Anthropic News | Claude Code v2.1.183 Safety Constraints and Deprecation Warnings
The Clauder Navi editorial team brings you the latest news from Anthropic on June 20, 2026, in a news broadcast style. Today's featured update is Claude Code v2.1.183. Safety constraints have been added to block the risk of unintended destructive operations being executed by agents. Deprecation warnings have also been strengthened, making this an update that directly affects engineers using Claude Code in production.
In Claude Code v2.1.183, safety constraints have been added to auto mode that block agents from executing destructive operations such as git reset --hard and git clean -fd without explicit user instruction. git checkout -- . and git stash drop are also restricted, and infrastructure management tools like terraform destroy are blocked unless a stack name is explicitly specified. As situations where tasks are delegated to agents become more common, this is an important change to prevent unintended destructive operations.
Warnings for teams using deprecated models have been strengthened across three display channels. Warnings are now delivered via standard error output, print mode (the -p option), and the header section of agent configuration files, making it easier to catch misconfigured settings pointing to old models. Teams working through model migration plans will benefit from being less likely to miss migration timing.
The /config --help command has been added, allowing users to view a list of available shorthand keys on the spot. Additionally, errors during sub-agent creation and full-screen display corruption in Windows Terminal have been fixed, further improving the stability of everyday Claude Code usage.
目次 (3)
[1] Claude Code v2.1.183 — Auto Mode Safety Constraints and Enhanced Deprecation Warnings
Claude Code v2.1.183 was released on June 19, 2026 at 01:20 UTC (source: https://github.com/anthropics/claude-code/releases/tag/v2.1.183 ). The impact rating for engineer readers is ★★★.
The biggest change in this release is the addition of safety constraints when agents operate autonomously. Destructive operations such as git reset --hard, git checkout -- ., git clean -fd, and git stash drop can no longer be executed without explicit user instruction. These operations carry the risk of losing uncommitted changes and untracked files if run accidentally. A mechanism is now in place to prevent the kind of trouble where "running some command while delegating a task to an agent caused work-in-progress files to disappear."
Constraints on infrastructure management tools are another notable addition. terraform destroy, pulumi destroy, and cdk destroy are now blocked when attempted without explicitly specifying a stack name. This serves as a preventive measure against the risk of agents accidentally deleting production infrastructure. This change is particularly important for teams that automate cloud resource management.
The strengthened warnings for deprecated models are also not to be overlooked. In cases where an old model is still specified in configuration files, warnings will now appear through one of three channels: standard error output, print mode (when using the -p option), or the header section of agent configuration files. This ensures that even when old model names remain in multiple configuration locations across a team, none are missed. Teams that have been putting off migrating to newer models are advised to take this opportunity to review their configurations.
On the usability improvement side, /config --help now lets you instantly check a list of shorthand keys directly from the command line. Configuration keys that previously had to be looked up in documentation can now be referenced on the spot, reducing friction when making configuration changes. Bug fixes include the resolution of an error that occurred during sub-agent creation, and a fix for full-screen display corruption in Windows Terminal. Engineers using Claude Code on Windows can expect an improved experience.
What to Watch Next
The addition of these safety constraints is an opportunity for teams using Claude Code in agent mode to review their existing processing workflows. Checking whether any calls to blocked operations are made without explicit instruction will help avoid unexpected behavior changes. The strengthened deprecation warnings are also creating momentum for teams to take stock of their configuration management. Now is a good time to verify within your team whether the models you are using have been updated to the latest recommended versions. Since the pace of Claude Code updates continues to be rapid, developing a habit of regularly following release notes will help engineers maintain quality.
Sources
- GitHub — Claude Code v2.1.183 Release Notes: https://github.com/anthropics/claude-code/releases/tag/v2.1.183
