academic-harness
================

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

Renamed in v0.16.0: this project was formerly `dsh-researchcraft` and displayed
as ResearchCraft. References to ResearchCraft below identify upstream content
and are retained for attribution.

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) is vendored as
`skills/modal-usage`, and since v0.25.0 its frontmatter `name` is also
`modal-usage`, so it does not collide with this plugin's own first-party
`modal` skill (specific to the bundled `modal_run` tool). The skill body is
otherwise unchanged; the file carries its own `Apache-2.0` license header as
published upstream, which takes precedence for that bundle.

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 `co-scientist-tournament` skill (`skills/co-scientist-tournament`), the
`coscientist` / `coscientist_rank` tools, and their backing modules
(`coscientist-models.js`, `coscientist-store.js`, `store-lock.js`) port the
algorithms and role prompts of the open-source `raktim-mondol/co-scientist`
project (MIT License, © 2025 Raktim Mondol) — itself an implementation of the
Gottweis et al. *Nature* (2026) paper cited above. No file from that repository
was copied: its TypeScript was rewritten as this plugin's own ES modules, and
its Bun / `bun:sqlite` / `sqlite-vec` / ONNX-embedding / MCP-OAuth / ink-TUI
stack was deliberately left behind.
What is carried over is the deterministic core — the Glicko-2 update and
review-score seeding, the citation-integrity penalty, both-order judge
debiasing, the scheduling weight tables and termination rules, matchup
selection, the save-time duplicate gate, the dual-use severity gate, and the
experimental-feedback reward mapping — with the upstream constants kept so a
run here stays comparable to a run there, plus subagent briefs written from its
prompt templates. Upstream code is MIT; the *Nature* paper it implements is
CC BY 4.0.

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 `paper2agent` skill (`skills/paper2agent`) is vendored from the
open-source `jmiao24/Paper2Agent` project (MIT License) —
https://github.com/jmiao24/Paper2Agent, commit
`8c2d059165ef8cdcb70dbea76655b9c2b55b38e6`. The skill tree (`paper2agent` router plus sibling `paper2skill`,
`paper2mcp`, and `paper2agent-paper` catalog skills, and Python helpers)
is copied; a ResearchCraft mapping was added at the top of the three
component `SKILL.md` files and to `paper2mcp/references/orchestration.md` and
`output-delivery.md` (DSH `subagent`/`subagentId`, `uv`, `paper_download`,
`paper_mcp`, no `claude mcp add`). Components are top-level skills because
DSH's skill catalog is one directory deep (nested `paper2mcp` was
`skill`-unknown). Codex UI metadata (`paper2mcp/agents/openai.yaml`) and
the Claude-only `scripts/connect_remote_mcp.sh` were not copied. Python
verification helpers are unmodified.

Academic Harness (`vendor/academic-harness/` plus `skills/academic-harness/`
and `skills/ah-*`) is the Python engine from
`raktim-mondol/academic-harness` (commit
`dbc7414662c776b55e9da87c798d1ad08bf1c941`), vendored without the Pi
adapter, Claude Code adapter, examples, phase reports, or tests. ResearchCraft
drives that engine through native DeepSeek Harness tools and interceptors
(`academic-harness.js`, `academic-harness-hooks.js`). The optional T3
verifier backend that can shell out to a `pi` binary is unused unless a
project sets `verifier.backend: pi`. Skills were rewritten for DSH tool
names; the engine Python is unmodified.

open-deepreview (`vendor/open-deepreview/`) is vendored from
`raktim-mondol/open-deepreview` (MIT License) —
https://github.com/raktim-mondol/open-deepreview — at commit
`c48e29450134d7a41559a73112962377f9602bce`. The TypeScript source under
`src/` is unmodified. The plugin adds only a Node runner, a TypeScript loader
hook, and `SOURCE.txt`, and does not expose the Bun-only web viewer, the
development fixture command, or upstream tests. See
`vendor/open-deepreview/SOURCE.txt`.

PaperMemory (`vendor/papermemory/` plus `skills/papermemory/SKILL.md`) is
vendored from the open-source `raktim-mondol/papermemory` project (MIT
License) — https://github.com/raktim-mondol/papermemory. The Python package
(CLI, SQLite store) is unmodified except `mcp_server.py`: DSH's MCP SDK
speaks newline-delimited JSON on stdio, while upstream only spoke LSP
Content-Length headers, so the handshake deadlocked and tools never
registered. The vendored server now detects the inbound framing and replies
in kind. PDF/markdown ingest also accepts `--doi` / `--pdf-path` so a
downloaded PDF and its `pdf_to_markdown` output merge onto one verified
PaperMemory record instead of a second unverified local-pdf row. The skill body was rewritten for this plugin's MCP tool names
(`mcp__papermemory__*`), `paper_download` auto-ingest, and the ResearchCraft
preset; citation-integrity rules are unchanged. Tests from that repository
were not copied.

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.

Citation formatting uses the open-source `citation-js` packages
(`@citation-js/core`, `@citation-js/plugin-bibtex`, `@citation-js/plugin-csl`,
`@citation-js/plugin-ris`, all MIT License) as npm dependencies, not vendored
code. Mermaid rendering (`mermaid_render`) likewise uses `pele` (MIT License,
https://github.com/obsidianmd/pele) as an npm dependency, unmodified; the only
code this plugin adds around it is CSS-variable flattening for file output.
The CSL style files under `styles/` (`ieee.csl`,
`chicago-author-date.csl`, `modern-language-association.csl`) are unmodified
files from the Citation Style Language styles repository,
https://github.com/citation-style-language/styles, retrieved 2026-10-23 and
licensed under CC BY-SA 3.0 (https://creativecommons.org/licenses/by-sa/3.0/).
See `styles/README.md`.

The `remove-ai-signs` skill under `skills/remove-ai-signs` is adapted from the
local skill of the same name. Its reference guide, `references/signs.md`, is
adapted from Wikipedia's "Signs of AI writing" (WP:AICAT) and is licensed
CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). It is a prose
style-editing aid; it makes no detector-evasion or provenance-removal claim.
A Supervisor-Skills routing note pointing to `paper-polish` was added; that
modification remains under CC BY-SA 4.0.

The reply-folding engine in `fold-core.js` is adapted from TextFold's
`plugins/textfold/hooks/fold.ts` (Apache License 2.0), part of
https://github.com/chenxiachan/textfold — version 0.10.31, commit 2182d9a,
© 2026 Xia Chen. Changes for this plugin: string level names, structured blocks
for the browser panel, removal of Pi-specific drawing-row detection, and
DSH-oriented tests. The deletion-only fold, sentence splitter, word-count gate,
and CJK reading-effort accounting are retained. TextFold is not vendored; its
Python package is not used by `fold_text`.

The preset-only `paper_fold` tool uses TextFold as an optional runtime
dependency: on first preset start it installs `textfold==0.10.31` from PyPI into
`$DSH_HOME/textfold` and calls the documented `get --json` CLI. The package is
not vendored, and no TextFold source is copied for this tool; `TEXTFOLD_PYTHON`,
`TEXTFOLD_CLI`, and `TEXTFOLD_SKIP_INSTALL` let a user supply or skip it.

The 12 Supervisor-Skills research-lifecycle skills under `skills/`
(benchmark-paper-template, deep-research, drawio-reconstruction,
figure-designer, idea-evaluator, intro-drafter, paper-polish, paper-writer,
pre-submission-reviewer, rebuttal-guidance, tech-paper-template, and
vibe-research-workflow) are vendored from
https://github.com/HKUSTDial/Supervisor-Skills at commit
`a20d76ef3979b804d7dbf599d577826957e0a8a9` (2026-10-15, upstream v2.1.0).
Their licences are mixed:

- **CC BY 4.0**: benchmark-paper-template, figure-designer, idea-evaluator,
  pre-submission-reviewer, tech-paper-template, vibe-research-workflow.
- **CC BY-NC-SA 4.0**: deep-research, intro-drafter, paper-polish, paper-writer,
  rebuttal-guidance. These are non-commercial and share-alike; they remain
  under that licence inside this package, and this package's MIT licence does
  not relicense them. The full licence text is in
  `vendor/supervisor-skills/LICENSE`.
- **MIT**: drawio-reconstruction (Copyright (c) 2026 VCG-Bench Authors; its
  `LICENSE` is kept in the skill directory).

License texts: https://creativecommons.org/licenses/by/4.0/ and
https://creativecommons.org/licenses/by-nc-sa/4.0/; the latter is also included
in `vendor/supervisor-skills/LICENSE`. The suite is © 2026 Yuyu Luo
(methodology) and contributors, except drawio-reconstruction
(© 2026 VCG-Bench Authors).

The upstream `handbook/` and `handbook-en/` curricula, the repository and
per-skill READMEs, and drawio-reconstruction's examples/assets/agent metadata
were not copied; the skills are self-contained, and handbook references point
upstream. Local changes: "(separate plugin)" cross-references were removed;
`license:` frontmatter was added to drawio-reconstruction and
rebuttal-guidance; short DSH runtime notes were added to drawio-reconstruction
(use `uv run python`) and vibe-research-workflow (this agent is the runtime);
and every skill gained a `Routing in Academic Harness` section that names the
existing skill to use when a task crosses a stage boundary. Reciprocal
routing notes were added to the plugin's `academic-harness`, `ah-write-unit`,
`ah-plan-section`, `scientific-figure-making`, `diagram-design`,
`manuscript-editor`, `peer-reviewer`, `literature-researcher`,
`literature-review`, `surveying-prior-work`,
`generating-scientific-hypotheses`, `agentic-data-science-pipeline`,
`remove-ai-signs`, `abstract-writer`, and `systematic-review-ledger` skills;
`seed.js` refreshes those known-stale copies on existing installs.
See `vendor/supervisor-skills/SOURCE.txt` for the exact file list and changes.

Deep research pipeline design

The `deep_research` tool's pipeline shape — step-per-phase skills, the frozen
canonical query, tiers and gears, independence clustering, patch-only editing
with per-hunk caps, the ship gate, and untrusted-source fencing — was adapted
from `jordan-gibbs/hyperresearch` (MIT License), a Claude Code/Codex research
harness. No upstream code is vendored: the implementation is first-party
JavaScript (`deep-research.js`, `deep-research-store.js`,
`deep-research-steps.js`, `independence-cluster.js`, `untrusted-fence.js`)
written against this plugin's existing retrieval, ledger, and lock helpers, and
the step set, gate semantics, tiers, and verification seams differ from the
upstream design.

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