dsh-researchcraft
==================

This plugin packages ResearchCraft's research persona, scientific skills
catalogue, lab notebook, and specialist subagent briefs as a DeepSeek Harness
(DSH) profile bundle.

Scientific skills are seeded from the open-source catalogue
`K-Dense-AI/scientific-agent-skills` (MIT License) — the full catalogue
(140 skills) is vendored under `skills/` alongside this plugin's own
first-party skills. Its `modal` skill (general Modal usage) was renamed to
`modal-usage` here to avoid colliding with this plugin's own `modal` skill
(specific to the bundled `modal_run` tool); no other content was changed.

Research-discipline skills (framing-research-questions, preregistering-analysis,
verifying-results-before-claiming, requesting-red-team-review, and 12 more) are
vendored under `skills/` from the open-source `K-Dense-AI/science-superpowers`
catalogue (MIT License), unmodified. That project's Claude Code
`SessionStart` hook, which force-injects its `using-science-superpowers`
skill into every session, was not ported — DSH has no equivalent hook
mechanism — and its guidance was folded into this plugin's own system prompt
instead (see the "Research discipline" section of `prompt.js`).

503 profession-specific expert-reasoning skills (accelerator-physicist,
zoologist, marine-biologist, and so on) are vendored under `skills/` from the
open-source `K-Dense-AI/scientific-agents` catalogue (MIT License). Each was
mechanically converted from that repo's `agents/<slug>.md` (Claude Code
subagent format) into `SKILL.md` by adding a `license`/`metadata` block to the
existing `name`/`description` frontmatter; the body content is unchanged.

The `docx-editor-zotero` skill (`skills/docx-editor-zotero`) is adapted from
the `extensions/docx-editor-zotero/` extension in the open-source
`K-Dense-AI/claude-scientific-writer` repository (MIT License): its
`docx-editor-zotero.md` became `SKILL.md` (frontmatter added; content
otherwise unchanged), and its two Python scripts and `EXAMPLES.md` were
copied into `scripts/` and `references/` — with one fix:
`zotero_docx_preserver.py` had a malformed f-string (a stray `}` in a
`.get(...)` call) that failed to parse under Python 3.12+; corrected to match
the identical pattern used everywhere else in that file. That repository's 26 other
skills were not vendored — every one of them duplicates a skill already
vendored here from `K-Dense-AI/scientific-agent-skills` above.

23 small-molecule and protein-therapeutics skills (autodock-vina,
binding-site-analysis, retrosynthesis, protein-binder-design, and 19 more)
are vendored under `skills/` from the open-source
`K-Dense-AI/drug-discovery-agent-skills` catalogue (MIT License), unmodified.
That catalogue's other 14 skills (rdkit, datamol, medchem, pytdc, esm, and 9
more) were not vendored — this plugin already has a same-named skill for
each from `K-Dense-AI/scientific-agent-skills` above; the two catalogues
wrote independent, differently-detailed versions of those particular skills,
and rather than pick a winner or maintain both under different names, the
already-vendored version was kept as-is.

The `agentic-data-science-pipeline` skill (`skills/agentic-data-science-pipeline`)
adapts the plan → implement → verify → reflect agent architecture from the
open-source `K-Dense-AI/agentic-data-scientist` project (MIT License): its
Google ADK / Claude Agent SDK orchestration code was not ported (it
reimplements what this plugin's `subagent`/`todo`/`notebook` tools already
do), but its 11 per-role prompt files (`prompts/base/*.md`) and its two
bioinformatics domain-specialization notes (`prompts/domain/bioinformatics/`)
were copied into `references/` largely verbatim as detailed role briefs,
alongside a new `SKILL.md` mapping the loop onto this plugin's own tools.
One near-duplicate file (`plan_generator.md`, an alternate draft of
`plan_maker.md`) was dropped to avoid two conflicting versions of the same
role.

The `hyperparameter-optimization` skill (`skills/hyperparameter-optimization`)
is a fresh adaptation of the agentic-ML-engineer loop described in the
open-source `K-Dense-AI/karpathy` project (MIT License) — no file from that
repository was copied. Its Python package (Google ADK + Claude Agent SDK,
a `delegate_task` tool) reimplements this plugin's own `subagent` delegation
and was not ported; only the workflow idea (resource-aware, pre-registered,
verify-before-claiming hyperparameter search) was carried over, rewritten
for this plugin's own tools and narrowed to the hyperparameter-tuning use
case.

The `generating-scientific-hypotheses` skill (`skills/generating-scientific-hypotheses`)
is a first-party procedure distilled from two *Nature* papers (no paper text
was copied): Gottweis et al., “Accelerating scientific discovery with
Co-Scientist”, *Nature* (2026), https://doi.org/10.1038/s41586-026-10644-y
(CC BY 4.0); and Ghareeb et al., “A multi-agent system for automating
scientific discovery”, *Nature* (2026),
https://doi.org/10.1038/s41586-026-10652-y (CC BY-NC-ND 4.0). The
evaluation axes, literature-before-claiming-novelty, and
mechanism→assay→candidate shape were rewritten for this plugin’s own
tools. The existing `hypothesis-generator` specialist brief points at that
skill; it is not a second copy of the procedure.

The `scientific-figure-making` skill (`skills/scientific-figure-making`) is a
fresh adaptation of the publication-figure workflow described in Chen Liu's
`figures4papers` repository (CC BY-NC 4.0) —
https://github.com/ChenLiu-1996/figures4papers. No file from that repository
was copied (its license is non-commercial). The house-style conventions
(semantic palette, spine/legend/export policy, grouped-bar and trend
patterns) were rewritten for this plugin's tools, and the helper API that
the upstream skill specifies but does not ship is implemented here as
`scripts/publication_figure.py`.

The `diagram-design` skill (`skills/diagram-design`) is vendored from the
open-source `cathrynlavery/diagram-design` project (MIT License). The
example-gallery HTML (`example-*.html`, `assets/index.html`) was not copied
— templates, type references, import/export scripts, and the self-check are
included, along with that project's `LICENSE` and `THIRD_PARTY_LICENSES.md`
(Tabler Icons MIT, Simple Icons CC0). A short ResearchCraft mapping was
added at the top of `SKILL.md` (paper-first PNG/SVG export, skip website
brand onboarding unless asked, `uv run python`); the rest of the skill is
unmodified.

The `python-helpers/` scientific-file-format decoders (chemistry, structure,
mass spec, arrays, imaging, AnnData) and the `workflows.json` research-task
template catalogue are adapted from the ResearchCraft codebase (MIT), a
white-labeled fork of K-Dense BYOK — see
https://github.com/K-Dense-AI/k-dense-byok.

ResearchCraft branding and content © ResearchCraft.
Site: https://researchcraft.dev