dsh-files

When using DeepSeek Harness for real tasks, the gap between local files and something the model can reason about can be surprisingly large. dsh-files closes that gap by bringing uploads, image input, and document reading into one workflow: you can attach files from the composer toolbar, import a whole folder, or drag files directly onto the page, then reference uploaded or workspace files through @ candidates.

The plugin focuses on three connected capabilities. First, uploads are stored in a session-isolated way with deduplication, expiry, and size controls. Second, JPEG, PNG, WebP, and GIF images are handed to the Harness attachment pipeline, so models that declare image input can actually see the picture. Third, the read_document tool reads text, PDF, DOCX, and XLSX files with content sniffing, encoding fallback, pagination, and sheet-level access for spreadsheets.

dsh-files is a good fit if you want to do document Q&A, inspect reports or screenshots, feed code and configuration files to the agent, or use vision-capable models inside DeepSeek Harness. It is especially useful for users who need file-aware inference, multimodal reasoning, and clear session boundaries with safer upload handling.