dsh-document: a plugin for reading PDF / Office documents

[jiaoqsh]

Today read rejects PDF and Word files as binary and web_fetch rejects PDFs by content type. dsh-document adds a read_document tool that turns PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, and CSV files into line-numbered Markdown, paged with offset/limit like the built-in read.

PDF extras

• Page selection — pages: "1-3,7" reads only those pages; <!-- Page N --> marks where each page starts
• Page facts — page count, text-based vs scanned, and exactly which pages have no text layer (content that would need OCR), instead of a vague failure
• Worker thread — large files never block the dsh process, and Stop terminates the conversion

Under the hood

Local only, no API key, no network: anydoc (Rust napi) for office formats, pdf-inspector (WASM) for PDF. Files are read through the ctx.fs seam, so the profile's filesystem provider and sandbox policy apply unchanged; input size, lines/bytes per call, and the PDF page cap are cordis.yml config.

Install

npx @deepseek-ai/dsh plugin --profile web add @jiaoqsh/dsh-document
npx @deepseek-ai/dsh web

https://github.com/jiaoqsh/dsh-document

Status

0.1.1, tested against dsh 0.1.0-rc.6; MIT, not affiliated with the project. Known limits: no OCR (the tool only reports which pages lack text), and layout-heavy PDFs may collapse paragraphs into long lines. Both engines sit behind a DocumentConverter interface, so an OCR or hosted engine can be added without changing the tool.