DSH (DeepSeek Harness) embraces the philosophy of “everything is a plugin.” In mathematical research, maintaining research state, reading literature, accumulating tool experience, and performing formal verification often involve cumbersome configuration. math-research-dsh one-way adapts the parent plugin math-research into four independently usable skills, aiming to help researchers complete literature reading, experience accumulation, task continuation, and Lean verification in the DSH environment.

Feature Overview

This plugin decomposes the research workflow into four core capabilities. Researchers or Agents can choose the corresponding entry point based on the specific problem:

  • math-research-workflow
    • Select research tools and maintain current progress.
    • Continue sessions and actual tasks.
  • manage-math-research-program
    • Read literature and manage tool cards and annotations.
    • Compare experiences and maintain human-editable understanding pages.
    • Manage accepted knowledge in the Blueprint.
  • rigorous-open-math-research
    • Develop proofs or counterexamples.
    • Explain successful and failed research routes.
    • Audit arguments as needed by the problem.
  • lean-verify
    • Provide compile feedback.
    • Precise goals and transitive axiom checks.
    • Perform semantic review and generate reproducible evidence.

Installation and Enabling

There are two installation methods; choose one.

Use DSH’s plugin registration feature to directly install the bundle:

dsh plugin --profile web add github:xsoc1/math-research-dsh

This command registers the package.json, cordis.patch.yml, and index.mjs files in the package. DSH will use FileSystemSkillProvider to load skills.

2. Local Development and Junction Installation

If local debugging is required, you can clone the repository locally and run the installation script:

if($env:DSH_HOME)
{
    $DshHomePath = $env::DSH_HOME
}
else
{
    $DshHomePath = Join-Path $HOME '.dsh'
}
$DshRepoPath = Join-Path $DshHomePath 'math-research-dsh'
git clone https://github.com/xsoc1/math-research-dsh.git $DshRepoPath
powershell -ExecutionPolicy Bypass -File (Join-Path $DshRepoPath 'install.ps1')
python (Join-Path $DshRepoPath 'scripts/dsh-doctor.py')

install.ps1 links the four directories to the user skills root directory. Replacing plain directory copies requires explicitly selecting the -Force parameter.

Typical Usage

After installation, load a specific skill using DSH’s skill tool:

dsh skill math-research-workflow

On the first line of a user message, you can also use the /skill-name syntax to directly activate a specific skill.

Architecture and Maintenance

The repository design follows a one-way sync principle:

  • The parent repository is the sole content source: The research methods and mathematical implementations in the math-research parent plugin are the only content source.
  • Adaptation layer: This repository only maintains the DSH loading method, layout mapping, packaging, and test adaptation.
  • Sync process: Upstream changes must first enter the parent repository, then be replayed through scripts. The synchronizer copies the four skills, merges scripts/ and assets/, and places runtime/ at the root of the corresponding skill.

During maintenance, it is recommended to run the following check commands:

python scripts/sync-from-parent.py --upstream <clean-parent-clone> --expect-commit <full-parent-commit> --check
python scripts/validate_all.py .
python scripts/dsh-check-bundle.py

Testing

The project retains 21 root smoke tests covering DSH adaptation, continuation, literature management, Blueprint and Git operations, Lean compatibility, and legacy pipeline compatibility.

Plugin tests within the package include:
* skills/manage-math-research-program/scripts/tests/: research experience, research library Q9 reuse, and so on.
* skills/lean-verify/scripts/tests/: verifier logic and real Lean compilation tests.

Notes

  • The plugin runs with the permissions of the current DSH process. Check the source code and license before installation.
  • Legacy sealed checkpoints and pipelines still have compatibility tools, but legacy full validation requires explicitly using the --legacy-v1 parameter.
  • Historical experiment stages, roles, and quota rules in old records belong only to the corresponding old records and do not apply to the current 2.0 version.
  • Directory page: https://www.skillhub.cn/plugins/xsoc1/math-research-dsh
  • GitHub repository: https://github.com/xsoc1/math-research-dsh