@brainpilot/dsh-neuro-previewer
适用于 DeepSeek Harness (DSH) 的只读神经科学数据预览插件
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @brainpilot/dsh-neuro-previewer@0.1.0Neuro Previewer 以 @brainpilot/dsh-neuro-previewer 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
Preview neuroscience data directly inside DeepSeek Harness.
NeuroPreviewer is a neuroscience data preview plugin for DeepSeek Harness (DSH). It is installed as a DSH bundle with a Web client extension—not as a standalone viewer. The model inspects local data through the read-only neuro_preview tool, while the DSH Web client renders the result as a dedicated preview card.
@brainpilot/dsh-neuro-previewer · GitHub · MIT License
Compatibility: Version
0.1.0targets DeepSeek Harness0.1.0-rc.6. DSH remains a developer preview and may introduce breaking interface changes.
NeuroPreviewer gives a DSH agent a safe, compact way to answer basic questions about a neuroscience file without sending the full dataset to the browser. The Host plugin reads and validates the file through DSH's filesystem interface, creates a bounded 2D preview, and returns structured metadata plus a Web-friendly grayscale frame.
Current priorities are reproducibility, explicit resource limits, and graceful text output when the Web client is unavailable.
| Capability | Status | Details |
|---|---|---|
Single-file NIfTI-1 .nii | ✅ | Validates sizeof_hdr=348 and the n+1 magic |
| 3D MRI | ✅ | Axial, coronal, and sagittal slices |
| 4D fMRI | ✅ | Selectable zero-based volume |
| Numeric data | ✅ | uint8/int8/int16/uint16/int32/uint32/float32/float64 |
| Intensity processing | ✅ | Applies scl_slope/scl_inter and a 2%–98% percentile window |
| Endianness | ✅ | Little-endian and big-endian files |
| DSH Web preview card | ✅ | Canvas image, dimensions, datatype, plane, and intensity range |
.nii.gz and NIfTI-2 | Planned | Real fixtures are available; decompression/parsing is not implemented yet |
| BIDS metadata and CSV/TSV | Planned | Real BIDS sidecars and events files are available |
| EDF/EDF+, BrainVision, EEGLAB | Planned | Real EEG/PSG fixtures are available; adapters are pending |
| NWB and FIF | Planned | Intended for an optional Python worker |
The current renderer follows voxel storage order and does not yet reorient images using qform/sform. NeuroPreviewer is intended for research-data inspection and development—not clinical interpretation or diagnosis.
^22.19.0 or >=24.0.00.1.0-rc.6dsh plugin --profile web add @brainpilot/dsh-neuro-previewer@0.1.0
dsh --profile web --dump-config
dsh --profile web
git clone https://github.com/NeuroAIHub/NeuroPreviewer.git
cd NeuroPreviewer
npm install
npm run check
dsh plugin --profile web add "$(pwd)"
dsh --profile web --dump-config
dsh --profile web
The dumped configuration should include:
- id: neuro-previewer
name: '@brainpilot/dsh-neuro-previewer'
config:
maxFileBytes: 268435456
maxSlicePixels: 4194304
neuro_previewExample tool input:
{
"path": "/absolute/path/to/image.nii",
"axis": "axial",
"index": 48,
"volume": 0
}
| Parameter | Required | Default | Description |
|---|---|---|---|
path | Yes | — | Path to a .nii file accessible through the DSH filesystem |
axis | No | axial | axial, coronal, or sagittal |
index | No | Middle slice | Zero-based slice index |
volume | No | 0 | Zero-based volume index for 4D data |
Without the Web extension, the tool still returns a text summary containing dimensions, voxel size, datatype, slice position, intensity range, and warnings.
Real datasets are downloaded to the gitignored test-data/real/ directory and are never included in the repository or npm package. Every downloaded file is checked against scripts/real-data.sha256.
Download the complete corpus—approximately 190 MiB—and run the real-data smoke test:
npm run data:download
npm run test:real
Individual format groups can also be downloaded:
bash scripts/download-real-data.sh nifti
bash scripts/download-real-data.sh edf
bash scripts/download-real-data.sh brainvision
bash scripts/download-real-data.sh eeglab
bash scripts/download-real-data.sh nwb
| Source | Domain and format | Local fixtures | Preview support |
|---|---|---|---|
OpenNeuro ds000005 | Human structural MRI and task fMRI; BIDS/NIfTI | 3D T1, 240-volume BOLD, JSON/TSV, compressed and uncompressed NIfTI | ✅ Two .nii files |
| PhysioNet Sleep-EDF Expanded | Human sleep EEG/PSG; EDF+ | PSG and Hypnogram | Adapter pending |
| PhysioNet EEGMMIDB | Human motor-imagery EEG; EDF+ | 64-channel baseline recording | Adapter pending |
OpenNeuro ds007629 | Human natural-reading EEG; BrainVision | .vhdr/.vmrk/.eeg triplet | Adapter pending |
| EEGLAB sample data | Human EEG; .set/.fdt and BrainVision | EEGLAB pair and a compact BrainVision regression sample | Adapter pending |
DANDI 000006 | Mouse ALM extracellular electrophysiology; NWB | Two compact .nwb sessions | Python worker pending |
The current smoke test parses a real 160 × 192 × 192 3D T1 image and a real 64 × 64 × 34 × 240 4D fMRI image. Unsupported formats remain in the corpus as explicit negative fixtures so future adapters are tested against real files rather than synthetic substitutes.
See docs/real-datasets.md for pinned download URLs, licenses, citations, privacy notes, and per-file hashes. Public or de-identified human data must never be used for re-identification attempts.
npm run typecheck # Strict TypeScript checking
npm test # Synthetic unit and contract tests
npm run test:real # Smoke tests against locally downloaded real NIfTI data
npm run build # Host ESM and DSH Web client bundles
npm run check # typecheck + unit tests + build
The synthetic suite covers header validation, truncated inputs, little/big endian data, all three slice planes, 4D volumes, slope/intercept scaling, invalid indices, pixel limits, cancellation, DSH tool registration, filesystem adaptation, text output, and client presentation metadata.
DSH neuro_preview Tool
│
▼
NeuroPreview Interface
│
├── NIfTI Adapter (current)
├── EDF / BrainVision Adapters (planned)
└── Python Worker Adapter (planned: NWB/FIF/EEGLAB)
│
▼
Unified PreviewDocument
│
▼
DSH Web NeuroPreviewRow + Canvas
The format-neutral core is kept separate from the DSH integration:
src/core/preview.ts defines the NeuroPreview interface.src/core/nifti.ts detects and parses NIfTI-1 files and extracts slices.src/dsh/source.ts adapts DSH ctx.fs into a bounded BinarySource.src/index.ts registers the Host tool and model-facing output.src/client.tsx implements the DSH Web tool card.ctx.fs.AbortSignal cancellation..nii.gz, NIfTI-2, qform/sform reorientation, and interactive slice sessions..set/.fdt support.NeuroPreviewer is released under the MIT License. Real test datasets are not redistributed with the source code and remain subject to their original licenses, citation requirements, and privacy terms.