dsh-science: A Research Workbench Plugin for DeepSeek Harness, Managing Research Loops and Artifact Provenance

dsh-science is a research workbench plugin (DSH plugin) for DeepSeek Harness (DSH), designed for genomics, pathogens, human health, and bioinformatics projects. It addresses the problems of scattered experimental records, hard-to-reproduce results, and missing data provenance in scientific workflows, providing a complete closed-loop research management capability. Core features include a ReAct-based research loop engine (research_* tools), versioned artifact management with provenance (artifact_* tools), and 10 science skills (e.g., research-loop, literature-connector, manuscript-writing, bioinformatics-toolkit). Maintained by biociao under the MIT license, last updated in 2026-08, current version v0.1.1 focuses on concurrency safety and error handling.

What research loop tools does the dsh-science DSH plugin provide?

The research loop engine implements a complete research state machine: Question → Hypothesis → Experiment → Observe → Analyze → Conclude → Next Question, driven by tools such as research_init, research_state, research_hypothesis, research_experiment, research_findings, research_phase, research_review, and research_report. All state is persisted in research-manifest.json, supporting a hypothesis state machine (proposed → testing → supported/refuted/inconclusive) and forward-only phase transitions (rewind requires configuration).

How does dsh-science manage versioned artifacts with provenance?

Artifact management tools include artifact_save, artifact_list, artifact_show, artifact_diff, artifact_verify, artifact_deprecate, and artifact_reproduce. Each result is saved under artifacts/<name>/v<N>/ with per-file SHA-256 hashes, artifact.json provenance (command, inputs, environment, envFile), and an append-only provenance.md. v0.1.1 adds streaming SHA-256 for large files, identical-content deduplication via hardlinks, upgraded artifact_diff/artifact_verify/artifact_deprecate, and envFile + input hashes in provenance.

What science skills are included in dsh-science?

The plugin bundles 10 science skills: research-loop, science-project-setup, artifact-provenance, scientific-reviewer, literature-connector, parallel-delegation, manuscript-writing, bioinformatics-toolkit, conda-environments, and data-inventory. These cover the full research workflow from project setup, literature review, experiment execution, to manuscript writing, helping researchers organize their work efficiently.

How to install dsh-science?

There are two installation options: as a profile bundle (community standard) or as an agent preset (science mode). The recommended profile bundle method uses the following command:

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

After installation, restart the profile or refresh the Web GUI, and the research_* and artifact_* tools become available to all agents on that profile. For the full science mode experience, clone the repository to ~/.dsh/.agent-presets/science and run bash scripts/install.sh.

What are the concurrency safety and error handling features of dsh-science?

v0.1.1 hardens concurrency safety: all manifest and artifact writes go through a lightweight file lock (O_EXCL + stale reclaim) and atomic tmp+rename, so parallel subagents cannot corrupt or lose updates on research-manifest.json / artifacts.json. Error codes are structured (e.g., ERR_NOT_INIT, ERR_NOT_FOUND, ERR_VALIDATION, etc.), and it supports manifest schema v1→v2 migration, structured JSON outputs (research_report, artifact_diff, artifact_verify), and an audit log at <root>/.science.log.

View more DSH plugins at dsh-plugin.org