Claude Code v2.1.221 Changelog

This update is all about making Claude Code smoother and safer to use, whether you're coding in a text editor, running automated background tasks, or using advanced cloud setups. It matters most to developers who use Claude to write or review code, because it fixes a lot of annoying bugs, speeds up how the AI thinks, and adds better protections for your private files.

What's new

• Focus view in VSCode: A new toggle (using Ctrl+Alt+F) hides all the messy background tool activity behind a clean summary. You'll see a live indicator showing what the AI is currently doing, keeping your screen clutter-free.
• File masking for security: On Linux and Windows (WSL), you can now hide secret credentials in a "sandbox" (an isolated, safe space for running code). Claude gets a fake, placeholder file to look at, but when it actually needs to send the real password to a server, the system safely swaps it in behind the scenes. On Mac, it just blocks access to the file entirely for safety.
• Plugin warnings: If you try to name a plugin something that the Claude Desktop app will reject, the system will now warn you ahead of time.
• Prompt auditing: A new tool checks your written instructions to see if they were designed for older AI models and might need an update for newer ones.

Improvements

• Smarter auto mode: When Claude runs several tools at once, it now checks permissions much faster. If you switch settings while it's thinking, it will reliably ask you for approval instead of using outdated choices.
• Cheaper AI calls: Claude now reuses parts of the conversation to save money and computing power when checking if it has permission to do things.
• Better stats panel: The stats screen now shows a complete breakdown of how many words (tokens) were used, including how many were read from or written to the memory cache.
• Faster Windows startup: On Windows, Claude used to open a separate PowerShell window just to check the time a process started. Now it asks the computer's core system directly, which means your security software won't accidentally pop up warning messages.
• Smarter background tasks: When Claude works on a task in the background, it now automatically saves and uploads your work to keep it safe. It only opens a "draft pull request" (a preview of code changes) if you specifically ask for one, and it always tells you exactly where your work was saved.
• Easier plugin installs: Installing plugins is smoother. If your plugin list is out of date, Claude will automatically refresh it and try again before telling you a plugin doesn't exist. Plus, safe plugins turn on instantly without needing a manual refresh.
• Better session info: The /status command now tells you exactly what kind of session you're in—like whether it's an interactive chat or a background job.
• Emoji shortcuts: You can now use common emoji codes like :thumbsup: or :love: instead of memorizing exact names.
• Smoother code branching: Using /fork to copy a chat session now creates a separate, isolated folder for your code, so it doesn't mess up the files from your original session.
• Browser cleanup: If Claude opens tabs in Chrome to look something up, it now closes them when it's done.
• Clearer credit warnings: If you run out of usage credits while Claude is working in "fast mode," it will now tell you on the screen instead of just failing silently.
• Better error messages: A tool called Monitor used to confusingly say "stream ended" if a watch command finished without printing anything. It now clearly says it produced no output. Also, a code review tool (/ultrareview) gives much better advice if you try to use it on a brand-new project with no history.

Fixes

• Hidden command bug: Fixed a security hole where the Zsh terminal could secretly run hidden commands without asking for your permission. It now prompts you first.
• Windows path bug: Fixed an issue on Windows where file paths with quote marks in the name confused the permission system.
• Thinking toggle: Fixed a bug where turning "thinking" off at the start of a chat accidentally turned it off for the whole session. Also, turning off a server connection mid-startup no longer breaks things.
• Server connection bug: Fixed an issue where custom servers weren't connected before Claude's first reply, which caused the AI to accidentally type out tool commands as regular text instead of running them.
• Dropped files: Fixed a bug where if you pressed Esc to cancel a message that mentioned a file (using @), the file would get deleted from your message when you tried to send it again.
• Crash fix: Fixed a crash that happened when preparing API requests for tools that had names like constructor (a reserved coding word).
• Web search fix: Fixed an issue where searching the web failed with an error if you had "thinking" turned off but set the search effort to maximum.
• Large uploads: Fixed a bug where uploading large files through the safe sandbox failed with a network (TLS) error.
• Confusing limit messages: Fixed an error message that blamed your company's monthly budget when you actually hit your own personal spending limit.
• Windows AWS login: Fixed an issue where logging into AWS on Windows failed if your computer had a leftover HOME environment variable.
• Resume setting: Fixed a bug where setting CLAUDE_CODE_RESUME_INTERRUPTED_TURN=0 didn't actually stop Claude from trying to resume interrupted messages.
• Sleep bug: Fixed a rare bug where waking your computer up from sleep could cause two Claude processes to try to refresh a login token at the exact same time, forcing you to log in again.
• Session renaming: Fixed an issue where renaming a chat from the desktop app or website didn't update the name in the terminal app.
• Skill commands: Fixed a bug where custom skills named after built-in commands (like /help) couldn't be used in automated, non-interactive sessions.
• Notification bug: Fixed a "Plugins changed" alert that wouldn't go away even after you reloaded your plugins.
• Vim mode fixes: If you use Vim keyboard shortcuts, copying (yanking) text now works properly when you open menus or search your history. Also, undoing all the way back to a blank prompt now correctly warns you before exiting.
• Gateway validation: Fixed a backend issue where invalid model names were sent through the system instead of being rejected immediately with an error.
• Annoying popups: Removed a repetitive warning message that popped up every time you changed permission modes while Claude was thinking.

Breaking changes

No breaking changes in this release.