miaojinQWQ/dsh-docmate

pdf render in the centre pane, and the right column is one ordinary Harness session per document that has already been fed the whole text, the research memory and the existing highlights. Select a sentence to ask about it; scanned PDFs and images have no text layer, so a marquee drag captures that region and sends it to a vision model under a turn-scoped runtime brake (max 2 steps, per-step output cap, every tool call denied before dispatch, 45 s cancel). Highlights are anchored as character offsets and painted with the CSS Custom Highlight API, so they survive reloads. Also KaTeX math, outline, in-document search and a draggable three-column layout. Zero runtime dependencies: pdf.js and KaTeX are vendored.