting and Consistency Management: Provides a fact ledger (novel_ledger) and a world book (novel_worldbook); conflicting value changes within the same chapter are rejected directly. Provides language voice cards (novel_character voice) to solve the “many characters, one voice” problem. Uses pre-writing briefings (novel_briefing) to inject context on demand and prevent long-context explosion.
  • Workflow and Gate Control: Uses a nine-phase state machine (novel_project phase) to enforce the order of settings, outline, chapter outline, and main text; phase gate (novel_outline approve) requires chapter outline approval before writing; three consecutive rejections of the same chapter trigger a circuit breaker and refuse writing.
  • Quality Control: Provides six-dimensional structural scanning (novel_noai_scan) for de-AI-tone checks; provides a machine review threshold (default 2000–4000 words) and 8-character shingle repeat detection to prevent cross-chapter repetition; provides end-of-chapter hook heuristics and chapter outline contract coverage checks.
  • Review and Compliance: Supports platform review (novel_audit platform, supporting Qidian/Tomato) and sensitive self-checks in seven categories (novel_audit censor); separates machine review from model review, and reviews output deterministic numbers with cited evidence; revisions follow a proposal workflow (novel_propose), and the main text is overwritten only after the user confirms in the panel.

Installation and Activation

Prerequisite: DeepSeek Harness (DSH) is already installed and the model is running; Node ≥ 22 (22.19+ / 24+ can enable retrieval indexing, otherwise it will automatically degrade).

Install:

dsh plugin --profile web add npm:dsh-novel-forge

Activate: After installation, you must restart the DSH web process and perform a hard browser refresh. The client bundle is snapshotted when the process starts, so not restarting is the same as not installing. After restarting, an independent “🔨 Forge” tab will appear in the right sidebar, and the model in the session should be able to see the 20 novel_* tools.

Typical Usage

  1. Create a book: Call novel_project init in the conversation to create a project.
  2. View the phase: Use novel_project phase to view the nine-phase board, and complete the missing items listed by the code (for example, premise first, then settings).
  3. Write a chapter: First save and approve the chapter outline with novel_outline save_chapter and approve, then run novel_briefing to get the context package, and finally use novel_write_chapter to write the chapter and pass the gates.
  4. Review: Use novel_noai_scan to scan for AI tone, and use novel_audit for platform review or sensitive self-checks.
  5. Revise: Use novel_propose propose to submit a revision proposal, then apply it after confirming in the right-side panel.

Panel and Tools

  • Forge Panel: The right sidebar provides project list, whole-book health check, batch drafting (including resume from interruption draft-batch resume), proposal approval, chapter listening, and other features.
  • 20 tools: Core tools include novel_project (phase management), novel_outline (outline and approval), novel_character (characters and language voice), novel_worldbook (world book), novel_scene (scene contracts and hidden characters), novel_write_chapter (chapter writing and gating), novel_ledger (fact ledger), novel_search (local retrieval index), etc.

Notes

  • Pure local and zero cost: Scanning and diagnostic features run fully locally and do not consume model tokens.
  • Revision mechanism: All revisions go through proposals; older versions are never overwritten and must become effective only after confirmation in the panel.
  • Repository: Public GitHub repository (SSH source requires a configured SSH key), npm version from 0.16.0.

  • Compatibility: The code-enforcement logic depends on DSH process permissions; it is recommended to check the source code and license before installation.

  • GitHub repository