dsh-plugin

Local OCR has long been a gap in the DeepSeek Harness Web profile: there is no off-the-shelf plugin that speaks both the DSH plugin API and PaddleOCR. dsh-plugin bridges that gap by building on the official DSH bundle and tool-registry interfaces and calling the PaddleOCR 3.x Python API, so all recognition runs locally with zero cloud dependency and no reliance on any external DeepSeek-OCR model.

The plugin registers a paddle_ocr model tool that the DSH Agent can invoke within the current workspace, and also exposes an /ocr [ch|en] human command for quick manual runs. Accepted inputs are strictly confined to PDF, PNG, JPG, WEBP, BMP, and TIF/TIFF files inside the workspace; symbolic links are resolved and must not escape that boundary. Each run produces a unique Markdown file under artifacts/paddle-ocr/, never overwriting the source document. All you need is a Python 3 environment with paddleocr 3.x installed, or point PADDLEOCR_PYTHON to the relevant interpreter if OCR lives in a separate virtualenv.

It is built for individuals or small teams who regularly work with scans, invoices, contracts, and similar documents inside the Harness Web console. Just tell the Agent something like "recognize scans/invoice.pdf and tidy up the result," and you get a clean, structured Markdown text without the manual copy-paste and reformatting overhead.