* chore(docs): add video demo to README.md for enhanced project visibility
* refactor(ai): centralize Claude Agent SDK env resolution with debug logging and model retry
Extract shared buildClaudeAgentEnv() and getClaudeAgentDebugFilePath() into
resolve-claude-agent-env.ts to eliminate duplicated env setup across all
server endpoints. Add model fallback retry logic (retry without explicit
model on exit code 1) and diagnostic hints from debug log tail on failures.
* fix(canvas): normalize layout justify/align values from CSS aliases
AI-generated nodes may use CSS-style values like "flex-start",
"space-between", "middle" etc. Add normalizeJustifyContent and
normalizeAlignItems to map these to internal enum values in both
the layout engine and the property panel.
* fix(canvas): improve text and path rendering accuracy
- Remove icon-specific bounding box override on path scaling to prevent
pathOffset drift that visually offsets icons in logo containers
- Only use Textbox for explicit fixed-width text modes instead of any
node with width > 0, preventing unwanted word wrapping on auto-width text
- Add widthSafetyFactor for Latin text estimation (1.14x vs 1.06x CJK)
to reduce accidental line wraps from font width variation
* feat(ai): add input icon affordance rules and trailing icon alignment
Prompt AI to include semantic icons in form inputs (search, password,
email). Add normalizeInputTrailingIconAlignment post-pass in
role-resolver to auto-set justifyContent="space_between" on input
frames with a trailing icon node.
* feat(ai): persist model selection across sessions via localStorage
Add preferredModel and selectModel to ai-store with localStorage
read/write. Hydrate on mount to restore user's last chosen model.
Add isHydrated flag to agent-settings-store to prevent race
conditions during provider list construction.
---------
Co-authored-by: Fini <fini.yang@gmail.com>
Install @iconify-json/lucide and @iconify-json/simple-icons as server-only
dependencies. Rewrite /api/ai/icon to read from local bundled JSON instead
of fetching from api.iconify.design, eliminating the external network
dependency that caused brand icons to silently fail.
Search order: simple-icons first (brand logos only — no UI name collisions),
then lucide (UI icons). Covers 3670 brand icons and 1729 Lucide icons
with sub-millisecond lookup. No impact on client bundle size.
- Add IconSection panel shown when a path node with iconId is selected
- Add IconPickerDialog: anchored popover with collection browsing, search,
current icon highlight, and outside-click/Escape to close
- Add iconId field to PathNode to track Iconify icon identity
- Fix canvas path sync to update d in-place (no object recreation/focus loss)
- Fix multi-element SVG icons: convert circles, rects, lines to path commands
so shapes like emoji face circles are not dropped during icon replacement
- Filter invisible bounding-box paths (Tabler's M0 0h24v24H0z) before merging
- Preserve original node width/height on icon replace; carry over display color
- Populate iconId in AI icon resolver and server-side Iconify proxy
- Introduced a layout engine for auto-layout computation, including padding resolution and fit-content size calculations.
- Added a node creator for various drawing tools (rectangle, frame, ellipse, line, text) to streamline object creation.
- Implemented object modification handling to sync changes back to the document store, improving interaction with Fabric.js.
- Enhanced text measurement utilities for better width and height estimation, including support for CJK characters.
- Added new API endpoints for AI icon retrieval and validation, supporting multimodal content analysis.
- Add mcp-server.cjs to extraResources in electron-builder.yml for production packaging
- Modify package.json build script to include MCP server compilation
- Enhance main.ts to set ELECTRON_RESOURCES_PATH for resource access
- Update mcp-install.ts to resolve MCP server path from resources in production
- Introduce support for .op file format alongside .pen
- Add MCP server functionality for document management and tool operations
- Implement batch processing tools for design and variable management
- Enhance save and open dialogs to accommodate new file format
- Update dependencies and scripts for MCP server compilation and execution
- Support thinkingMode/effort params in chat and generate endpoints
- Add Codex (OpenAI) provider streaming via codex-client utility
- Forward thinking config to both Anthropic SDK and Agent SDK paths
- Remove the complexity classifier as its functionality is no longer needed.
- Update design generator to always route through the orchestrator, simplifying the logic.
- Enhance error handling during orchestration to ensure fallback to direct generation is clear.
- Introduce a new sub-agent prompt for improved output formatting and clarity in design generation.
- Add Electron configuration and main process setup for building a desktop application.
- Implement IPC communication for file operations (open, save) between the renderer and main processes.
- Create a preload script to expose Electron APIs to the renderer.
- Update package.json to include Electron and related dependencies.
- Enhance the build process with electron-builder for packaging the application.
- Introduce a new electron-builder.yml configuration file for build settings.
- Modify Vite configuration to support Electron-specific builds.
- Update UI components to accommodate Electron's window management and drag regions.
- Add support for OpenCode as a new AI provider, allowing users to connect to a local OpenCode server.
- Implement new API endpoints for generating and streaming chat responses via OpenCode.
- Update UI components to include OpenCode in provider selection and display.
- Introduce OpenCode logo and enhance the agent settings dialog to manage OpenCode connections.
- Refactor existing chat and generation logic to accommodate provider-specific routing.
- Include type definitions for OpenCode SDK to ensure type safety and improve developer experience.
- Path icons: use uniform scaling (preserve aspect ratio instead of squishing),
fillRule 'evenodd' for compound paths with cutouts, transparent fill for
stroke-only icons, strokeUniform to keep stroke width constant
- Text centering: use actual Fabric rendered height (fontSize * lineHeight)
for cross-axis centering instead of declared height
- Hover cursor: use default arrow instead of move/crosshair on elements
- Stream reliability: server sends keep-alive pings every 15s during API TTFT,
forwards thinking_delta events, client resets timeout on ping/thinking chunks
- Remove silent fallback from streamChat to generateCompletion (was causing
double requests when server is unresponsive)
- Add 3-minute timeout to generateCompletion
- AI prompts: instruct model to preserve icon aspect ratios
- Add support for 'thinking' state in chat message streaming to indicate processing.
- Update toolbar to parse SVG files into editable nodes, allowing for better integration of SVG graphics.
- Refactor image handling to accommodate both SVG and raster images, ensuring proper placement and scaling on the canvas.
- Improve error handling and user feedback during file uploads.
- Add agent settings dialog to connect local Claude Code and Codex CLI
- Connect button fetches real models from Agent SDK / Codex models cache
- Provider logos (Claude, OpenAI) as React icon components in top-bar and dialog
- Model selector in AI chat panel grouped by provider with logos and Best badge
- MCP CLI integration toggles (Claude Code, Codex, Gemini, OpenCode, Kiro)
- Persist agent settings to localStorage with hydrate on mount
- Redesign AI chat panel: header, empty state pills, input area, color layering
- Add shadcn Switch component, agent-settings types and Zustand store
Enhance AI chat and generation functionalities by introducing model selection. Update the chat and generation endpoints to accept an optional model parameter. Implement a new API endpoint to fetch available AI models from the server, and integrate model selection into the AI chat panel for improved user experience.
Add floating AI assistant panel with streaming chat, design keyword
detection, and automatic PenNode JSON extraction to canvas. Server
endpoints use Anthropic SDK (with API key) or Claude Agent SDK (local
OAuth) as fallback. Panel supports minimize, drag-to-snap, and
quick action presets.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
