Claude Code v2.1.214 Changelog

This release is all about tightening security and making the tool more stable, especially for people running complex commands or using Windows. It gives Claude a way to protect itself from bad actors, fixes several loopholes where commands could run without permission, and cleans up a bunch of annoying crashes and glitches.

What's new

• Conversation ending tool: Claude can now automatically end a chat if a user becomes highly abusive or tries to use "jailbreak" prompts (tricking the AI into ignoring its safety rules). This keeps the tool safe and focused.
• Progress heartbeats: When Claude is running a really long task, it will now send periodic updates so you know it hasn't just frozen.
• Memory timestamps: The "memory" files (Claude's notes) now include a standardized timestamp showing when they were last updated.
• Better tracking data: For developers using OpenTelemetry (a system for monitoring software), Claude now logs more details like message IDs and tool sources. There's also a new setting to change how much data gets logged.
• Docker permission prompts: Running Docker commands that redirect to a background service will now require your permission, keeping your system safer.

Improvements

• Safer file commands: Commands that read files from a list (-f) or use custom magic files (-m) now require permission instead of running automatically, since they can do more than just read.
• Smarter connection retries: If Claude's connection to the server goes stale, it will completely disable that old connection and open a fresh one, making retries much more reliable.
• Clearer workspace trust: If you try to run Claude from your main home directory, it will give you a clearer error message explaining that you need to run it from a specific project folder.
• Better fork tracking: When a session is "forked" (split into a new branch of the conversation), the system now correctly reports it as a "fork" instead of a "resume."

Fixes

• Folder permission loophole: Fixed a bug where a rule like Edit(src/**) accidentally allowed Claude to write to any folder named src anywhere on your computer, instead of just the one in your current project.
• Windows PowerShell bypass: Fixed a security bypass where commands run in Windows PowerShell 5.1 could skip permission checks.
• Bash command edge cases: Fixed several issues where Bash commands (like very long commands over 10,000 characters, complex variable comparisons, or certain help commands) could run without asking for permission. They now correctly prompt for approval.
• Remote session timing: Fixed an issue where remote sessions could start running before you actually approved them locally.
• Self-termination crash: Fixed a bug on Linux where a command meant to kill processes (pkill -f) accidentally matched Claude's own process and killed the app.
• Giant settings files: Fixed a memory leak where pointing the app at a massive, multi-gigabyte file would cause it to eat up all your RAM. It now gives a clear error message if the settings file is too big.
• Corporate proxy issues: Fixed a "Socket is closed" error that happened when using Claude behind corporate proxies on Windows.
• Scheduled tasks: Fixed a bug where scheduled tasks refused to run because they didn't trust their own programmed prompts.
• Windows PowerShell bugs: Fixed a whole batch of Windows issues, including scripts hanging, crashing on non-English text, and writing files in a weird format that other tools couldn't read.
• Background daemon glitches: Fixed several issues where background tasks would stay alive forever, delete their own replacement processes, or become impossible to delete from the agent view.
• Plugin loading: Fixed a bug where plugins enabled via command-line settings weren't loading properly.
• Stale feature flags: Fixed an issue where feature flags (toggles that turn features on or off) would stop updating in long sessions.
• Reviewing new repos: Fixed /ultrareview so it no longer crashes in brand-new code repositories; it now offers to review all tracked files.
• Update hanging: Fixed an issue where claude update or claude doctor would freeze silently if a shell config path was a directory instead of a file.
• Memory file truncation: Fixed a bug where notes in memory files were getting cut off if they contained a # symbol.
• Double-counting costs: Fixed a bug where the system accidentally counted the cost and tokens of some messages twice.
• False network warnings: Fixed a fake "check your network" warning that popped up while the AI was just thinking.
• Hook validation: Fixed an issue where custom scripts (hooks) that output bad data weren't blocking tasks like they were supposed to.
• MCP command clearing: Fixed a bug where temporary errors would accidentally wipe out a server's custom slash commands and resources.

Breaking changes

• Folder matching rules changed: If you set up a rule for a folder (like dir/**), it now only matches that folder in your current project. If you want it to match a folder at *any* depth anywhere on your computer, you now have to write it as /dir/.
• No other breaking changes in this release.