Qwen Code v0.19.3 Changelog

This release is all about making the AI coding assistant more user-friendly and stable. It matters most to developers who use the tool daily, especially those who rely on the web interface or use voice commands to control their coding environment. The updates focus on better visual organization, preventing the AI from getting stuck, and cleaning up background processes so everything runs smoother.

What's new

• Excel-style tables: The assistant can now format markdown tables so you can interact with them like a spreadsheet.
• Voice control and keyterms: Added new APIs (communication hooks) for voice control. You can now set up a custom file of technical words so the voice dictation understands you better.
• Browse connected tools: In the web interface, you can now easily browse resources from MCP servers (external tool servers connected to the AI).
• Smarter tool display: The terminal now groups read-only tools together to save space, but shows tools that actually change your code individually so you can track them.
• Custom loading phrases: Hosts can now override the "loading..." text with their own custom phrases while the AI is thinking.
• Built-in status line: New users get a helpful status bar turned on by default.
• Extension mentions: You can now use @extension in your typing autocomplete to quickly reference add-ons.
• Thinking duration: The web interface now shows how long the AI spent "thinking" before it answered.

Improvements

• Cleaner startup: The startup text for "auto mode" is simpler and emojis were removed for a cleaner look.
• Better token speed tracking: The math for figuring out how fast the AI generates words (tokens) is now more accurate.
• Smoother loops: The self-paced /loop feature now relies on background notifications, making it run leaner.
• Faster serving: Sped up the qwen serve command by skipping an unnecessary wrapper step.
• Safer memory: The /remember command is now separate from auto-extraction, and it no longer writes directly to your main QWEN.md file without being asked.
• Configurable privacy: You can now adjust how much sensitive information is tracked in the app's telemetry (usage data).
• Live tool updates: MCP servers now update instantly if your settings change.
• Clearer Chinese text: In the web shell, the Chinese summary for tool groups was changed from "executed" (执行了) to "called" (调用了) to be more accurate.

Fixes

• Flickering terminal: Fixed an annoying screen flicker and scroll snap-back that happened when multiple AI agents were running.
• Frozen AI responses: Added an inactivity timeout to the OpenAI pipeline. Previously, if the AI stopped streaming data, it could hang forever; now it recovers.
• Duplicate text: Stopped the AI from repeating the same provider responses over and over.
• Lost reasoning: When merging assistant turns, the AI used to lose its "reasoning_content" (its internal logic). That is now preserved.
• WebFetch security: Rejected URLs with user info (like user:pass@site.com) to prevent security issues, and added a JSON fallback if a web fetch fails.
• macOS shortcuts: Fixed the alt+t shortcut so it correctly expands the AI's thinking process on macOS, and changed the display to show the proper ⌥T symbol.
• Windows shell leaks: Used "tree-kill" to properly shut down PowerShell (pwsh) processes on Windows so they don't leak memory.
• Gateway timeouts: Made chat-compression side-queries survive gateway timeouts so they don't crash your session.
• Web shell loading: Stabilized the active prompt loading state and deferred local commands so they don't mess up the transcript while the AI is streaming.
• Abort button: When you abort a self-paced loop, it now properly cancels pending wakeups so the AI doesn't keep trying to run.

Breaking changes

No breaking changes in this release.