Claude Code 2.1.283 starts in auto mode on third-party providers and lets admins block new models

Most of the 2.1.283 release notes fix plugins, vim mode, and list navigation. Four groups of changes matter more: a new default permission mode, tighter control over models, two security fixes, and the reversal of a name reservation added one release earlier.

Auto mode becomes the default in two setups

Interactive sessions on third-party providers, or with telemetry turned off, now start in auto mode when no permission mode is configured. A permissions.defaultMode value in settings still overrides the new default.

The release notes do not name the providers. This suggests that teams using Claude Code through Amazon Bedrock, Google Vertex AI, or a gateway will see fewer approval prompts after the update, because the classifier now judges those actions instead. Teams that want every action approved by hand should set the permission mode explicitly.

Administrators can hold back new models

Two managed settings decide which models a team may pick:

• availableModelsMatch set to "exact" makes an availableModels entry allow only the model version it names, so new releases stay blocked until someone lists them;
• deniedModels blocks specific models even when availableModels allows them.

Two security fixes

On Windows, the PowerShell tool could delete drive roots, the home folder, and other folders that Remove-Item refuses to touch, by calling cmd /c rd, rmdir, del, or erase. Version 2.1.283 closes that path.

Managed sandbox settings were ignored entirely when one nested value was invalid. The invalid value now fails closed, and the rest of the block still applies.

The claude-ai name is available again

Version 2.1.282 reserved the anthropic-skills and claude-ai namespaces for skills synced from claude.ai. Version 2.1.283 reverts the claude-ai half of that change. Skills, commands, workflows, and MCP server skills and prompts with that name load again, and Skill(claude-ai:*) rules are ordinary prefix rules.

Prompt audit and review billing

A new /doctor prompt-audit command checks CLAUDE.md files, skills, agents, and commands for prompting patterns written for older models. The /ultrareview launch dialog now warns that reviewing a local branch may upload uncommitted changes to tracked files.

A Code Review run that stops at its time limit with nothing verified now shows as incomplete, is not charged, and is retried once.