dsh-openpencil

Stop settling for a generated image: let the agent drive a real, editable, interactive design canvas and preview, inspect, and modify .op documents in-conversation.

What problem it solves

Asking an agent for a design usually yields a single generated image — easy to look at, painful to change, since every revision means regenerating, screenshotting, and round-tripping. DSH OpenPencil connects DeepSeek Harness with OpenPencil, the open-source AI-native vector design tool, so the agent drives a real, editable .op document instead of a bitmap.

Inside the conversation you get exact multi-frame previews: the OpenPencil headless exporter renders a large first-frame PNG plus a scrollable thumbnail rail. “Open interactive canvas” lazily mounts a read-only Web SDK with pan, zoom, and fit, so you can inspect any page or nested node without leaving the chat. With editable: true, the edit action opens the managed editor in a resizable right-hand workbench — selection, layers, properties, drawing tools, undo/redo, and explicit save semantics. Five agent-native tools create, modify, and read the real canvas through transactional batch_design programs, with capability-gated grants and transactional safety: a document publishes only after the whole batch succeeds, never overwrites an existing path, and a failed batch leaves no empty file.

Key features

• Exact multi-frame previews: a design-faithful first-frame PNG and thumbnail rail, with previous/next navigation
• Interactive canvas: lazily mounted read-only Web SDK with pan, zoom, and fit for any page or nested node
• Managed editor: selection, layers, properties, drawing tools, undo/redo, and explicit saves in a resizable, full-screenable workbench
• Agent-native design tools: five tools (openpencil_new / openpencil_create / openpencil_edit / openpencil_render / openpencil_selection) drive the canvas through batch_design programs
• Capability-gated grants: signed, hash-bound image/document capabilities that never expose arbitrary host paths
• One complete workflow: “requirement → agent edits the real canvas → live preview and validation → keep iterating”, with no screenshot round-trips

Quick start

Restart dsh web after installing and describe your design in plain language. On a first request the agent calls openpencil_new with a workspace-relative .op path and the first complete batch_design program, then openpencil_render with editable: true and autoOpen: true to show the gallery and expand the editor once. For an already-open canvas, openpencil_create and openpencil_edit apply edits that stay unsaved until you click Save. The core tools:

openpencil_new — create a fresh .op from a batch_design program, saved atomically

openpencil_edit — modify an explicit node or the single node selected by the user

openpencil_render — produce an immutable snapshot and render every top-level frame

openpencil_selection — read the exact nodes selected in the live canvas

Who it is for

• Creators using AI for UI/visual design who want editable design files instead of bitmaps
• Designers and frontend developers who want the agent inside the full iterate loop (preview, inspect, edit)
• Teams that need design output to be verifiable and auditable via hash-bound, transactional saves