Qwen Code v0.21.14 Changelog

This release is all about making the system smarter and less frustrating to use. It matters a lot to developers and teams who use Qwen Code to review code, manage AI agents, and collaborate on projects, giving them better tools to see what the AI is doing and keep their workflows running smoothly.

What's new

• Session tracking: You can now use a new command (qwen sessions ps) to list all currently running interactive sessions. You can even get this data in JSON format, which is a clean, organized way for computers to read it.
• Second opinions: There's a new /advisor slash command. It lets you ask for an independent, read-only second opinion on what the AI is doing, without messing up your original chat history.
• Pinned directories: Workflow agents (the automated helpers) can now be pinned to a specific working directory. This means you can force them to operate only in a specific folder, outside their usual boundaries.
• GitLab support: The code review tool now supports GitLab repositories (via something called Aone Code), not just GitHub.
• Expandable HTML exports: When you export your chat history to an HTML file to view in a browser, you now get an "Expand all" and "Collapse all" button to easily hide or show long sections of AI thinking or tool outputs.
• Flakiness gate: When the system tests your code in a sandbox (a safe, isolated space), it now runs modified tests multiple times to catch "flaky" tests—those that randomly pass or fail for no good reason.
• Privacy-safe diagnostics: The system can now track if tool results are too big or have been altered using secure "hashes" (like digital fingerprints) without actually looking at or exposing your private data.

Improvements

• Web Shell reliability: The Web Shell no longer locks up your typing if the data stream disconnects. You can keep typing your prompts, and it will send them when the connection comes back.
• Image previews: Pictures you reference in a session now survive page refreshes, so they don't just disappear.
• Smarter reviews: The review system now analyzes your shell and CI (Continuous Integration) scripts based on the exact environments where they actually run, making the feedback more accurate.
• Less spam: The system now checks new comments against a blocklist the moment they are created. Before, spam comments could stay visible for over an hour; now, the gap is practically zero.
• Cleaner comments: When the system posts review comments, it now writes them in plain, natural language instead of using robotic templates.
• Better team prompts: The instructions for AI agent teams were updated to correctly explain that they will automatically deliver a final answer when teammates are done working.

Fixes

• Crash prevention: Fixed a bug in the Web Shell where certain image formats would cause the system to crash.
• Duplicate messages: Fixed an issue where the Web Shell would accidentally show the same message twice.
• False failures: Fixed a problem where foreground agents (active helpers) were incorrectly marked as "failed" just because some routing data was missing when they started.
• Accurate cost analysis: Fixed the context usage details so it correctly ignores skills you've turned off, giving you a true picture of what's using up your context limits.
• Review pipeline: Repaired seven different defects in the review system found during live runs, including broken logic for finding data and allowing multi-step build-and-test checks.
• Interrupted reviews: Added a --resume flag to the fetch-pr command, so if a code review gets interrupted, it can pick up right where it left off instead of starting over.
• Retryable errors: The system will now automatically retry when it hits temporary issues like running out of resources or a quick read error, instead of just giving up completely.

Breaking changes

No breaking changes in this release.