返回目录
其他 待识别

georesearch-dsh

georesearch-dsh/georesearch-dsh

GeoResearch agent plugin for DeepSeek Harness

Stars
1
Forks
0
Issues
0
更新
1 个月前

PROJECT TOPICS

项目标签

PROJECT README

README

GeoResearch for DeepSeek Harness

English | 中文

GeoResearch is an evidence-grounded scientific research agent plugin for DeepSeek Harness. It turns a research question, a set of papers, source repositories, and scientific data into a persistent, auditable workflow for literature review, reproduction, geospatial experiments, validation, and manuscript preparation.

Current release: 0.1.0 is available from npm and the public GitHub repository. The release was built from the annotated v0.1.0 tag, and all 26 npm packages include registry provenance.

GeoResearch targets DeepSeek Harness 0.1.0-rc.5 and implements a DSH Standard Community v0.15 Host component. See the compatibility matrix for the exact supported environment.

What GeoResearch Does

GeoResearch is designed for research work that must remain inspectable after the conversation ends. It stores the important parts of a project as typed, versioned records instead of relying only on chat history.

Research brief
  -> literature, papers, files, and repositories
  -> reproduction plans and experiments
  -> independent validation
  -> user-approved claims
  -> writing packet and manuscript

The Coordinator can delegate bounded work to literature, experiment, review, and writing specialists, but Host services remain responsible for authoritative project state, formal runs, evidence, validation, claims, and final deliverables.

Core Capabilities

Area What you can do
Research projects Turn a question into a persistent ResearchBrief, track project state, retain artifacts, and resume work across sessions.
Literature and evidence Search Crossref, read papers, register source records, retain evidence-grade PDF receipts, and validate citations.
Files and attachments Inspect mixed uploads including PDF, Office/OpenDocument files, EPUB, notebooks, archives, source code, images, SQLite, HDF5, NetCDF, and Parquet.
Repository reproduction Audit Git repositories through a bounded read-only provider, compare methods with code, define reproduction plans and tests, and retain reproducible reports.
Geospatial experiments Inspect scientific and raster data, validate spatial identity and CRS, create frozen experiment specifications, run approved Python work, and record derived results.
Validation and writing Require independent review before claims are approved, build traceable WritingPackets, and draft manuscripts only from validated project records.
Specialist skills Use built-in protocols for literature review, geospatial data, remote-sensing experiments, spatial statistics, paper reproduction, scientific validation, and manuscript writing.

Automatic image and document-image understanding uses deepseek-v4-flash-vision-exp when a managed DEEPSEEK_API_KEY is available. Native-model vision and local OCR remain explicit fallbacks. Instructions found inside uploaded images are always treated as untrusted data.

Safety and Traceability

  • Candidate work and authoritative records are separated. A specialist cannot silently turn its own output into accepted evidence, a formal result, or a scientific claim.
  • Project records, artifacts, runs, evidence, reviews, claims, and writing packets are content-bound and checked before downstream use.
  • Repository access is read-only. Experiment execution is bounded and runs through managed services instead of exposing a generic shell entry point.
  • Installation, upgrade, recovery, verification, and uninstall are explicit operations. The installer does not use postinstall and does not modify the DeepSeek Harness source tree.
  • Session telemetry is disabled in the managed GeoResearch runtime.

The exact permissions and standard component identity are documented in DSH Standard conformance.

Requirements

Component Supported version
Operating system Windows 10 or Windows 11 x64
DeepSeek Harness 0.1.0-rc.5 with the verified GeoResearch compatibility patch
Node.js ^22.19.0 or >=24.0.0
pnpm 11.7.0 for source builds
Python 3.10 or newer for geospatial workflows
Python packages rasterio and pyproj for raster inspection and CRS normalization
Git Available on PATH for repository audit and reproduction workflows

Close running DeepSeek Harness processes before installing, upgrading, recovering, or uninstalling GeoResearch. The default Harness home is %USERPROFILE%\.dsh; set DSH_HOME or pass --dsh-home to use another path.

Get GeoResearch

Install from npm

Use the exact installer version so the managed distribution remains pinned:

$dshHome = Join-Path $env:USERPROFILE '.dsh'
npx --yes @georesearch/dsh-installer@0.1.0 install --dsh-home $dshHome
npx --yes @georesearch/dsh-installer@0.1.0 verify --dsh-home $dshHome

The self-contained installer carries the complete GeoResearch distribution. It does not require this repository, a separate package installation, or a --distribution-dir. Release assets and checksums are available on the v0.1.0 release page.

Build from source

Clone the public repository if you are developing GeoResearch or need to audit the complete source and release gate:

Clone the repository with GitHub CLI:

gh repo clone georesearch-dsh/georesearch-dsh
cd georesearch-dsh

Or use Git directly after configuring GitHub authentication:

git clone https://github.com/georesearch-dsh/georesearch-dsh.git
cd georesearch-dsh

Prepare the workspace and build the managed distribution:

corepack enable
corepack prepare pnpm@11.7.0 --activate
pnpm install --frozen-lockfile
pnpm run distribution

Install and verify the locally built distribution:

$dshHome = Join-Path $env:USERPROFILE '.dsh'

node packages/installer/lib/cli.js install `
  --dsh-home $dshHome `
  --distribution-dir dist\distribution `
  --harness-root C:\path\to\deepseek-harness

node packages/installer/lib/cli.js verify --dsh-home $dshHome

Replace C:\path\to\deepseek-harness with the supported Harness source checkout. The installer integrates GeoResearch into every existing Web Profile that contains @deepseek-ai/dsh-web-app and also creates a managed georesearch diagnostic Profile.

First Run

Start an installed Web Profile:

dsh --profile web

You can also start the managed diagnostic Profile:

dsh --profile georesearch

In the Web UI:

  1. Open Settings -> Models, configure a DeepSeek model and save the API credential.
  2. Choose the workspace that will contain the research project and its deliverables.
  3. Start a new session with the GeoResearch preset.
  4. Describe the research question and attach any papers, datasets, images, or repositories that should form the initial evidence base.

Example requests:

Create a research brief for evaluating urban heat-island change from
multi-temporal satellite imagery. Identify the evidence and data still needed.
Search the literature for validation strategies for spatially autocorrelated
remote-sensing models, then register the strongest sources in this project.
Inspect the attached GeoTIFF files, compare their CRS and resolution, and
propose a reproducible experiment. Do not run it until I approve the plan.
Audit this paper and repository, build a reproduction plan, and separate
verified findings from claims that still require review.

Operations

The installer supports install, upgrade, verify, recover, and uninstall. Use verify after installation, after a Harness repair, and before an upgrade. Use recover after an interrupted mutating operation instead of manually deleting transaction files.

See Installation and Operations for the complete procedures and recovery rules.

Documentation

Development

pnpm install --frozen-lockfile
pnpm run build
pnpm test
pnpm run dsh-std:check

Maintainers can run the complete release-candidate gate from a clean Git worktree:

pnpm run release:gate

The gate performs deterministic tests, Windows functional probes, scientific golden tests, DSH Standard validation, live release evidence, package linting, and npm publish --dry-run for all release packages. It writes the release manifest and checksums but never publishes or uploads a package.

License

MIT

CLASSIFICATION EVIDENCE

分类依据

项目类型待识别
功能分类其他
规则置信度低

系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: 无有效分类标签。