Introduction

The core philosophy of DeepSeek Harness (DSH) is «everything is a plugin». In long-text production scenarios, a single model often struggles to balance evidence depth, logical rigor, and multi-perspective critique. Lunheng registers an on-demand agent skill, transforming long-form production into a 9-role pipeline with human checkpoints, aiming to address weak evidence chains and logical inconsistencies in long-form writing.

Plugin Positioning

Lunheng is a DeepSeek Harness (DSH) plugin package (v18.8.0), containing package.json, cordis.patch.yml, and lib/index.js. It is not used as a text generator, but as an orchestration skill that decomposes long-form deliverables into a pipeline of 9 independent roles (T1–T9).

Core Features

The plugin registers a skill that orchestrates tasks through DSH’s subagent invocation mechanism, with the following characteristics:

  1. 9-Role Pipeline:

    • T1–T3 (Retrieval): literature reconnaissance, data reconnaissance, case reconnaissance.
    • T4 (Analysis): generates an analysis outline.
    • T5 (Writing): generates a first draft.
    • T6 (Critique): provides a critical report.
    • T7 (Audit): independent audit, outputting an audit report and revision tasks.
    • T8 (Final Check): final finalization (executed by the main coordinator).
    • T9 (Review): peer review and AI-trace gate check.

  2. Triangulated Evidence Base: every claim must be mapped to literature [Lxx], data [Dxx], and case [Cxx] evidence.

  3. Independent Audit and Human Checkpoints:

    • The auditor (T7) only reports and does not edit, ensuring objectivity.
    • Four human checkpoints are set: topic confirmation, outline confirmation, first-hand background input, and final draft review.

  4. Read-only Tools and Security Mechanisms:

    • Provides two read-only tools: lunheng_m_gate (M-gate mechanical precheck) and lunheng_char_count (pure character count).
    • Mechanism File Write Protection: globally rejects write/edit tool calls targeting the skill package, preventing silent rewrites of pipeline rules. It can be bypassed via the environment variable LUNHENG_ALLOW_MECH_EDIT=1 or the configuration item allowMechanismEdit.

Installation and Configuration

For the installation process, please refer to the official documentation. It typically involves deploying the plugin package to the DSH environment.

  1. Installation Method: the materials mention that the installation verification documentation is located at docs/installation.md.
  2. Configuration Items: set the following parameters in the DSH configuration file (or environment variables):
    • quiet: silent mode control.
    • allowMechanismEdit: whether to allow mechanism file editing (protection is enabled by default).
    • scriptTimeoutMs: script timeout duration.
    • scriptMaxOutputBytes: maximum script output bytes.
  3. Environment Variables: the corresponding environment variables LUNHENG_QUIET / LUNHENG_ALLOW_MECH_EDIT for the above configuration items are also valid.

Typical Usage and Output

This plugin is suitable for generating content exceeding 2000 characters and usually requires a wait time of 1-3 hours.

  1. Input: topic, length requirements, and citation format.
  2. Process:
    • Parallel retrieval of literature, data, and cases (T1-T3).
    • Human confirms the outline (Phase 2.5).
    • Human provides first-hand background (Phase 3.5).
    • Critique and audit (T6, T7).
    • Revision (Phase 4.2).
    • Review and final check (T8, T9).
  3. Output:
    • Evidence Cards: literature cards [Lxx] (with A/B/C confidence), data cards [Dxx] (with conflicting data comparison), and case cards [Cxx].
    • Analysis Document: argumentation threads, claim-evidence mapping, and rebuttal plan.
    • Drafts and Reports: multiple draft versions, critical report (C1–C7), audit report (G0–G14), and peer review report.
    • Final Deliverables: final/定稿.md, charts, evidence package, and delivery notes.

Applicable Scenarios and Limitations

Applicable Scenarios:
* Content exceeding 2000 characters that must withstand scrutiny.
* Reliance on published evidence rather than opinions alone.
* Human involvement for outline confirmation and final draft review.

Explicit Limitations (cannot be completed autonomously):
* Not a Text Generator: it is a pipeline, does not directly generate text, but orchestrates agents to generate content.
* Data Sources: it can only collect published evidence; users must provide first-hand materials themselves.
* Statistical Analysis: it does not run SPSS/R/Python and can only cite results.
* Charts and Multimedia: it does not perform scraping/OCR/speech-to-text, has no built-in text-to-image (T2I), and has no raw image/video processing capability.
* Code Execution: it only runs whitelisted scripts; other code execution requires explicit approval.

Summary

Lunheng structures the long-form production process through strict 9-role division of labor and triangulated evidence validation. It combines independent audit with human checkpoints, making it suitable for scenarios requiring high-trust long-form content. Developers should review the source code and license (MIT), then deploy it according to their own permissions (especially code execution and file write permissions).

  • Repository URL: https://github.com/zuoyunlai/lunheng-article-pipeline-dsh
  • Plugin Directory: https://www.skillhub.cn/plugins/zuoyunlai/lunheng-article-pipeline-dsh