Qwen Code v0.19.0 Changelog

This release is all about making the AI assistant feel more like a helpful coworker and less like a plain text terminal. It matters a lot to developers and power users who use this tool daily, because it introduces cooler ways to interact (like voice and interactive web pages) while cleaning up the visual clutter and fixing behind-the-scenes glitches that caused crashes or weird behavior.

What's new

• Voice dictation: You can now talk to the AI instead of typing. It captures your voice natively, streams the text as you speak, and even biases the transcription to understand tech words better.
• Interactive HTML pages (Artifacts): The AI can now publish interactive web pages directly in a window. You can even set them to auto-open so you can immediately play with things like charts or mini-apps the AI just wrote.
• Dynamic Workflows: This is a big upgrade for automating tasks. You can now save workflows, trigger them with specific keywords, get notifications when they finish, and resume them if they get interrupted.
• Background sub-agents: Think of these as helper bots. If they get interrupted or time out, they are now "revivable," meaning you can bring them back to life instead of starting all over.
• Cleaner interface: The visual layout got an upgrade. Completed tool results collapse out of the way, bulky borders were removed, and file previews now open in a resizable side panel instead of taking up the whole screen.
• Timestamps: You can turn on an optional timestamp (like [14:32:05]) before every AI response so you can track exactly when things happened.
• Browse MCP resources: MCP (Model Context Protocol) is a way for the AI to connect to outside tools. You can now browse exactly what those outside tools offer right in the /mcp chat menu.
• Stricter file naming: The system now enforces "kebab-case" for filenames (words-lowercase-with-dashes). This keeps everything organized and prevents bugs when sharing code across different computers.

Improvements

• Smarter model hiding: Developers can now use special tags (fastOnly or voiceOnly) to hide specific AI models from the main menu, keeping the list clean and only showing the models you actually need for standard tasks.
• Better AI "thinking" viewer: When the AI is reasoning through a problem, you can press Alt+T to expand or collapse its "thinking block," making it easier to see exactly how it arrived at an answer.
• Custom ignore files: You can now set up custom files to tell the AI exactly which folders or files to ignore, similar to a .gitignore file but specifically for the AI agents.
• Cleaner provider setup: Behind the scenes, the code for connecting to outside AI providers (like OpenRouter) was simplified. Instead of needing clunky custom code, it just uses standard custom headers, making it faster to maintain.

Fixes

• Stopped infinite loops: The AI used to sometimes get stuck calling the exact same tool over and over. A new guardrail now catches this and stops it. It will also safely fail if it detects a loop during non-interactive runs.
• Fixed voice recorder crashes: If your computer was missing certain audio pre-build files, the voice recorder used to fail silently. Now it warns you and smoothly falls back to a backup method, and it fixed a bug where the recorder wouldn't retry after an error.
• Prevented settings double-counting: There was a bug where usage stats were being counted twice. This is fixed, so your data tracking is now accurate.
• Stopped web session glitches: If you deleted a chat session in the web interface, it used to annoyingly auto-recreate itself. That no longer happens.
• Fixed missing web content: When using certain AI models (like GLM on DashScope), the AI was accidentally dropping content it fetched from the web. It now keeps all the text.
• Safer settings migrations: When updating your settings to a new version, the system now ensures the update won't break if run twice by accident.
• Fixed broken commands: If a tool call was left dangling (like a paused task that got cut off), the system used to choke. It now fails gracefully instead of crashing.
• Better environment variable parsing: If you typed in a weird or invalid format for things like API timeouts or rate limits, the system used to break. It now strictly checks those formats and handles them safely.

Breaking changes

No breaking changes in this release.