Research Mode for DeepSeek Harness
Evidence-first research workflow — local-first, file-based, and honest about what it can prove.
把科研还给具体的人:一个本地优先、证据优先的科研工作流。 它不是"会写论文的聊天框",而是把**产物(Artifact)和证据链(Provenance)**做成一等对象。
Why this exists
Most AI research tools optimise for answers. That is the wrong unit for research, because an answer cannot be re-checked, handed over, or challenged. This project optimises for the thing that can be: a content-addressed artifact with a provenance record that says where it came from, what was actually run, and what remains unknown.
It is a set of plugins for DeepSeek Harness (a plugin-based agent runtime), plus skills that encode research discipline rather than research marketing.
Designed for social-science work first (small samples, textual sources, local literature) — the first preset ships a Chinese social-science persona — but the evidence model is field-agnostic.
What it actually does
Ten bundles, 24 tools, no third-party runtime dependencies (Node builtins only).
| Bundle | What it provides |
|---|---|
research-host |
Artifact store with 7-face provenance, review records with per-round content signatures, .science archive export/import |
research-workspace |
Left-sidebar research workspace: project templates, Host project list, launch a session scoped to a project |
research-hub |
Single right-pane entry with four tabs: artifacts · provenance · reviews · local library |
research-connectors |
8 read-only connectors (OpenAlex, Crossref, DataCite, OSF/SocArXiv, PKU Dataverse, World Bank, ILOSTAT, NBS snapshot) |
research-phase2-litdb |
Read-only search over a local Chinese social-science sample database (FTS5 with LIKE fallback) |
research-phase2-notebook |
Read-only .ipynb inspection: cells, execution counts, imports, reproducibility gaps |
research-phase2-replay |
Read-only replay plan: artifacts, reviews, exports, gaps |
research-preset / research-preset-cn-social |
Agent presets: evidence-first persona + reading guardrails |
skills/ (10) |
The discipline: literature review, evidence audit, review, reproduction, design, figures, narrative, litdb, environment |
Data layout — everything lives in a plain directory you own:
.research/
projects/<id>/
project.json
artifacts/<sha256> # immutable payload, content-addressed
artifacts/<sha256>.json # sidecar: 7-face provenance
reviews/<artifactId>.json # per-round history + signature
exports/<name>.science # tar.gz, manifest + RO-Crate stub + payload
Quick start
# 1. clone into your harness plugins area
git clone https://github.com/peishenglai3-hash/dsh-research-mode.git
# 2. add the bundles to a profile (see docs/INSTALL.md)
dsh plugin --profile <your-profile> add link:./plugins/active/research-host
# ... repeat for the other bundles, then
dsh --profile <your-profile> --dump-config
Read docs/INSTALL.md for the full profile recipe, and docs/EVIDENCE-MODEL.md for what the
seven provenance faces mean.
Verification (run it yourself)
node --test plugins/active/*/test/*.test.mjs # 66 tests
node tools/w01/verify-acceptance.mjs # 13 acceptance checks
node tools/w01/probe-containment.mjs # 12 path-containment vectors
node tools/w01/s1-check-skill-contracts.mjs # skill <-> tool contract + mirror drift
All of these are zero-dependency Node scripts. They exist because "it works" should be a command, not a claim.
What this software does not do
This section is longer than usual on purpose. A provenance tool that oversells itself is worse than none.
provenance.reproducibility.availabledoes not mean "reproduced". It means a checksum, byte count and source were recorded. Nothing was re-run.provenance.environment.availabledoes not mean "reproducible environment". If you did not pass one in, the tool fills in{platform, arch, node, python}automatically. Compare the keys before trusting it.provenance.review.availabledoes not mean "reviewed by a human". Signature = unkeyed content hash. It proves the record was not edited; it does not prove identity.review verify → validdoes not prove history is complete. A single file cannot prove older rounds were not deleted or reordered. Rounds without signatures are reported as legacy, not valid.- There is no per-tool authorization gate. A tool call is not approved by a human at runtime.
confirm=trueis a tool parameter, not human consent. It is a speed bump, not a control..scienceimport read-only is declared, not enforced. On Windows the copied tree is still writable.- Nothing here claims scientific validity. It records what happened; it does not judge whether your argument is true.
Design principles
- The artifact is the unit, not the answer. Conversation is a channel that produces artifacts.
- Unknown is a first-class state. Missing evidence is recorded as
unavailable, never inferred. - Content addressing over mutable files. Identity = sha256 of the bytes.
- Local-first, file-based, no database server. You can
grepyour own research. - Discipline lives in text, not in a workflow engine. The skills say what good practice is; the tools record what actually happened. We do not pretend the two are the same.
Status
Early — v0.1.0. The evidence layer and the local data layout are exercised by a real acceptance
suite; the reading/parsing path depends on an external PDF parser (Docling) and is documented as
such, including its known timeout issue. Interfaces may change before 1.0.
This repository is the source of truth. Development, issues and releases happen here. See
CONTRIBUTING.md for the workflow and for the two things the test suite does
not cover (bundle mounting and the GUI).
Packaging: the bundles are currently installed by link: from a clone. npm publication is in
preparation so they can be installed as ordinary packages; until that lands, follow
docs/INSTALL.md.
License
MIT — see LICENSE.
Built as part of a sociology-of-AI practice: the questions that drove it were about digital inequality, social atomisation, and what accountability looks like when knowledge production is automated. The code is the smallest possible answer we could make honest.
No comments yet. Be the first to write one.