Qwen Code v0.21.11 Changelog

This release is all about teamwork and making the developer experience smoother. It matters to anyone building apps with Qwen Code, especially if you are using the Web Shell (the browser-based interface) to manage multiple AI agents. You now have more power to customize agents, connect them to chat apps like DingTalk, and keep an eye on what they're doing behind the scenes.

What's new

• Agent Plugins v1: You can now build and add plugins (mini add-ons) to give your AI agents extra skills and tools.
• Multi-agent teamwork: A new /coordinate command lets a "leader" agent assign tasks to "read-only" helper agents. The helpers do the research and automatically send their findings back to the leader.
• Chat app channels: The Web Shell sidebar now has a "Channels" view. You can track tasks that are hooked into chat apps like DingTalk, Feishu, and WeCom right next to your normal tasks.
• Thinking controls for Qwen 3.8: When using the newest Qwen model, you can click a "model chip" on the screen to turn "Thinking mode" on or off and pick how much effort the AI puts into its answer.
• Terminal status symbols: If you use a terminal that doesn't show colors well, the window title will now show symbols (like ◐ and ✳) so you know what state a task is in.
• Memory doctor checkup: The /doctor memory command now tells you if the AI is holding onto too much text in its memory (over 30,000 characters) and gives you stats to help manage it.
• Better tracking (Observability): Added standard "OpenTelemetry" events (a way to track software performance) so you can see exactly when an AI chat session starts and ends.

Improvements

• Smarter text selection: Double-clicking and dragging now selects text word-by-word, and triple-clicking grabs a whole line, just like a normal word processor.
• Web Shell upgrades: Background task pop-ups now look like clean chat bubbles and can be translated into different languages. Subagent rows are more interactive when you hover over them.
• Better task management: ACP sessions (agent communication channels) now use a unified "Goal v3" system. This means you can smoothly create, edit, pause, resume, or clear tasks without confusing the AI's schedule.
• Smarter code reviews: The AI's review skill can now look at "sandboxes" (safe testing environments) and compare what the code is actually doing versus what the model thinks it should do, helping it catch hidden bugs.

Fixes

• Qwen 3.8 setting clashes: Fixed a bug where sending two conflicting "thinking" settings at the same time caused the AI request to fail completely.
• False error reports: If you manually cancel a request, the system now correctly realizes you meant to do it, instead of popping up a fake "API Error" message.
• Broken text streams: Fixed an issue where unclosed "thinking tags" (hidden code tags) would break the text stream when using OpenAI-compatible providers.
• macOS path mix-ups: Fixed workspace path checks on Mac computers so they correctly understand where files are located.
• Stale code repairs: The repair agent is now properly warned to rebuild its files if a task fails, so it doesn't accidentally trust old, broken code.
• Desktop app fixes: Fixed a bunch of minor glitches (regressions) from the last version, made the Mac app icon look better, and hid the internal workspace files until the app is totally finished loading.
• Web Shell keyboard fixes: Improved how the keyboard interacts when the Web Shell asks you a question, and made sure the workspace picker menu stays closed when you don't need it.

Breaking changes

No breaking changes in this release.