Gemini CLI v0.52.0 Changelog

This update is mostly about making the Gemini command-line tool smarter and less frustrating to use. It matters most to developers who use Gemini in their terminal to write code, because it fixes several annoying bugs with how the AI saves files and handles errors, while also laying the groundwork for new background automation.

What's new

• New "Caretaker" background tools: The developers added the first building blocks for a new automated system (called "caretaker-triage" and "caretaker-egress"). Think of this as a robot assistant that will eventually help sort through project issues and talk to GitHub automatically. It’s not fully running yet, but the foundation is now there.
• Updated login software: The tool that handles your Google login (an authentication library) was upgraded to a newer version to keep things secure and running smoothly.

Improvements

• Less clutter for the AI: The tool will now ignore temporary setup files created by automated testing systems (CI). This means the AI won't get confused by files that don't actually matter to your project.
• Simpler file planning: When you are just planning out changes (Plan Mode) before actually writing them, the tool now handles simple, relative folder paths (like ./folder/file instead of long, full paths) much better.

Fixes

• Stops the AI from "correcting" code files: Previously, when saving or replacing files, the AI would sometimes try to "fix" the formatting inside strict code files like JSON or Jupyter Notebooks (IPYNB). This would actually break the code! Now, it leaves those files exactly as they should be.
• Clearer account warnings: If your account doesn't have the right "Code Assist" permission level, or if you share a project and hit a usage limit, the tool used to give confusing errors. Now, it shows a clear message explaining the problem and gives you a hint on how to fix your setup.
• Better understanding of your commands: If the AI got confused about what you asked it to do, it used a fallback summary that was sometimes misleading. Now, it labels your previous intent clearly so it doesn't get off track.
• Stopping tasks actually stops them: If you asked the tool to do a long, automated task (on an "a2a-server") and then hit cancel, the AI used to keep working in the background. Now, hitting cancel actually stops it right away.

Breaking changes

No breaking changes in this release.