Claude Code v2.1.259 Changelog

This update is all about making the software better for large teams and automated setups. If you're an IT manager running this for a whole company, or a developer using GitLab and running heavy automated tasks, this release makes your life easier and stops a lot of frustrating glitches from happening.

What's new

• Company-wide custom tools: Organizations can now automatically provide custom AI tools (called MCP servers) to all their users at once.
• Unattended mode: A new setting (--permission-prompts none) lets the program run on autopilot. If it tries to do something that normally requires a human to click "OK," it will just automatically deny the action and keep going. This is perfect for background tasks.
• Better GitLab support: The software now recognizes GitLab merge request commands. It will correctly label them as "MR !N" in the summaries and update the status badge at the bottom of your screen.
• Machine-readable plugin checks: You can now use --json with the plugin validator. This outputs the results in a format that other computer programs can easily read, rather than just plain text.
• VSCode session filters: If you use VSCode, the sidebar list of your active sessions now has filters. You can sort by what's currently active, what needs your input, what's working, and what's done.

Improvements

• Faster terminal display: Long responses now load and resize much faster because the program got smarter about reusing text measurements instead of recalculating them.
• Cleaner workflow results: When background agents finish a task, their JSON results (a way computers format data) are now nicely color-coded and easy to read, with long sections neatly tucked away behind an "expand" button.
• Faster startup for automated tasks: Headless (invisible, background) sessions start up to 50 milliseconds faster once their custom tools are connected.
• Smarter GitHub App installer: If you try to install the GitHub App while inside a GitLab project, the software now realizes its mistake, explains it's for GitHub only, and points you to the right GitLab instructions.
• Better background agent memory: If a background agent spawns another background agent, those results are now properly saved in the main log. If you pause and resume your work, you won't lose what the nested agents did.
• Clearer tool permissions: The rules for allowing custom tools (allowedMcpServers) now only apply to tools that users add themselves. Company-managed tools won't accidentally get blocked by user rules anymore.

Fixes

• Overwriting settings: Fixed a big bug where running multiple sessions at the same time caused them to silently overwrite each other's settings and lose project data.
• Stuck thinking: Fixed an issue where if the AI's internal thinking was rejected once, it would get rejected on every single turn after that.
• File security loopholes: Tightened up security rules that were accidentally letting the AI read forbidden files. It used to miss files hidden behind certain command flags (like -f.env) or inside folders. Now, it properly asks for permission or blocks them.
• Cache clearing: Fixed a bug where refreshing your login token accidentally wiped the AI's "short-term memory" (the prompt cache), making it slower, but only if you had telemetry (usage tracking) turned off.
• Blank full screen: Fixed a glitch where using full-screen mode after a really long task with hundreds of steps would just show a blank screen.
• Wrong AI models: Fixed an issue where custom commands tried to force the AI to use a specific model it didn't support, or ignored the model you asked for in interactive sessions.
• Token limits ignored: Fixed a bug where your custom token limits were ignored if you used a specific dating format (@YYYYMMDD) for AI models the software didn't officially recognize.
• Hidden terminal text: Fixed a visual bug where the newest lines of a running command would get hidden if an older line of text wrapped around the screen.
• Slow startup checks: Fixed a background check for GitHub connections that ran every single time you launched the app. It now remembers the result, making startup faster.
• Broken resume: Fixed an issue where trying to resume or continue a session would crash or open empty if it contained a broken file attachment.
• Artifact publishing error: Fixed a weird "unexpected parameter" error that happened when trying to publish an Artifact in a conversation started on an older version of the software.
• Managed settings enforcement: If a company's mandatory settings file is broken and can't be read, the software will now refuse to start and tell you exactly which file is broken, rather than just ignoring the rules.
• Git worktree errors: Fixed an issue where "worktrees" (isolated working folders in Git) created by automated hooks were rejected if your computer gave a slightly different error message than expected.
• Missing tracking data: Fixed a bug where cloud sessions weren't properly tagging usage data with user emails and organization IDs.
• Failing tool connections: Fixed an issue where if a custom tool disconnected while the software was starting up, it would falsely show as "connected" but with zero tools, instead of just telling you the connection failed.
• Cut-off permission prompts: Fixed a glitch where the popup asking for permission to edit a file would cut off the line of text it was showing you.
• Lost repo identity: Fixed a bug where a temporary glitch caused the software to completely forget which code repository you were working in.
• Unstoppable background tasks: Fixed an issue where clicking "Stop" on remote background tasks didn't actually stop them. Now, killed tasks stay visible and are truly stopped.
• Duplicate agents: Fixed a bug where resuming a workflow right after stopping it caused the software to accidentally run two copies of the same agent.
• Bad clone links: Fixed an issue where copying a GitHub link with a trailing slash or extra symbol (? or #) created a broken download link.
• Lost reasoning: Fixed a bug where "Stop hooks" (scripts that run when you stop a task) caused the AI to lose its reasoning memory and run slower on the next turn.
• 60-second delays: Fixed a massive 60-second delay that happened when starting a remote session after a browser-based tool's page had closed.
• Blocked bash loops: Fixed an issue where isolated Git sessions blocked normal bash loops and commands that couldn't reach the main project folder.
• Paused remote sessions: Fixed a bug where remote or scheduled sessions would just freeze and do nothing if you approved a permission prompt while the session was paused.

Breaking changes

No breaking changes in this release.