dsh-doc-toolkit

Out of the box, DeepSeek Harness AI assistants can only handle plain text. When a conversation involves a PDF, Word document, spreadsheet, or CSV, users must manually convert the file before feeding it to the model. DSH Doc Toolkit closes that gap by registering two native tools, read_document and write_document, so the assistant can directly read and generate PDF, DOCX, XLSX, and CSV files. Parsed results are returned as plain text or TSV, giving the LLM zero-friction access to document content.

On the write side, generating DOCX, XLSX, or CSV from structured data is straightforward. PDF generation goes further: pure ASCII output uses the standard Helvetica face for a few-KB file, while CJK text triggers automatic discovery of a system TTF/TTC font with subsetting, producing a Type0 + Identity-H + ToUnicode structure whose text remains selectable and searchable. Tables repeat their header row across page breaks. Large files are read in pages via limit and offset parameters to protect the context window; any single read exceeding 50,000 characters is truncated with a truncated flag for follow-up paging. CSV encoding auto-detects UTF-8 and falls back to GBK/GB18030 for Chinese exports from Excel.

At load time the plugin registers a doc-toolkit-usage Skill that teaches the assistant when and how to call the two tools, removing the need for users to explain parameters each time. DSH Doc Toolkit is best seen as the document read/write infrastructure layer of the Harness ecosystem. It suits developers, analysts, and knowledge workers who want to keep their entire document workflow inside DeepSeek Harness without switching to a separate office suite.