dsh-document

Large language models have no native ability to parse the binary internals of Word, PowerPoint, Excel, or PDF files. The dsh-document plugin bridges that gap for DeepSeek Harness by giving the model a read_document tool: it converts nineteen-plus document and e-book formats into line-numbered Markdown, then the model pages through the result with offset and limit, reading only the section it needs rather than loading the entire document at once. Every conversion step stays on the local machine. Office formats are handled by the Rust-core anydoc engine, and PDFs run through a WASM build of pdf-inspector inside a dedicated worker thread. No API keys, no network calls, no external binaries. PDF handling includes per-page selection, automatic reporting of total pages, document type, and pages with no extractable text. Because files are read through the Harness ctx.fs seam, whatever filesystem provider and sandbox policy a deployment mounts applies without modification. This plugin suits teams and developers who want a fully local, closed-loop document-reading experience: meeting-minutes Q&A, contract summary extraction, spreadsheet comprehension, multi-page PDF digestion, and similar workflows, with no separate document-parsing service to stand up or maintain.