Introduction

DeepSeek Harness supports extending Agent workflows through plugins. Mathematical modeling competitions usually require team collaboration, involving three stages: modeling, programming, and paper writing. The dsh-math-team plugin package integrates role division and quality gates into two presets and provides a team collaboration protocol, aiming to provide a standardized execution framework for multi-person collaboration.

Core Features

Role Presets

The plugin provides two role Agent presets, corresponding to different competition roles:

  • model-code: For modeling and programming members. Includes capabilities such as problem analysis, model construction, Python/MATLAB solving, result and chart production, and reproducibility checklists.
  • paper: For paper-writing members. Includes capabilities such as paper writing (Word/LaTeX), evidence retrieval and verification, and independent review quality gates.

Team Collaboration

Supports multi-folder collaboration through Gitee or GitHub. The presets include independent workspace protocols for member-a, member-b, and member-c. Each member commits files in their own independent workspace without interfering with the others.

Quality Gates

The plugin defines a complete set of quality checkpoints:

  • M1: Modeling final check
  • P1: Minimal runnable final check
  • M2: Robustness attack final check
  • P2: Programming final check
  • W1: Evidence outline final check
  • W2: Paper final check

M2 Robustness Attack Final Check

After P1 passes, a systematic “attack” is performed on each core model. This includes proposing attack points, verifying with alternative methods, introducing errors, out-of-sample validation, and clearly defining applicability boundaries, preventing model conclusions from not withstanding scrutiny.

Independent Model Review

Requires using two different models (e.g., Kimi K3-256K and DeepSeek V4.1 Flash) to independently review the content. Review dimensions include factual accuracy, consistency, completeness, model attacks, and citation verification, helping cover blind spots through cross-review.

Image Analysis Subagent

Provides a mandatory image review gate. All formal figures must be reviewed and recorded by the image analysis subagent before P2 or W2. If the main model does not support image input, the subagent automatically detects and uses an image-capable model (e.g., kimi-for-coding) for review.

Installation and Configuration

Environment Requirements

Requires DeepSeek Harness version 0.2.0 or higher. Starting with 0.2.0, the ~/.dsh/.agent-presets directory is no longer read; presets must be declared by a plugin package.

Installation Command

Reference this repository as a local path using the link: method. For Windows, for example, run the following command in the desktop profile directory:

& 'F:\dsh\DeepSeek Harness.exe' --expose-internals 'F:\dsh\resources\app.asar\dsh\node_modules\@deepseek-ai\dsh-desktop-host\lib\cli.js' plugin --profile desktop add "link:<本仓库绝对路径>"

After installation, restart the desktop client. The preset selector should show “Modeling and Programming Agent” and “Paper Agent”.

Typical Usage

3-Person Team Collaboration

  1. Members A/B: Clone the repository in their respective independent workspaces and enable the model-code preset. Produce modeling and programming deliverables into their own member-a or member-b folders, then push to the remote repository.
  2. Member C: Enable the paper preset and pull the deliverables from A and B. Write the paper in the member-c folder, then push to the remote repository.

Mandatory Image Review

After enabling any role preset, the image analysis subagent must be invoked to review formal figures before the P2 or W2 checkpoint. Only figures with a PASS review record can pass the final check.

Notes

  1. Version dependency: You must use DSH version 0.2.0 or higher.
  2. Installation method: You must reference the local absolute path of this repository using link:. Moving or deleting the directory will invalidate the presets.
  3. Compatibility: The ~/.dsh/.agent-presets directory used by 0.1.x versions is no longer read in 0.2.0. Old configurations need to be reinstalled.