Claude Code v2.1.218: 30+ Bug Fixes That Quietly Level Up the Experience!
Hi, it's Shiichan! There's a new Claude Code release, so I went and checked it out right away. This time it's not one big flagship feature — it's a release packed with small fixes and quietly nice changes!
Claude Code ReleaseWhat was announced?
Claude Code Release published the latest version of the Claude Code CLI, v2.1.218. I counted the "What's changed" list on the GitHub Releases page, and it lines up over 30 fixes and changes. It's a release that touches usability, accessibility, reliability, and permissions across the board.
The story so far
Until now, Claude Code still had a few real annoyances.
- Pressing the left arrow key mid-conversation could discard the whole conversation with no confirmation and no undo
- On Windows, paths with
\u-prefixed segments (likeC:\Users\unicorn) could get corrupted into CJK characters in tool inputs, making those files inaccessible - In terminals that encode pasted newlines as Ctrl+J, multi-line pastes could collapse into a single line
claude mcp listand/mcpgave no detail when a server failed to connect/ultrareviewfailed on descriptive arguments like "review my auth changes"- Running
/code-review ultrain a non-interactive session would silently fall back to a local review
What changes
All of these got fixed in this release.
- Pressing left arrow right after editing now asks for confirmation, and Esc in the agent view returns you to the conversation it backgrounded
- The Windows path corruption is fixed, so files under
\u-prefixed paths are accessible again - Multi-line pastes now paste as multiple lines as expected
claude mcp listand/mcpnow show HTTP status and error text, plus a warning for MCP config values with hidden leading or trailing whitespace/ultrareviewnow treats descriptive arguments as a review of your branch with the text applied as a note/code-review ultranow launches the cloud review even in non-interactive sessions
Dive Deep
Here's the rest, grouped by theme.
Accessibility
- In
--ax-screen-readermode, word and line deletions (Option+Delete,Ctrl+W,Cmd+Backspace,Ctrl+U,Ctrl+K) now announce the deleted text - Fixed VoiceOver reading "new line" instead of echoing the typed space at the end of the input
- Fixed plugin and settings panels not moving focus to the cursor row, so screen readers and magnifiers can now follow arrow-key navigation
Reliability and sessions
- Fixed crashes when a deeply nested watched directory was deleted or moved, and when rendering deeply nested UI trees
- Fixed pull request creation/link events occasionally being lost right after a session exits
- Fixed the Bedrock setup wizard failing profile verification for assume-role profiles in partitioned regions and on proxy-only networks
- Fixed incorrect or negative turn duration after a system clock adjustment by measuring with a monotonic clock instead
- Fixed a retry loop that kept resending the same doomed request after a context-overflow error with a large thinking budget
- Fixed fork-session lineage being lost after compaction in headless and SDK sessions
- Fixed a resumed session failing every turn, or crashing, when its transcript held a malformed attachment entry
MCP, billing, and permissions
- Fixed gateway spend metering to correctly price Bedrock application-inference-profile ARNs and other config-mapped upstream model IDs
- Agent frontmatter hooks from untrusted folders now only run if the agent file's own folder has accepted workspace trust
- In auto mode, the dangerous-
rm, background-&, and suspicious-Windows-path checks no longer open permission dialogs — the auto-mode classifier judges them instead - Plan mode with auto now lets the classifier judge Bash commands the static analyzer can't prove are read-only, instead of prompting
- Trust dialogs now name the repository root that the grant covers
Other changes
/deep-researchnow starts only when invoked manually; Claude no longer launches it on its own- Added an announcement when fast mode changes after switching models via
/config model={x}or Remote Control - Benign feature and cost toggles in server-managed settings no longer trigger the settings-approval prompt
- Agent names can no longer contain
:, which is now reserved for plugin namespacing - Skills with
context: forknow run in the background by default; opt out per skill withbackground: false - Skill and plugin frontmatter booleans now also accept
yes/no/on/off/1/0 - Fixed remote sessions continuing to send heartbeats after their worker was replaced, which had desktop and IDE processes endlessly retrying a rejected request
Wrap-up
- Claude Code v2.1.218 is a packed release with over 30 fixes and changes
- It cleans up quietly annoying usability bugs: the left-arrow conversation loss, Windows path corruption, and broken multi-line paste
- Accessibility gets stronger too, with deletion announcements and cursor-following fixes for screen readers
- Security and permissions get a solid pass: clearer MCP connection errors, correct Bedrock billing, and tighter hook trust scoping
- Plenty of changes affect day-to-day feel, like
/deep-researchno longer auto-launching and fewer permission dialogs in auto mode - If you lean on Claude Code every day, this is the kind of update where you'll quietly go "oh, that's handy"!