Introduction

DSH’s plugin mechanism allows workflows, presets, and skills to be synced locally. Mathematical modeling competition tasks usually span multiple stages: reading the problem, framing, derivation, coding, verification, writing the paper, and reviewing. A single round of prompts tends to break the workflow into parts.

dsh-ultramath is a multi-agent solving plugin tailored for mathematical modeling competitions. It is maintained by Andiii208 under the MIT license, with the goal of completing the aforementioned workflow using native DeepSeek Harness agent presets and skills.

What is this

Below introduces what problems it solves and what capabilities it provides.

dsh-ultramath packages award-winning mathematical modeling methodologies into DeepSeek Harness native agent presets + skills. It covers:

Reading the Problem → Framing → Derivation → Coding → Verification → Paper → Review

It provides the UltraMath master preset, as well as 4 single-stage role presets: Mathematician, Engineer, Writer, and Reviewer. The plugin is purely Node.js-based, has no browser UI, and does not modify DSH core source code.

Core Features

Workflow and Roles

One preset can run through the entire workflow: “Reading the Problem → Framing → Derivation → Coding → Verification → Paper → Review”. In addition to the master controller, the 4 single-stage role presets are used to start from specific stages:

  • Mathematician
  • Engineer
  • Writer
  • Reviewer

Model Library and Self-contained

The plugin comes with an embedded library of 33 models and distributes paper templates, review scripts, and the model library along with the package. It is completely self-contained, does not depend on any external repositories, and is ready to use upon installation.

Upon startup, the plugin idempotently syncs presets, the model library, and paper templates to ~/.dsh/, requiring no manual configuration.

The ultramath main entry skill is available, and the model library is located at:

~/.dsh/skills/ultramath/模型库/

The model library is loaded on demand; loading the entire library at once is prohibited.

Quality Gates

The workflow includes independent verification, blocking items, blind review, and academic integrity gates:

  • Phase 2.5 Independent Verification: Disagreement >30% is FAIL, 10-30% is SUSPECT, <10% is PASS.
  • Review phase blocking items B1-B9 have veto power.
  • Phase 4 Final Review assigns 3 judges who do not share context with each other for structured scoring; if the difference exceeds 20 points, evidence arbitration is triggered.
  • Academic Integrity 7 types of gates: Plagiarism / Data Fabrication / Fake Citations / Fake Images / Non-reproducible / Anonymity Violation / Unauthorized Authorship, any trigger results in a veto.

Digital Freezing is used to prevent inconsistencies between paper numbers and code outputs. 冻结数字.json (freeze numbers.json) allows every number in the paper to be traced back to CSV or script output; after changing code, grep detects stale references.

Progress and Method Selection

The entire workflow maintains 求解/进度.md (solution/progress.md) and is driven by todos, with a final output steps/progress for visualization.

Method selection requests will recommend and load the corresponding model library documentation based on 触发词索引.md (Trigger Word Index) and 方法选择决策树.md (Method Selection Decision Tree).

Installation and Enablement

First, confirm that the Node version meets the requirements:

^22.19.0 || >=24.0.0

Then execute the installation command:

dsh plugin --profile web add github:Andiii208/dsh-ultramath

After installation, restart DSH. When creating a new session, select the “UltraMath” master controller or the 4 single-stage roles in the preset selector.

Typical Usage

Full Automation

Create a new session, select the “UltraMath” master controller, paste the problem + data attachments directly, and run through without asking for anything in between.

Staged Execution

When executing only a specific step, you can use prompts to limit the scope:

只做问题 1 的建模和求解，先不写论文

已有推导和结果，直接帮我写论文并编译

已成稿，只跑评审收敛（多视角评审 → 修复 → 终审）

Manual Oversight

When you need to confirm key decisions, you can request to pause and choose:

用人工把关模式：每个关键决策点停下来给我编号选项确认，再继续

Method Selection

Just ask:

这类问题该用什么模型

The plugin will recommend and load the corresponding model library documentation based on 触发词索引.md + 方法选择决策树.md.

Resume Mid-process

Put the existing 求解/ (solution/) directory into the workspace, then say:

继续求解

The plugin will continue from the breakpoint based on 求解/进度.md (solution/progress.md).

Update

When updating to a new version, first execute the reinstall command, then restart DSH:

dsh plugin --profile web add github:Andiii208/dsh-ultramath

Scenarios and Notes

Suitable for DSH users who need to break down mathematical modeling competition solving into stages, require paper templates and review scripts, and want traceable results.

Notes:

  • The plugin runs with the current DSH process permissions; check the source code and MIT license before installing.
  • The model library is loaded on demand; loading the entire library at once is prohibited.
  • Digital freezing relies on 冻结数字.json (freeze numbers.json), CSV, and script output; pay attention to stale reference detection after changing code.
  • Academic integrity gates and blocking items will trigger a veto.

Conclusion

The value of dsh-ultramath lies in breaking down mathematical modeling competition solving into executable, verifiable, and traceable stages. The GitHub repository is as follows:

https://github.com/Andiii208/dsh-ultramath