Gemini CLI v0.55.1 Changelog

This release is all about making the Gemini command-line tool safer and more reliable. It matters most to developers who use this tool daily to write code, because it fixes several sneaky security bugs, stops the AI from repeating its own mistakes, and introduces new background automation to help manage software projects.

What's new

• Tool Registry Discovery: The tool can now automatically find and organize available add-on tools (like plugins) much more easily.
• "Caretaker" Bot: Added a new automated service that runs in the cloud. It acts like a robot assistant that reads, sorts, and can even auto-close GitHub issues for the project.
• PR Generator Upgrades: Added a new system using a database (Firestore) and an "agent runner" to help automatically generate pull requests (code submissions) safely, even when multiple things are happening at once.
• Evaluation Reports: Added a new command to generate reports on how well the AI is performing in tests.

Improvements

• Safer Mac Sandbox: Updated the security rules (called Seatbelt profiles) for macOS so that the tool is locked down tighter by default, making your computer safer.
• Better Login Checks: Improved how the tool checks your saved login credentials so you don't get locked out by mistake.
• Clearer Error Messages: If you hit a limit on your account or are missing a setup step, the tool now gives you a much clearer message on how to fix it.
• Cleaner Workspace: The tool now ignores temporary background files created by automated tests, so they don't clutter up your actual project.

Fixes

• Stopped Endless Loops: Fixed a bug where the AI would get stuck in a "ReAct loop"—basically repeating the same thought and action over and over forever.
• Stopped "Thought Leakage": Fixed an issue where the AI's private, internal reasoning was accidentally leaking into your chat history.
• Protected Code Files: Fixed a bug where the AI tried to "correct" JSON and Jupyter Notebook (IPYNB) files when saving them, which was accidentally breaking the code. It now leaves them exactly as written.
• Blocked Folder Escapes: Fixed a major security flaw where "symbolic links" (shortcuts) could be used to trick the AI into accessing private files outside of your project folder.
• Prevented Remote Code Execution (RCE): Patched a vulnerability in the server that could have allowed a malicious task to secretly run dangerous code on your machine.
• Forced Secure Connections: Fixed an issue to ensure login tokens are always sent over a secure HTTPS connection, preventing passwords from leaking in plain text.
• Fixed Cancelled Tasks: Fixed a bug where cancelling a task while the AI was working would cause the whole program to crash with a "400 Bad Request" error.

Breaking changes

No breaking changes in this release.