Claude Code v2.1.222 Changelog

This update is a big maintenance release focused heavily on security, accurate error reporting, and accessibility. It matters most to developers using Claude Code in complex setups (like background tasks, custom servers, or isolated workspaces), but it also includes a highly requested fix for visually impaired users.

What's new

• Safer auto-mode: When the AI sends messages to other background agent sessions, it now checks its own permission rules first to make sure it's actually allowed to send that specific message.
• Better skill refusals: If Claude isn't allowed to automatically run a certain "skill" (a custom tool or workflow), it will now simply ask you to run it instead of trying to secretly copy the workflow itself.
• Cleaner diffs: When you compare changes to your code (using /diff or web sessions), the tool now looks at the raw, unmodified file data. This prevents weird custom settings on your computer from messing up how the changes are displayed.

Improvements

• Remote Control safety: "Remote Control" lets you manage Claude from another device. It can no longer be automatically turned on by settings hidden inside a specific project folder (which could be a security risk). You now have to turn it on yourself in your main user settings. Project folders can still turn it *off* if needed.
• Removed ultraplan: An older feature called "ultraplan" has been completely removed from the software.

Fixes

• Isolated workspace security: Fixed a major bug where "worktree-isolated sessions" (temporary side-branches of your project) could run destructive Git commands that accidentally wiped out files in your main project. The isolation now strictly applies to file edits and terminal commands.
• Background task bypasses: Fixed a bug where "auto-allow hooks" (rules that let tools run without asking permission) were accidentally bypassing security restrictions when the AI was doing background tasks like summarizing or renaming things.
• Credit request glitch: Fixed an issue on Team and Enterprise accounts where the /usage-credits command would tell users they had "already sent a request" if an older request was dismissed by an admin, permanently blocking them from asking again.
• Firewall/proxy startup hangs: Fixed the app freezing and crashing when starting up behind an HTTPS proxy (a common network security filter used by schools and companies). It now uses the correct network path and gives you a clear error message if it times out.
• Fake connection errors: Fixed the app reporting "Connection closed mid-response" errors on messages that had actually finished perfectly fine.
• Inflated usage stats: Fixed the /usage command blaming MCP servers (external tools hooked into Claude) for using too much power. It now only counts the server when it actually uses a tool result, instead of blaming it for every single thing that happens afterward.
• Missing Pull Request links: Fixed sessions failing to link up with Pull Requests on GitHub if the PR was created after you pushed your code branch.
• Model step-down failure: Fixed a bug where if an organization restricted access to a high-tier AI model (like "opus"), the system would drop down to the basic parent model instead of finding the newest allowed model in that same family.
• Custom gateway timeouts: Fixed the app timing out on custom API gateways (using ANTHROPIC_BASE_URL) even when the server was actively sending "keep-alive" signals to say it was still there.
• False authorization warnings: Fixed claude.ai connectors (external apps linked to your account) falsely claiming they needed authorization when your login token simply expired. They now correctly tell you to use /login.
• Missing tool errors: Fixed the app hiding error messages for tools that were recently removed (like if you deleted an MCP server), leaving you confused about why things weren't working.
• Long summary crashes: Fixed the SendMessage tool completely crashing when trying to send a really long text summary. It now just shortens the text instead of failing.
• Wrong effort labels: Fixed a UI glitch where the loading spinner showed the wrong "effort" level for background sub-agents, showing the main session's setting instead of the sub-agent's actual setting.
• File watcher crashes: Fixed rare crashes that happened when the system watching your files for changes hit an error or shut down.
• Screen reader backspace bug: Fixed a highly annoying accessibility bug where screen readers would read your entire input line out loud every single time you pressed backspace. It now only reads out the specific letters you just deleted.
• Stale settings overriding hosts: Fixed a bug where old, saved settings on your computer were overriding the model choices made by the host system when a specific environment variable was set.

Breaking changes

No breaking changes in this release.