Introduction

DeepSeek Harness (DSH) extends agent capabilities through a plugin mechanism. When tool libraries are provided to an agent, if a dependency is missing, the agent often does not report the missing dependency; instead, it may fabricate a client or file path and produce a plausible result.

The dsh-biomni plugin prevents this “false availability”. It runs as an external plugin for DeepSeek Harness and does not modify Harness itself. The core principle is: everything told to the agent (functions, datasets, tools) is validated by environment checks, and only content that the current machine can actually deliver is provided.

Core Features

This plugin provides the following capabilities to the agent:

  1. Stateful Python interpreter
    One process per session. Imported data, dataframes, and fitted models persist across multiple calls, allowing the agent to execute scripts step by step instead of resending them each time. If the interpreter is idle for more than 30 minutes, it will be reclaimed to free memory, and the agent will receive a notification instead of encountering a NameError.

  2. 21 research function skills
    Based on Biomni’s 218 functions, the functions are split into 21 skills by module. Each skill carries real function signatures (parameter types, default values, and meanings). Skills are loaded on demand, saving context tokens.

  3. Data lake skill
    Displays locally downloaded biomedical datasets, including exact paths, sizes, and license flags. It prioritizes local data and reduces network requests.

  4. Software skill
    Reports bioinformatics packages and command-line tools actually installed on the current machine, enabling the agent to directly invoke tools such as samtools instead of reimplementing them.

  5. Honest environment reporting
    Provides the /biomni command and a settings page that accurately shows which capabilities are available, which are missing, and which packages are needed to repair them.

  6. Result output channels
    run_python returns text. The interpreter is bound to the output directory BIOMNI_OUT. Files written to that directory are read by the model; charts are displayed inline, CSV files are displayed as tables, and file downloads are supported.

  7. Data lake browser
    The settings page lists all datasets and their sizes, marking which ones are on disk. Explicit acknowledgment is required before using non-commercial datasets.

Installation and Activation

1. Add the Plugin

Run the following command in the terminal to add the plugin to the configuration:

dsh plugin --profile web add dsh-biomni

2. Build the Python Environment

The plugin does not automatically install Biomni. The data lake and software catalog are available immediately, but the 21 tool skills require Biomni.

git clone https://github.com/ZhuYanTech/dsh-biomni && cd dsh-biomni
bash scripts/setup-env.sh

This script prefers uv to install dependencies (about 11 seconds) and falls back to pip if uv is not installed. After installation, the script runs probe checks, and the probe declares success only when Biomni can actually be imported successfully.

3. Configure the Interpreter Path

After starting DSH, go to Settings → Biomni and set the Python interpreter to the virtual environment path created in Step 2 (for example, /abs/path/to/.venv/bin/python).

Alternatively, configure it manually in $DSH_HOME/settings.yaml:

biomni:
  python: /abs/path/to/.venv/bin/python
  idleTimeoutMs: 1800000  # 30分钟闲置后回收
  guardShellPython: true

4. Verify the Installation

/biomni

This command lists what the interpreter can run and the remediation options for any missing items.

Typical Usage

  • /biomni: View available skills and environment status.
  • /biomni-datasets: View local dataset details.
  • /biomni-out: View output files generated during work.
  • pnpm run install:preset: Install the preset biomedical researcher persona configuration.

Notes

  • Memory reclamation: The interpreter is reset after 30 minutes of idleness to free memory.
  • Data licensing: Explicit acknowledgment is required when using non-commercial datasets.
  • Non-invasive: This plugin does not fork or patch Harness.
  • Probe mechanism: The probe refuses to claim success if importing Biomni fails, ensuring that the environment description remains accurate.

Summary

dsh-biomni is a DeepSeek Harness plugin maintained by ZhuYanTech. By using a persistent Python interpreter and fine-grained skill splitting, it provides the agent with a reliable biomedical research toolchain. All capabilities are grounded in the actual environment, helping prevent agent hallucinations caused by missing dependencies.

GitHub repository: https://github.com/ZhuYanTech/dsh-biomni