Qwen Code v0.21.3 Changelog

This release is all about making the AI a better, more reliable teammate for developers. It focuses heavily on improving how the AI reviews code (making sure it actually tests what it claims to test), giving you better visual tools to see what the AI is doing, and fixing communication glitches so everything runs smoother.

What's new

• Smarter Code Reviews: When you ask the AI to review a pull request (a proposed set of code changes), it now checks the "Test Plan" in your description against the actual code. It won't just take your word for it anymore!
• Pictures in the Terminal: There's a new tool that lets the AI display images directly in your computer's text terminal, complete with clickable web links and right-click menus.
• Background Task Approvals: If an AI "workflow agent" (a background helper) needs to run a computer command or edit a file, it can now pop up and ask for your permission directly, rather than getting stuck.
• Visual Workflow Map: There's a new "Session Workflow view" that draws a map (a dependency graph) showing exactly how the AI's plan is running, what tools it's calling, and what it's doing step-by-step.
• Custom Project Profiles: You can now create special settings files using simple Markdown text. These let you control exactly which tools the AI is allowed to use for specific projects.
• Download Web Files: In the Web Shell (the browser-based terminal), you can now easily download files, like HTML or text files, straight to your computer.
• More Simultaneous Sessions: The system now lets you run up to 32 AI chat sessions at the same time in one workspace (up from 20), which is great if you are doing a lot of heavy lifting at once.
• Live Status Emojis: If you use GitLab, the AI will post a little 👀 emoji on your code while it's reviewing it, and remove it when it's done, so you know it's working.

Improvements

• Better Test Verification: The AI's "verify" skill (how it double-checks its own fixes) has been upgraded to measure its suggested fixes against live code changes, making sure its solutions actually work.
• Faster Triage: Running the /triage command to sort through a pull request now immediately starts building and testing the code in a safe, isolated sandbox at the same time, saving you wait time.
• Smarter Chat Queue: If the AI is currently generating a response and you type another message, the system will smartly queue your new message up and show you exactly where it is in line, rather than ignoring you.
• Safe Commands While Loading: You can now run safe, simple commands like /status or /help while the AI is still typing out a response to you, so you don't have to wait.
• Skill Cleanup: Auto-generated AI "skills" (custom shortcuts) are now tracked. If you haven't used one in 30 days, it's marked as stale, and after 90 days, it's archived to keep your workspace clutter-free.
• Longer AI Memory: You can now adjust settings to use an extended 1-hour "prompt cache" (memory) for the Anthropic AI model, which keeps the AI remembering context longer without having to re-read everything.

Fixes

• Anthropic API Glitches: Fixed several issues with how the software talks to Anthropic's AI. It was creating duplicate text blocks, mixing up the order of messages, and using messy tool IDs, which caused errors. All of this is cleaned up.
• Web Shell Paste Bug: Fixed an issue in the Web Shell where pasting text from your clipboard didn't show up properly on the screen, making it look like your paste failed.
• Web Shell Context Switching: Fixed a bug where looking back at previous session recaps in the Web Shell would cause the screen to glitch or switch contexts incorrectly.
• Cache Mix-ups: Fixed an issue in GitHub Actions where the AI writing to the memory cache and the AI reading from it would get their wires crossed, causing mismatches.

Breaking changes

No breaking changes in this release.