dsh-tool-doc

Getting an AI model to process a PDF or an Excel file in a chat session typically means shelling out to external CLI tools. That adds dependencies, widens the attack surface, and makes the tool surface opaque. dsh-tool-doc takes a different approach: it registers document read, create, and edit operations directly as model tools so the model can parse, generate, and modify office files entirely within the conversation, with zero external process calls.

Under the hood, the plugin exposes three tools on ctx.tools. read_document extracts the PDF text layer page by page, parses DOCX body text and tables, iterates every XLSX worksheet and cell value, pulls per-slide text from PPTX, and handles quote-aware CSV, Markdown, YAML, and other plain-text formats, with output capped by maxTextChars (default 200 000 characters). create_document builds a new file from Markdown-lite input and supports DOCX with headings, bullets, and tables, PDF with an embedded CJK font when one is discoverable on the system, multi-sheet XLSX, and CSV, and the target path must not already exist. edit_document applies a sequence of literal, order-preserving edits to an existing file: DOCX supports in-run text replacement and paragraph append, XLSX supports set_cell / append_rows / add_sheet with 1-based addressing, and CSV supports set_cell / append_rows; every operation is validated up front so a single rejected edit aborts the entire call before any write.

If you are building AI workflows that revolve around office documents-reviewing contract PDFs, drafting weekly-report DOCX files, or updating data in XLSX workbooks-and you want all of that to flow through a governed filesystem seam with sandbox policy and approval events, dsh-tool-doc is a lightweight, drop-in option.