Claude Code v2.1.223 Changelog

This release is mostly about tightening security and fixing bugs to keep the app from crashing or acting weird. It matters most to developers and IT admins who rely on this tool for daily coding help and need to keep their systems locked down and running smoothly.

What's new

• Wildcard blocking for GitHub orgs: Admins can now use a wildcard (like "owner/*") in the settings to instantly allow or block all repositories owned by a specific GitHub organization.
• New warning for restricted models: If a background task or custom command tries to use a specific AI model that it isn't allowed to use, the system will now warn you that it's falling back to the main model instead.
• Teleport hint: When you are working in a cloud session, the app will now show a hint on how to move that exact session to your local computer using the /teleport command.

Improvements

• Smarter context limits: The app uses a feature called "auto-compaction" to summarize your conversation so it doesn't get too long. It now handles models with massive memory limits (a 1-million-token context window) much better, capping them at 200K to save resources. It also stops unrecognized models from blowing past their memory limits.
• Easier code reviews: The /review command is now just a shortcut for /code-review. If you don't specify how deep you want the review to be, the app will just remember and reuse the depth level you used last time.

Fixes

• Hidden command security bypass: Fixed a security loophole where a sneaky, crafted command could hide parts of itself from the permission check.
• Invisible character trick: Commands can no longer use invisible spaces or tabs to hide what they're doing from the approval screen.
• Sandbox escape: Fixed an issue where automated workflow scripts could use a trick called import() to run code outside of their safe, restricted area (sandbox).
• Permission override gap: Fixed a problem where an agent's setting to skip permissions would ignore the company's strict "no skipping permissions" policy.
• Broken session resume: Fixed a bug where using the /cd command to change folders mid-session caused the session to come back completely empty if you tried to resume it later.
• Missing Claude models: Fixed an issue where the app couldn't find Claude models if they were listed with specific provider tags (like vertex_ai/claude-*).
• Bad model overrides: Fixed a bug where typing an incorrect model ID would accidentally trick the app into using it as the main model. Unknown IDs are now safely ignored.
• Admin settings clash: Fixed an issue where settings sent from a server would accidentally wipe out local environment settings on your machine. They now merge together smoothly.
• Linux sandbox crash: Fixed a bug where sandboxed commands wouldn't even start on Linux if the sandbox was told to block writing to the current folder.
• Stuck background agents: Fixed a glitch where a background task would get stuck saying "already resuming" forever if something went wrong while trying to rebuild its prompt.
• Crash from bad data: Fixed a bug where a corrupted piece of data in the chat history would cause the app to fail every single turn or freeze on an error screen.
• Git hang-up: Fixed a rare issue where the app would freeze if it encountered unusual text while running a git push command.

Breaking changes

No breaking changes in this release.