dsh-image-picker

The DeepSeek Harness Web GUI had no native attachment picker, and the official upload pipeline only accepted images - txt, md, and docx files hit a hard type error. Worse, browser security policies hide real disk paths, and inlining full text into messages can blow up the context window, silently truncating long documents.

dsh-image-picker adds a small clip button next to the input box that opens the system file dialog. Images (png, jpeg, webp, gif) flow through the official attachment pipeline with a thumbnail rail, size and count checks, and per-message upload. Text and docx files are read in the browser (.docx is unpacked via in-browser ZIP parsing plus DecompressionStream), POSTed to a local host endpoint, and staged as files under ~/.dsh/picker-attachments/. The message carries only a short receipt - absolute path, character count, and a brief preview - so the agent reads the staged file on demand with its normal tooling instead of burning context on raw text.

It suits anyone who mixes image and document attachments in a DeepSeek Harness web session: a quick screenshot plus a design spec or Markdown note, without switching to a terminal or pasting thousands of characters into the chat. A 25 MB per-file cap, automatic filename sanitization with a timestamp prefix to block path traversal, and a graceful fallback to inline text (truncated at 30k characters) when the staging endpoint is unavailable keep everyday use friction-free.