DeepSeek Harness (DSH) is an agent development framework. When running quantized small models locally, context-window and reasoning-capability limitations can make it difficult to handle long-form novels, which require multi-stage collaboration. The dsh-novel-solo plugin addresses this scenario by drastically streamlining the toolchain and building in a complete writing and review workflow.

This is a single-author novel-writing plugin designed for DeepSeek Harness and maintained by Tkingxiao. By pruning unnecessary tool calls and enforcing a strictly defined writing workflow, it enables quantized small models to reliably complete the full pipeline from outline to finished book on a local machine.

Core Features

1. Toolchain Streamlining and Quantization Safety
The plugin hard-disables tool categories such as shell, jobs, skills, goals, and web, retaining only file read/write, search, and basic subagent capabilities. The output protocol is restricted to pure Chinese and common punctuation, avoiding JSON formatting and escape characters, ensuring stable and parseable text on quantized models.

2. Complete Writing and Review Presets
The presets include a zero-to-one novel generation pipeline, covering: narrative method → core worldview → term index → outline → chapter catalog → character profiles → chapter-by-chapter writing → chapter-by-chapter review → final full-book review and compilation.

3. Multi-Dimensional Review Mechanism
Each chapter undergoes review across seven dimensions (setting, characters, catalog, narrative, text rules, AI clichés, and plot logic). Writing proceeds to the next chapter only after all dimensions pass.

4. AI Tone Elimination
The review stage includes AI-tone detection, covering six dimensions (syntactic templates, clichéd vocabulary, emotional telling, structural similarity, information dilution, and character distortion). It distinguishes narrator-level repetition from character-level inconsistency and is applied consistently in both the writing and review stages.

5. Versioned Writing
To prevent overwriting, each chapter’s output file is named 第N章-章节名字-vX.md. When rewriting, a new version is written first, while the old version is retained as a historical draft.

Installation and Activation

Use the official installation command to add the plugin:

dsh plugin --profile web add "dsh-novel-solo"
dsh web

After installation, the plugin generates a configuration card on the DSH plugin page.

  • DSH 0.1.7: The card is located under “Plugins” in the left sidebar.
  • DSH 0.1.6: The card is located under “Settings → Plugins”.

Configure the “Subagent count (1-12)” parameter on the card to determine the number of subagents working in parallel. Note that GUI changes take effect only after the host is restarted.

Typical Usage

  1. Install the plugin and restart the DSH host.
  2. Open the plugin page and set a value for “Subagent count” (for example, 3).
  3. Select the novel-solo preset in DSH.
  4. The preset automatically runs the complete pipeline described above, from worldbuilding to final book compilation.
  5. Drafts and review reports for each chapter are saved as Markdown files in the project directory.

Use Cases and Notes

This plugin is designed for use cases where quantized small models (such as lightweight versions of Qwen, Llama, and others) are run locally to process long-text tasks.

Notes:
* Configuration Change Delay: The concurrency setting in the GUI must be written into the prompt by restarting the DSH host.
* Host Version Differences: DSH 0.1.6 and 0.1.7+ differ in preset deployment methods. The plugin automatically adapts, but if you have manually modified presets, pay attention to version compatibility.
* Configuration Key Changes: For older host versions below DSH 0.1.5, the text key in the persona configuration has been removed and should be changed to the prefix key.
* Source Code and License: The plugin is open source under the MIT license. It is recommended to review the source code before use to ensure it meets local runtime requirements.

dsh-novel-solo turns local small models from simple conversational tools into execution engines capable of driving long-form narrative pipelines. For the project source and detailed documentation, see GitHub or the SkillHub Catalog.