dsh-file-upload

DSH Web's native drag-and-drop experience is limited: images are outright rejected when the current model lacks vision support, there is no unified archive for uploaded files, and tracing back to previously sent attachments is a hassle. dsh-file-upload drops an upload button—styled with official DSH design tokens, theme-aware—into the composer toolbar and takes over window-level drag events, giving you the most natural way to bring local files into a conversation. Images travel through the DeepSeek Files API official attachment rail with automatic file_id reuse and 7-day expiry re-upload; documents, code, and any other file type are archived into the attachment library while a path line lands in the draft; dragging a folder recursively reads the tree, rebuilds it under the library with its original layout, and inserts only a single folder path into the draft. Even when the model cannot see images, the plugin gracefully degrades to path text so you never hit a 400 error. Starting with v0.2.0, the plugin also ships an uploaded-file management panel (paged browsing, one-click @path copy or delete) and zero-copy clipboard path reading: copy a file or folder in your OS file manager, click the clipboard button beside the upload icon, and the real path appears as an @ reference in the draft—no disk copy, no size cap.

This plugin is a great fit for two kinds of users. First, developers who frequently reference local docs, screenshots, or code snippets inside DSH Web loops and want to skip the repetitive open-manager-copy-paste dance. Second, anyone who values a tidy archive: every upload lands under ~/Documents/DSH/Attachments organized by date, and the settings-card panel lets you browse, copy references, or clean up at any time—the plugin itself never auto-deletes your files. The implementation is pure Node (only clipboard reading shells out to a platform tool), cross-platform (fully tested on macOS, expected on Linux and Windows), released under MIT, and maintained independently by the community with no dependency on official DeepSeek teams.