About this project
dsh-science-workbench is a plugin for the DeepSeek Harness that provides a reproducible science workbench environment. It combines Jupyter-like cells with agent-driven execution and Nextflow-style provenance tracking, ensuring every figure and artifact is traceable and replayable.
Key features include:
- Code-to-figure workflow with inline display and structured feedback for iterative redrawing (v1 → v2 → v3).
- A plain-text manifest.json as the single source of truth for cells, artifacts, provenance, and feedback.
- Reproducibility by construction: self-contained scripts, fresh subprocess per cell, environment.lock, SHA-256 hashes, and fixed seeds.
- Automatic git versioning for each project (local commits only).
- Cross-platform support: bash on macOS/Linux, PowerShell on Windows.
Nine agent-facing tools are provided: bio_init_project, bio_run_cell, bio_rerun_cell, bio_add_feedback, bio_get_project, bio_list_projects, bio_set_projects_dir, bio_delete_cell, and bio_mark_cell. A browser-based workbench UI offers a three-panel layout with inline figure preview, searchable project picker, cell search, and provenance details.
The plugin also bundles two publication-grade figure skills (figure-style and figure-composer) adapted from Claude Science.
Installation is via the standard dsh plugin command, and the plugin automatically registers itself in the profile bundles. After restarting dsh web, the bio_* tools and workbench tab become available.
Each project follows a structured layout with manifest.json, environment.lock, index.md, code/, data/, figures/, and .git/. The architecture separates Host (execution and provenance) from Client (browser UI), communicating via same-origin fetch routes.
Development is straightforward with npm run lint for syntax checks and local plugin installation for testing.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.