Introduction

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” In the field of AI-Driven Drug Discovery (AIDD), purely conversational agents are often insufficient to meet the rigorous workflows, traceability, and data governance requirements of scientific research. dsh-aidd is an AIDD research workbench built for this purpose. It introduces a Claude Science-style research engine and toolset to DSH.

Plugin Overview

dsh-aidd is a research workbench plugin tailored for AIDD scenarios and maintained by developer panda_aidd. Built on DeepSeek Harness, it provides a standardized set of scientific workflow tools. Its core value is converting unstructured research conversations into structured research loops while ensuring reproducibility of experimental data and models.

Core Features

The plugin mainly consists of three core components:

  1. ReAct Research Loop Engine
    It provides a set of tools to drive the research workflow: research_init (initialization), research_state (state check), research_hypothesis (hypothesis generation), research_experiment (experiment execution), research_findings (findings recording), as well as research_review and research_report. These states are persistently stored in research-manifest.json, recording the complete path from problem to conclusion.

  2. Versioned Artifacts with Provenance
    It manages experimental outputs through the artifact_* series of tools (such as artifact_save, artifact_list, and artifact_verify). All artifacts are archived by version under artifacts/<name>/v<N>/, with each file including a SHA-256 checksum and accompanied by artifact.json provenance information (including commands, inputs, environment configuration, etc.) and a provenance.md log.

  3. 17 Research Skills
    The plugin has 17 built-in skills, divided into general skills (such as literature linking, parallel delegation, and manuscript drafting) and AIDD-specific skills (such as virtual screening, molecular generation, property prediction, DTI modeling, and model reproduction). These skills are automatically discovered from project-level .dsh/skills/ directories or a global installation directory.

Installation and Enablement

There are two installation methods; choose based on your use case.

Method 1: Profile Bundle (Community Standard)

Install the plugin as a Bundle into the current Profile. All Agents under that Profile can then use the related tools.

dsh plugin --profile web add dsh-aidd

Method 2: Agent Preset (Full AIDD Mode)

Clone the repository and install the Preset. Select the “AIDD Mode” preset in the DSH Web interface for a full experience that includes the research persona and data standards guardrails.

git clone https://github.com/panda_aidd/dsh-aidd ~/.dsh/.agent-presets/aidd
bash scripts/install.sh

Usage Example

After installation, the standard AIDD research workflow is as follows:

  1. Initialization: Call research_init. The system creates research-manifest.json and a project skeleton directory (including experiments/, artifacts/, analyses/, data/, etc.).
  2. State Check: At the start of each session, call research_state first to confirm current research progress.
  3. Execution Loop:
    • Use research_hypothesis to define a hypothesis (for example, “This sub-library contains hit compounds with scores exceeding the threshold.”).
    • Use research_experiment to design and execute a virtual screening experiment, generating experiment records.
    • After running the experiment, use research_findings to analyze enrichment and diversity, and update the conclusions.
    • Use artifact_save to save the hit list and scoring table.
  4. Modeling and Reproduction: Use the ml-experiment-tracking and model-reproduction skills for model training and reproduction. Key conclusions must be reviewed by scientific-reviewer and archived.

Repository and Validation

This plugin is forked from biociao/dsh-science v0.1.1. The engine layer replicates the upstream implementation, while the domain layer is rewritten for AIDD. The plugin has zero dependencies and uses only built-in Node.js modules.

The project provides automated testing scripts to verify logical correctness and concurrency stability:

node scripts/smoke-test.mjs
node scripts/stability-test.mjs

Notes

  1. Data Standards: The plugin includes built-in AIDD data discipline guardrails, including fixing random seeds, archiving data split files as artifacts, and ensuring that the training and test sets do not contain homologous targets or shared scaffolds. Performance improvements must not be claimed before the baseline has been reproduced.
  2. Installation Path: When using Agent Preset mode, ensure that the preset/engines/ directory stays synchronized with the preset configuration.
  3. Version Requirement: The runtime environment must meet Node.js version >= 18.

Summary

dsh-aidd introduces a structured scientific research workflow to the DeepSeek Harness ecosystem, making it suitable for research scenarios that involve molecular data, model reproduction, and experiment records. Through versioned artifacts and the ReAct loop engine, it addresses common issues in AIDD development such as lax processes and untraceable data.

Repository: https://github.com/panda_aidd/dsh-aidd