Qwen Code v0.21.13 Changelog

This update is all about making the AI assistant easier to use and much more reliable. It gives you more control over your chats (like splitting them into different branches) and makes the automated code review system smarter so it doesn't get stuck or lose your work. This matters a lot to developers who use Qwen to manage code changes and review pull requests (proposals to merge new code into a project).

What's new

• Forking chats: You can now "fork" a conversation from any specific AI response. Forking means you branch off a new chat from that exact point, letting you explore a different idea without losing your original chat history.
• Text file attachments: The Web Shell (the browser-based interface) now lets you drag, drop, or paste text files as attachments. Previously, it mostly handled images.
• External Context Providers: There's a new system that lets outside tools feed extra information to the AI, helping it understand your code better.
• New review flags: You can now use a --since flag in the review command to check only the newest changes, rather than re-checking the whole project.
• Bilingual release notes: Update summaries are now grouped by theme and provided in both English and Chinese, with helpful screenshots.

Improvements

• Smarter code reviews: The /review tool now uses safer, dedicated commands instead of risky raw text commands. It also stops posting repetitive comments after a few rounds to avoid getting stuck in an endless loop.
• Self-healing downloads: If the system's workspace gets corrupted (like a messy folder that breaks the AI's download process), it will automatically wipe and retry the download to fix itself.
• Saved progress: If a review job crashes or gets interrupted, the system now records a session ID so it can remember and credit the work the AI already did.
• Better dashboard tracking: "Takeover" pull requests (where an AI tried to take over a task but had to stop) are now tagged with a "needs-human" label and put in a special dashboard pool so developers can easily find and finish them.
• Smarter audits: The system does a "reverse-audit" (a double-check of the code). It now adjusts how long this takes based on the size of the code changes—spending more time on small changes and less time on massive ones.
• Safer shutdowns: If you try to close the app while the AI is still working, it will take up to 8 seconds to safely finish its current tasks before shutting down, so you don't lose your work.

Fixes

• Blocked review jobs: Fixed an issue where empty "no-op" (no operation) human review requests were blocking actual, active code reviews from running.
• Lost comments: Fixed a bug where valid review comments were accidentally dropped if they matched existing comment IDs.
• Broken text tags: Fixed the review pipeline so it correctly accepts text wrapped in brackets, and added a tool to clean up these inputs.
• Lost findings: Fixed a problem where verified issues found outside a pull request's main changes were lost; they are now saved in a dedicated follow-up issue.
• Overwritten files: Fixed a bug where writing overlap lists accidentally overwrote the main findings file. They are now kept in separate files.
• Bot confusion: Fixed an issue where the triage agent (a bot that sorts issues) was accidentally trying to process tracking issues created by the autofix bot. A guard was added to stop this.

Breaking changes

No breaking changes in this release.