dsh-file-fix

DeepSeek Harness Web still only accepts images for upload: a PDF, a CSV, or a script dropped into the chat is either rejected with a 'PNG, JPG, WebP, GIF only' notice or leaves the drag overlay stuck. Even when a file somehow reaches the server, the model cannot see it and can only guess file paths—sometimes burning through dozens of steps.
dsh-file-fix closes that gap. Any file type can be uploaded via drag-and-drop, paste, or a file-picker button; bytes land in a content-addressed attachment store (SHA-256 deduplication) that is fully decoupled from the workspace, so it works equally well in self-hosted deployments. At send time the file manifest is injected into the model's context, and a bubble with file name, size, and a download link renders beneath the corresponding message in history. On the agent side, the read_attachment tool supports chunked reads (with automatic workspace mirroring for large files) and place_attachment exports attachment bytes to any validated path in the session workspace. The interaction layer offers a two-tier drop zone (image lane vs. file lane), three-state upload chips (uploading / done / failed-retry), Esc-to-cancel drag, and auto-greying of vision-unsupported models when the session involves image content to prevent silent failures.
Best for users who regularly handle documents, code snippets, spreadsheets, or other non-image assets in DSH Web, and for teams that need the agent to read and operate on previously uploaded files across multi-turn conversations.