2 min read AI-generated

Claude Code 2.1.183: Auto Mode Now Blocks Destructive Commands

Copy article as Markdown

The new release draws a hard line in auto mode: 'git reset --hard', 'terraform destroy' and friends get blocked unless you explicitly asked for them. Plus a deprecated-model warning and a batch of bug fixes.

Featured image for "Claude Code 2.1.183: Auto Mode Now Blocks Destructive Commands"

Auto mode is convenient: Claude Code works through a task without asking permission at every step. But that’s also exactly the moment when a single wrong command can hurt. Claude Code 2.1.183 tackles that risk head-on and builds in a safety net.

No more accidental data loss

The heart of this release: in auto mode, destructive commands are now blocked unless you asked for them yourself. Concretely, git reset --hard, git checkout -- ., git clean -fd, and git stash drop no longer run unless you explicitly said you want to discard local work. git commit --amend is blocked when the commit wasn’t even made by the agent during this session.

And it doesn’t stop at Git: terraform destroy, pulumi destroy, and cdk destroy get blocked too — unless you asked for that specific stack. This is the kind of protection you only learn to appreciate after you’ve torn down infrastructure by accident once.

Clearer warnings and a tidier startup

There’s also a deprecated-model warning: if the requested model is retired or has been automatically bumped to a newer version, you now see it in print mode (-p) — and new, also when the model is set in an agent’s frontmatter. The attribution.sessionUrl setting lets you keep the claude.ai session link out of commits and PRs, which is handy for web and Remote Control sessions in particular. The new /config --help lists all the shorthand keys for /config key=value.

On top of that comes a broad stack of bug fixes: WebSearch was returning empty results in subagents, Windows Terminal suffered corrupted fullscreen views under heavy subagent load, and scheduled tasks plus webhook triggers were wrongly treated as keyboard input — all fixed.

My take

This is exactly the right direction. The more autonomously Claude Code works, the more important guardrails become that distinguish between “go for it” and “wipe my repo.” What I like is that the logic doesn’t bluntly block everything — it looks at intent: if you asked for the discard yourself, it runs. That’s the difference between an annoying gatekeeper and one that actually thinks along. For teams running auto mode in production, this update is basically a must.

Sources: Claude Code Changelog, Releasebot: Claude Code