dsh-research-suite

A DeepSeek Harness (DSH) plugin for the full clinical-research workflow — manuscript writing, language polishing, figures, statistics, research ideation, peer review, slides, literature search and journal formatting. Install it and it takes over research tasks automatically: no UI, no manual skill picking.

中文完整说明见 README.zh-CN.md（内容更全，含配置项、目录结构、占位符清单）。

• Package: dsh-research-suite · Version: 3.2.0 · License: MIT · DSH tested: 0.2.1-rc.5, 0.2.2-rc.1
• Requires Node ^22.19.0 || >=24.0.0, DSH >=0.2.0-0 <0.3.0
• Self-contained: all 450 source skill documents plus their bundled references/scripts/templates assets ship inside the package; nothing is read from any external skill directory at runtime.

Why capability cards instead of a skill router

The hard part of research work is not a lack of skills — it is that three to eight skills can each do the same task well in different ways. A router hands the model one "primary skill" name; the model then reads that skill's manual and improvises, losing the other skills' strengths along the way.

Instead, each capability card takes every competing skill in its group, records best_at (what it is best at), limits (what it explicitly cannot do) and take (the concrete rules, thresholds and templates lifted from it), then fuses them into one executable protocol. Every protocol step says which skill it comes from and why. The acceptance rule is strict: after reading one capability card you can finish the job without opening any original skill document.

The 450 source documents (and the 954 asset files they bring along) remain bundled as a deep-reading layer — for checking details, not for improvising.

How auto-takeover works

No UI, no buttons. Three pieces:

1. A system-prompt section — lib/prompt.js builds a policy named research-suite/policy at order: 650 ("after team policy, before tool docs"); lib/index.js registers it through systemPrompt.section({ name, order, text: () => buildPolicy(core) }). The policy tells the model: recognise a research task → call research_playbook first → follow the capability protocol → deep-read if needed → run research_gate before delivering.
2. 11 skills (skills/): one entry skill plus ten domain SOPs (writing / polish / figures / stats / ideation / review / slides / literature / format / data). Their content is inlined at registration time — passing a lazy load() hook fails DSH's runtime validation and silently drops every skill (the exact failure mode of v1).
3. 15 tools (7 query + 8 execution/verification/generation/memory) for capability lookup, searching the 450-document catalog, deep reading (including skill-bundled assets), standards, and mechanical self-checks.

When research_playbook finds no match, the policy falls back to research_catalog + research_kb_read rather than refusing the task.

Domains

Domain	Source skills	Covers
writing	64	paper types and sections (original / review / case report / grant)
polish	12	register, de-AI-flavour, reduction, proofreading, Word reports
figures	33	data plots, statistical charts, mechanism diagrams, layout
stats	30	descriptive, comparison, regression, survival, prediction, PSM, MR, Meta, missing data, sample size
ideation	37	topic selection, novelty search, hypotheses, grants, ethics
review	20	self-review, reviewer comments, revision, rebuttal, response letters
slides	37	conference, lecture, defence, lab meeting, poster
literature	26	search strings, screening, close reading, evidence tables, recall/precision
format	51	journal formatting, references, figure/table specs, submission package
data	40	databases and omics data: GEO / TCGA / dbGaP / genotype and variant databases, omics retrieval and curation
Total	450	103 capability cards across the ten domains

toolbox is an additional index-only bucket of 100 utility skills: it feeds research_catalog and research_kb_read, and deliberately has no capability cards — a tool is a means, not a task a user asks for.

Every one of the 450 source skills has a recorded verdict in data/ledger/<domain>.json (fused — and which cards it feeds / index-only / deprecated / out-of-scope), so the "all studied, best only" claim is auditable: node scripts/verify-ledger.mjs all re-checks coverage and that every fused skill is actually a member of the card it claims. Beyond the text, skill-bundled assets ship too: 186 skills contribute 954 files (~6.8 MB) of references/, scripts/ and templates/ under knowledge/skills/<domain>/<slug>__assets/; research_kb_read(name, asset='scripts/plot.py') reads them (omitting asset lists what the skill ships with). Nine reference standards live in references/.

Tools

Tool	Purpose
research_playbook	The only entry point. Pass the user's own words; returns the 1–3 best-matching capability cards with protocol, gates, deliverables and suggested deep reads
research_capability	Expand a full capability card (member comparison, protocol, decisions, gates, outputs, kb) by id or by domain
research_catalog	Search the 450-document catalog by name/description keywords (the 100 toolbox skills included), optionally filtered by domain; each hit reports how many bundled assets it carries
research_kb_read	Read one skill document or INDEX.md/FUSION.md; the asset parameter reads a skill-bundled file such as scripts/plot.py or references/*.md (omitting it lists the skill's assets); paging and path-escape rejection included
research_standard	List or read the nine reference standards in references/
research_gate	Gate and mechanical self-check: section = all / data / selfcheck / mutex / forbidden / hard / handoffs; text checks fullwidth punctuation, superscript citation numbers, AI-cliché phrases and absolute wording
research_suite_status	Load self-check: version, per-domain capability counts, 450-entry catalog completeness, asset statistics, references, package root
research_run	actually executes an R or Python script and returns stdout/stderr plus produced files; off by default — first calls return a preview only (see Security below)
research_data_check	mechanical half of the six data gates: rows/cols, missing rate, constant columns, duplicate IDs, impossible values, EPV<10
research_output_check	deliverable audit: real PNG DPI, PDF pages and text layer, DOCX/XLSX structure, CSV rows, MD word count
research_doi_check	reference authenticity: resolves each DOI against Crossref and returns title/journal/year; degrades honestly offline
research_ref_audit	citation reconciliation: [n] gaps/duplicates, mismatch with the reference list, missing DOIs, suspicious years, duplicate DOIs
research_project	project memory: read/write .dsh-research/project.json — target journal, glossary, finalized sections, decisions — surviving across sessions
research_docx	produce and audit .docx: build a Chinese-journal-formatted Word file (SimSun 12pt, 1.5 spacing, first-line indent, three-line tables, hanging reference indent, superscript citations) or audit an existing one
research_litsearch	real search: PubMed via esearch+esummary (PMID/journal/year/authors/DOI), or field-by-field DOI comparison against expected metadata

Install

Published on npm as dsh-research-suite@3.2.0.

# via plugin_manager (GUI) — recommended
plugin_manager(action="install_bundle", target="dsh-research-suite@3.2.0")

# via CLI (Windows)
dsh.cmd plugin --profile desktop add dsh-research-suite@3.2.0
# via CLI (POSIX)
dsh plugin --profile desktop add dsh-research-suite@3.2.0

# from a local checkout (plugin development)
plugin_manager(action="install_bundle", target="<absolute package path>")

# from GitHub, if npm is not reachable
dsh plugin --profile desktop add github:xuzhecomeon/dsh-research-suite

Fully quit and restart DSH Desktop afterwards (closing the window is not enough — the loader reads the bundle patch layer at startup). Verify with research_suite_status.

Configuration

The plugin deliberately does not export a Config: in v2.0.0 that export (toJSON() returning JSON Schema) was hydrated by DSH's Host settings service as a schemastery envelope, crashed on schema.meta, and prevented DSH Desktop from starting. Configuration is now validated inside apply() by validateConfig().

Key	Default	Meaning
autoTakeover	true	Inject the auto-takeover prompt section; false degrades the plugin to "15 tools + 11 skills" without automatic routing
sectionOrder	650	Order of the prompt section (after team policy, before tool docs)
verbose	false	Log warnings when skill/tool/prompt registration fails

Because there is no exported Config, the DSH settings panel shows no entries for this plugin (a deliberate trade-off). Override values in the profile's cordis.patch.yml entry instead:

- id: dsh-research-suite
  config:
    autoTakeover: true
    sectionOrder: 650
    verbose: false

Documentation

• docs/CAPABILITY-SPEC.md — capability-card content contract
• docs/ARCHITECTURE.md — modules, data flow, how auto-takeover is implemented, and the v3 knowledge layer & ledger
• docs/PUBLISH.md — npm release, DSH install paths, upgrade/rollback, troubleshooting table
• CHANGELOG.md — what changed in 3.0.0

Self-checks

npm test                          # scripts/selftest.mjs — fake-ctx load smoke test
npm run verify                    # scripts/verify-package.mjs — static release checks
npm run verify:caps               # scripts/verify-capabilities.mjs all
node scripts/verify-ledger.mjs all # ledger coverage (11 domains, 450 skills) + fused check
npm run pack:check                # npm pack --dry-run (~38 MB unpacked in v3; knowledge/ included by design)

Measured on the current v3 tree (final numbers follow once every domain lands): 1,481 files / 38.62 MB on disk (scripts/verify-package.mjs count, includes .github/); the bulk is knowledge/ — 450 source documents plus 954 bundled asset files.

Security and execution boundary

Execution is off by default. This plugin can genuinely run code (research_run), which is exactly why it ships a safety layer: scripts generated by the model are treated as untrusted input.

The flow is two-stage: preview → explicit confirmation with a matching previewHash → three gates (enabled flag / hash / static policy) → run.

• data/exec-policy.json ships with run.enabled: false. In that state research_run returns a preview only (first 60 lines, sha256 previewHash, policy summary, predicted violations) and never writes a file or spawns a process.
• Once enabled, scripts run with your user privileges inside <cwd>/.dsh-research/run/<runId>/. The static scanner enforces package allowlists (separate lists for R and Python), rejects the common network and subprocess APIs, confines writes to relative paths inside that directory, rejects PHI path keywords, and permanently blocks installers. This is a guardrail, not a sandbox — real isolation requires a container, a restricted account, or a VM.
• research_doi_check does make network requests: it sends each DOI to https://api.crossref.org and nothing else (no manuscript or local file content). No other tool performs network requests.
• The plugin is self-contained: 450 source skill documents and 954 bundled assets ship inside the package; at runtime it reads no external skill directory.
• The entry module never imports @deepseek-ai/*: those packages do not resolve from the profile's node_modules, and a single import would make the whole plugin fail to load (symptom: "installed, but zero skills"). scripts/verify-package.mjs enforces this.

Tool	Reads files	Writes files	Network	Subprocess	Blast radius
research_playbook / research_capability	yes (capability cards)	no	no	no	in-process, read-only
research_catalog / research_kb_read / research_standard	yes (index, skill docs & assets, references)	no	no	no	in-process, read-only, paths confined to the package
research_gate / research_ref_audit	no (analyses supplied text)	no	no	no	in-process text checks
research_data_check / research_output_check	yes (caller-supplied data/deliverable)	no	no	no	in-process, read-only
research_suite_status	yes (index, stats, policy)	no	no	where / which probes only	in-process + PATH probe
research_doi_check	no	no	yes (Crossref, DOI only)	no	sends DOI strings
research_run	yes (whatever the interpreter can read as your user)	yes (inside the run dir; out-of-bounds writes rejected)	no by default	yes (Rscript / python)	run dir + whatever your user account can reach

How to enable execution, the full policy schema, the three isolation layers, and the known bypass limits are documented in docs/SECURITY.md.

Honest capability boundary

An external expert review pointed out that this plugin used to describe its abilities far better than it could verify them. This section states both sides plainly.

What it actually does

1. Gated task routing — on a 50-utterance golden set: takeover decision 50/50 correct, all 5 non-research cases rejected, capability-card top-3 100%; reproducible via node scripts/eval-match.mjs --strict (also run in CI).
2. Real execution — research_run actually spawns R (Rscript) or Python in a scratch directory and returns stdout/stderr; the model cannot pretend it ran something.
3. Mechanical data gates — research_data_check: missing rates, constant columns, duplicate IDs, impossible values, EPV < 10.
4. Deliverable audit — research_output_check: real PNG DPI, PDF page count and text layer, DOCX/XLSX structure.
5. Reference authenticity — research_doi_check resolves DOIs against Crossref; research_ref_audit reconciles in-text [n] against the reference list.

What it does not do

1. No plagiarism or AI-text detection — both require uploading full texts to paid external services. Not provided, and not faked.
2. No final statistical judgement — tools cover mechanical checks only; model choice, confounding control and sensitivity analysis still follow the capability protocol with human judgement.
3. No full-text verification of claims — a DOI proves existence, not that the cited conclusion was read correctly.
4. No taste — the suite gives verifiable numbers (DPI, font size, palette), not "what an editor will like".
5. Limited golden set — the 50 utterances are self-authored; the mechanism generalises but the sample is small.
6. Host interface dependency — if the host exposes no tools / systemPrompt service, the plugin degrades instead of dying (skills and policy still load); check with research_suite_status or the "装载自检" line in the policy section.

License

MIT — see LICENSE.

Scope note: MIT covers the plugin itself (lib/, data/, skills/, references/, scripts/, docs/ and the index files). The 450 skill texts and 954 bundled assets under knowledge/skills/ are third-party material vendored verbatim from the author's local skill collection; their licenses belong to their original authors and are not covered by MIT. Read docs/ATTRIBUTION.md before redistributing.