Introduction

Common pain points when working on genomics, pathogen, or bioinformatics projects within DeepSeek Harness (DSH) include: research processes scattered across conversation logs, experimental results lacking version control and traceability, and the need to connect to laboratory workstations or HPC clusters for long-running bioinformatics tasks. General-purpose agent plugins often only cover code editing or document retrieval, making it difficult to string together the “question—hypothesis—experiment—conclusion” research cycle along with remote compute orchestration into a reproducible workflow.

Here we introduce the community plugin dsh-science (maintained by biociao). It packages the ReAct research cycle, versioned artifact provenance, SSH remote computing, and 11 scientific skills into DSH, positioning itself close to Claude Science’s research workbench, targeting genomics / pathogen / human health / bioinformatics scenarios. It currently has about 27 stars and 2 forks on GitHub; the latest release is 0.2.0, MIT licensed, and requires Node.js >= 18.

What It Is

dsh-science is a research-oriented plugin bundle in the DSH ecosystem. It can be installed as a profile bundle or enabled via the agent preset “Science Mode.”

The plugin provides four engines with zero third-party dependencies (Node.js built-in modules + system OpenSSH) within its repository, registering standard cordis tools. The accompanying dsh-model-tier model tier routing is also zero-dependency and is automatically included when installing dsh-science. It solves three core problem types:

  1. Organizing research conversations into a persistable ReAct research cycle.
  2. Storing experimental results as versioned artifacts with SHA-256 hashing and provenance.
  3. Submitting, monitoring, and retrieving long-running bioinformatics jobs on remote workstations or SLURM clusters via SSH.

Core Features

ReAct Research Cycle Engine

Provides tools like research_init, research_state, research_hypothesis, research_experiment, research_findings, research_phase, research_review, research_report, etc. State is persisted in research-manifest.json, progressing through “Question → Hypothesis → Experiment → Observation → Analysis → Conclusion → Next Question.” The hypothesis state machine progresses: proposed → testing → supported / refuted / inconclusive. Phases default to forward progression only.

Versioned Artifacts & Provenance

Provides tools like artifact_save, artifact_list, artifact_show, artifact_diff, artifact_verify, artifact_deprecate, artifact_reproduce, etc. Each saved result resides in artifacts/<name>/v<N>/, including per-file SHA-256 hashes, an artifact.json (command, inputs, environment, envFile), and an append-only provenance.md.

Remote Compute Engine (SSH / HPC)

Provides 16 remote_* tools. Connects to laboratory workstations or HPC clusters via ~/.ssh/config aliases, with no additional components required on the remote end.

  • Host Registration: remote_host_add registers an alias, remote_host_probe detects CPU, memory, GPU, CUDA, conda, SLURM partitions, etc. The registry is located at $DSH_HOME/remotes/hosts.json.
  • Job Submission: remote_run uploads scripts and inputs to <scratch>/<jobId>/ (default ~/dsh-scratch). Workstation jobs run as detached processes using nohup+setsid; SLURM cluster jobs use sbatch. Default timeout is 30 minutes; approval is required by default before submission.
  • Monitoring & Retrieval: remote_status batch-probes and automatically migrates states (running → succeeded / failed / killed); remote_logs views logs; remote_pull retrieves outputs and writes pulled-manifest.json (files > 100 MB remain on the host with their path recorded); remote_cancel cancels jobs. The job registry is located at <project_root>/.dsh/remotes/jobs.json.
  • Project-level Allowlist: First-time connection to a host requires project-level authorization. Once approved, it’s written to .dsh/remotes/allowlist.json, isolated per project.

Remote Host Configuration UI

Adds a “Remote Hosts” page to the DSH settings panel, corresponding to Claude Science’s Settings > Compute > SSH hosts. The host-side REST API route is /dsh-science/remote-hosts/*. The client bundle shares the same data files with the remote engine. Changes require restarting the web process to take effect.

Model Tier Routing (dsh-model-tier)

Starting from v0.2.0, the companion bundle dsh-model-tier provides cross-provider tiered routing: auxiliary requests (session titles, compaction) and subtasks use the lightweight tier; the main dialogue retains the model selection for each session; deep-chain subtasks and very long inputs can be upgraded to the strong tier. It can also be installed independently into any profile:

dsh plugin add dsh-model-tier

Configuration example:

- id: model-tier
  name: dsh-model-tier
  config:
    tiers:
      strong: { provider: zai-coding-cn, model: glm-5.3 }
      default: { provider: deepseek-official, model: deepseek-v4-flash }
      light: { provider: opencode-go, model: minimax-m2.7 }
    routing:
      auxiliary: [session-title, compaction]
      subagents: light
      subagentDepthStrong: 3

When tiers are not configured or the target provider is not registered, routing defaults to no-op, not affecting the main flow.

11 Scientific Skills

research-loop, science-project-setup, artifact-provenance, scientific-reviewer, literature-connector, parallel-delegation, manuscript-writing, bioinformatics-toolkit, conda-environments, data-inventory, remote-compute. Skills can be auto-discovered from <project_root>/.dsh/skills/ or installed globally to ~/.dsh/skills.

Installation & Enabling

The DSH ecosystem follows “everything is a plugin.” The community directory SkillHub is an independent site with no official affiliation with DeepSeek / High-Flyer. Before installing, it’s recommended to read the GitHub repository source code and MIT license, confirming the plugin runs with the current dsh process permissions.

Method A: Profile Bundle (Community Standard)

Inserts the research cycle, artifact registry, remote computing, remote host UI, and model tier routing into a profile with a single command. All agents in that profile can use the research_*, artifact_*, remote_* tools.

dsh plugin --profile web add dsh-science

If not published on npm or needing to track the latest code, install directly from GitHub:

dsh plugin --profile web add "github:biociao/dsh-science"

Restart the profile after installation (or refresh the Web GUI; new engines or modified engine code require a profile restart).

Method B: Agent Preset (“Science Mode”)

Suitable for scenarios requiring a complete scientific persona, isolated per agent:

git clone https://github.com/biociao/dsh-science ~/.dsh/.agent-presets/science

Or use the repository’s install script:

bash scripts/install.sh

Then create a new session in DSH Web and select the Science Mode preset.

Skill Installation

bash scripts/install-skills.sh

Skills are installed to ~/.dsh/skills (following $DSH_HOME).

Typical Usage

Below is a reproducible initial session flow, excerpted from the official README.

  1. Call research_init to create research-manifest.json and the project skeleton (experiments/, literature/, artifacts/, analyses/, figures/, manuscript/, reviews/, data/, envs/).
  2. Start each session with research_state to persist the research cycle state across sessions.
  3. Run the research cycle: research_hypothesis (H1/H2/…) → research_experiment (E01/…, creating experiments/<id>/{design.md,log.md,code/,results/}) → execute code → research_findings (append to log.md, update hypothesis state) → call artifact_save for results worth citing or reproducing.
  4. When GPU or cluster is needed: remote_host_add to register a host → remote_run to submit a remote job (requires approval) → periodically use remote_status to monitor and remote_logs to view logs → after completion, remote_pull to retrieve outputs → artifact_save to archive.
  5. For key assertions: extract the assertion → sub-agent review cross-references execution logs (see scientific-reviewer skill) → research_review archives to reviews/R0n/report.md.

Remote computing details are documented in the repository docs/remote-compute.md and the remote-compute skill.

Applicable Scenarios & Notes

Who It’s For:

  • Researchers working on genomics, pathogen, human health, or bioinformatics projects in DSH.
  • Teams needing to structure and consolidate hypotheses, experiments, and conclusions, rather than just keeping conversation logs.
  • Users who already have ~/.ssh/config configured to connect to laboratory workstations or SLURM clusters and want to submit and monitor long-running jobs directly within the agent.

Usage Notes:

  • The plugin runs with the current dsh process permissions. Operations like remote_run will actually connect to remote hosts and submit jobs. Review the source code and license before installation.
  • Remote job submissions and first-time host accesses require approval by default. Large files (> 100 MB) remain on the remote end during pull, with only the path recorded.
  • Choosing between bundle and preset: The bundle installs with one command and provides profile-level shared tools. The preset offers a complete “Science Mode” experience isolated per agent. Profile-level persona configuration is commented out by default in cordis.patch.yml and must be uncommented to apply to the entire profile.
  • Remote host UI and new engines require restarting the web process to take effect.

Conclusion

dsh-science integrates the ReAct research cycle, artifact provenance, SSH remote computing, and model tier routing into DSH, providing a reproducible research workbench for genomics and bioinformatics scenarios. If you’re already using DSH for research assistance, you can start with a profile bundle installation and gradually enable remote computing and skills as your project needs evolve.