v0.21.0 — Virtualized Chat, Mermaid Diagrams & Shared Lib Extraction
Pre-release
Warning
This is a pre-release and can only be downloaded manually due to potential issues found out after testing.
What's New
⚡ Much Faster Chat
Long conversations are dramatically faster now. We replaced the old rendering approach with a proper virtualized list — only the messages you can actually see are rendered at any time. Previously, a 500-message chat would keep hundreds of React components in memory. Now it's just the ~20 visible ones, regardless of how long the conversation gets. Scrolling is smoother, switching sessions is snappier, and the app uses less memory overall.
📊 Mermaid Diagrams #29
When Claude draws a diagram using a mermaid code block, it now actually renders as a visual diagram — flowcharts, sequence diagrams, pie charts, git graphs, and more. Diagrams adapt to your light/dark theme automatically and are cached so they don't re-render on scroll. While Claude is still typing, you see the raw source; once the message is complete, the diagram appears.
📂 Deep Folder Inclusion (@#)
You can now use @#foldername in the composer to include the full contents of a folder — not just the file tree, but every file inside it. Regular @folder still gives you the structure overview. Use @# when you want Agent to actually read everything in a directory. If the folder is large (over ~50k tokens), Harnss will warn you before sending.
⌨️ /clear Command #15
Type /clear in the composer and hit Enter to instantly open a fresh chat. Quick way to start over without reaching for the sidebar.
🛡 Better Error Feedback
Permission prompts (allow/deny tool use) now show a notification if something goes wrong responding to them, instead of failing silently. Fixes several edge cases where permission state could get stuck.
🐛 Bug Fixes
- Fixed a bug where switching out of plan mode could sometimes reset the permission level incorrectly
- Fixed markdown characters occasionally getting eaten during streaming (apostrophes, backticks, etc.)
- Clipboard copy now works correctly in all cases including sandboxed contexts
- Codex web search cards now show meaningful details (queries used, URLs opened) instead of blank tool inputs
- And many other small irritating bugs I don't remember to mention
Full Changelog: v0.20.0...v0.21.0