Claude Code 2.1.183 Blocks Destructive Git Commands in Auto Mode

Claude Code 2.1.183 shipped on June 19 with a set of safety improvements for auto mode: the agent now refuses to run destructive git and infrastructure commands unless you’ve explicitly asked for them.

Auto Mode Now Guards Against Destructive Operations

The specific commands now blocked without explicit user intent:

• git reset --hard, git checkout -- ., git clean -fd, git stash drop — blocked when you didn’t ask to discard local work
• git commit --amend — blocked when the commit being amended wasn’t created by the agent in the current session
• terraform destroy, pulumi destroy, cdk destroy — blocked unless you asked for the specific stack destruction

The --amend restriction is the most nuanced one. If you have commits in your history that you made yourself, the agent won’t rewrite them even if amending would be the technically clean thing to do. It will only amend its own work from the current session.

This matters because auto mode runs without pausing for permission prompts. The agent is faster, but it also has more latitude to take actions that are hard to undo. These guardrails narrow what it will do unilaterally.

Model Deprecation Warnings

If you’ve configured a specific model and that model gets deprecated or automatically swapped for a newer one, Claude Code now surfaces a warning. The warning shows in stderr in print mode (-p) and applies to models set both in flags and in agent frontmatter configurations.

This fills a gap that was causing silent confusion: your agent instructions might specify a model that no longer exists, and previously the tool would just upgrade it without telling you. Now you’ll know when that happens.

Config and Startup Changes

/config --help now lists all available shorthand keys for the /config key=value syntax, making it easier to discover what you can set without digging through documentation.

The /config toggle UI also changed. Previously, pressing Escape reverted your changes. Now Enter and Space both toggle the selected setting, and Escape saves and closes. It behaves more like a standard settings panel.

The startup output is cleaner too. The “setup issues” line that appeared under the logo is gone. If you want to see configuration problems, run /doctor or pass --debug.

Bug Fixes

The release addresses several edge cases that have been tripping people up:

• Fixed thinking.disabled.display: Extra inputs are not permitted 400 errors when spawning subagents
• Fixed WebSearch returning empty results inside subagents
• Fixed terminal cursor getting stranded in vim mode after navigating history
• Fixed fullscreen TUI corruption in Windows Terminal under heavy nested-subagent load
• Fixed turns completing silently when the model returned only a thinking block
• Fixed user-level skills appearing multiple times in slash-command autocomplete when multiple plugins are enabled
• Fixed MCP servers exposing auth-stub tools to the model in headless/SDK mode
• Fixed tmux teammate pane launch failures with slow shell rc-file initialization
• Fixed background tasks being killed when the teammate that started them finishes a turn

Claude Code 2.1.183 is available via npm install -g @anthropic-ai/claude-code.

Source: Claude Code v2.1.183 release notes