dsh-anydoc-markdown

In a DeepSeek Harness workflow, when an agent processes Word, PowerPoint, Excel, PDF, EPUB, and other documents, two problems stand out: plain-text extraction loses table structure, formulas, footnotes, and speaker notes, while embedded images remain completely invisible to a text-only model. dsh-anydoc-markdown solves both by converting virtually every common document format into clean GitHub-Flavored Markdown, faithfully preserving heading hierarchy, tables, lists, equations, links, and footnotes.

Every embedded image is passed to a vision-language model in controlled batches (up to 10 per request by default, avoiding a single oversized prompt that degrades inference). The resulting description replaces the image placeholder in the Markdown so a text-only model can reason about charts, screenshots, and diagrams. When no vision endpoint is configured, the plugin gracefully falls back to honest offline metadata descriptions with zero network calls and no API key needed. The output (a Markdown file plus every extracted image saved as a PNG) lives together in a single subfolder, making it portable out of the box.

This plugin suits teams that want their agent to summarise business reports, parse formula-heavy academic papers, or interpret architecture diagrams embedded in slide decks. Interaction is drag-and-drop: drop documents onto the Composer page, see them as cards in the input dock, and submit. The agent then calls the convert_document tool on demand with no pre-configuration required.