lubenweimeiyoukaig/dsh-file-upload-local

A local file-upload plugin for DeepSeek Harness (dsh). Upload files straight into the conversation via a paperclip button or drag-and-drop; files are stored per-session under .dsh-uploads/<sessionId>/ inside the session workspace, and the agent reads them with the read_document tool.

中文

Features

• Upload surface: a paperclip button in the composer's left slot, plus a full-viewport drag-and-drop overlay.
• Attachment rail: uploaded images render 64px thumbnails (click to open), other files render type glyphs with names; each tile has a hover remove button. References are appended invisibly to the sent text on Enter or the send button.
• read_document tool: pages through text files with line numbers (offset/limit), and OCRs images with tesseract.js (chi_sim+eng) so a text-only model can still read screenshots and photos.
• Per-session isolation: files land in .dsh-uploads/<sessionId>/ under the session workspace; TTL sweep keeps it from growing unbounded.

Install

dsh plugin --profile web add dsh-file-upload-local
# restart dsh web

Use

1. Click the paperclip button, or drag files anywhere onto the window.
2. Files appear as thumbnails above the composer; remove any with its ✕ button.
3. Send the message — the agent reads uploaded files with read_document <path> (text pages, image OCR).

Configuration

- id: dsh-file-upload-local
  config:
    uploadMaxBytes: 26214400   # single upload body cap
    uploadTtlMs: 604800000     # 7 days
    sweepIntervalMs: 3600000   # TTL sweep interval
    maxConcurrentUploads: 4
    maxFileBytes: 26214400     # read_document read cap
    readLimit: 2000            # max lines per read_document page
    ocrLangs: 'chi_sim+eng'    # tesseract language pack

Notes

• The OCR language pack (chi_sim+eng) is downloaded by tesseract.js on first use; keep network access or pre-seed the cache.
• Images travel through the OCR path (read_document), so a text-only model works out of the box. If the active model supports image input, you can still use the official read_image tool directly.

License

MIT