Qwen Code v0.21.4 Changelog

This release is all about making the software more reliable and independent. It matters most to developers who use Qwen Code to automate their work, because it gives them a smoother desktop experience, faster AI responses, and stricter security when multiple AI agents are working at the same time.

What's new

• Web Shell becomes a real desktop app: Instead of just running in a browser, the Web Shell is now a standalone desktop application. It manages its own updates and only lets you open one window at a time.
• Smarter code reviews: The review tool now has a special checklist for Java/JVM code to catch common performance mistakes. It also gives big pull requests (over 300 lines) up to 4 hours to be reviewed, instead of cutting them off too soon.
• Save summaries anywhere: You can now use the /summary command to save a project summary to a specific folder on your computer, and it will automatically create any missing folders for you.
• New "Session Workflow" mode: There's a new experimental setting you can turn on to help plan and review your work step-by-step.
• Better memory limits: You can now set a maximum number of "turns" (back-and-forth chats) an AI memory agent can take, with an option to turn the limit off completely.
• GitHub and GitLab management: You can now set up your security tokens and rules for GitHub and GitLab directly inside the Web Shell, rather than digging through settings files.

Improvements

• Faster chat compression: When a long chat gets shortened to save memory, the system now reuses cached information from your main conversation. This makes the AI respond faster without accidentally triggering background tools.
• Smarter review verification: When the AI checks its own code review findings, it now treats "unverified" issues as low-confidence warnings instead of flat-out rejecting them. It also double-checks the original sources first.
• Better daemon memory tracking: The background service (daemon) now shows exactly how close it is to running out of memory, so you aren't caught off guard by crashes.
• More accurate mutation testing: When testing how strong your code is by slightly changing it (mutation testing), the system now correctly labels tests as "inconclusive" instead of "survived" if the original tests were already broken. This prevents false alarms.

Fixes

• Windows paste bug: Fixed an issue on Windows where pasting multiple characters at once—like a secret API key—was completely ignored by the app.
• Web Shell history crashes: If one single chat message was too big for a page, it used to break the history viewer. Now it handles oversized messages gracefully and gives you a "retry" button if a page fails to load.
• Agent instruction leaks: Fixed a bug where "forked" subagents (AI helpers split off from the main task) could see the secret instructions of other sibling agents working at the same time. Now they are isolated and only see their own instructions.
• Language settings in reviews: The /review skill now correctly uses your chosen language for tips, saved reports, and labels, instead of defaulting back to English.
• Broken JSON responses: Fixed an issue where messy code responses with leaked protocol tags would pop up in the user interface. They are now safely routed through a retry system.
• Interrupted GitHub tasks: If a GitHub task gets interrupted, it can now safely recover and finish pending comments without having to start the whole AI agent over from scratch.
• Recovered long responses: If a long AI response gets cut off because the internet connection drops, the system can now save the part that was already delivered, separating it from the internal thinking steps that got lost.

Breaking changes

No breaking changes in this release.