Claude Code auto mode now refuses or asks when its server-side check is unreachable
Claude Code 2.1.319 removes the fallback to the built-in classifier when telemetry or feature flags are off, including behind LLM gateways.
Starting with Claude Code 2.1.319, auto mode no longer falls back to the built-in classifier when telemetry or feature flags are turned off. This includes setups where Claude Code runs behind an LLM gateway. According to Anthropic's changelog, calls the server cannot check are now either sent to the user for approval or denied outright. Teams that want the old behavior can set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep using the built-in classifier. The changelog says Bedrock, Vertex and Foundry deployments are unchanged.
The change comes down to where the safety decision gets made. Auto mode lets Claude Code approve tool calls without asking the user each time. In the setup the changelog describes, a server-side check decides whether a given action is acceptable. When that check cannot be reached, Claude Code previously used a classifier built into the client. The 2.1.319 entry removes that path for the configurations it names, so an action the server cannot review is either put to the user or refused. The changelog also notes that with CLAUDE_CODE_AUTO_MODE_SERVER=1 set, the server-side check is turned on even when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set. Managed settings that set the latter must also set the former.
The release contains several related changes to where that setting can come from. A project's settings files or a plugin can no longer switch the server-side check back on once a user has switched it off outside the project. The changelog also says that CLAUDE_CODE_PROXY_RESOLVES_HOSTS can no longer be set from settings files where the host manages the network. It must be set in the process environment instead.
The release also contains related fixes for auto mode. The entry says that claude -p used to exit 0 in auto mode when its actions were refused because no safety check was available. Text and JSON output now exit 1. Another fix says /status could show "Auto mode server: Enabled" when the server-side review was not reaching the session, for example behind a gateway that drops it. Before this release, then, a headless run could report success even though its actions had been refused for lack of a check. A status display could also show the server check as on when it was not reaching the session.
The changelog does not say how often the server-side check is unreachable in practice. It also does not compare the built-in classifier with the server-side one in accuracy. Readers who depend on auto mode behind a gateway will need to check the setting for their own deployment, since the changelog says Bedrock, Vertex and Foundry are unchanged.
What changed
Claude Code 2.1.319 no longer falls back to the built-in auto mode classifier when telemetry or feature flags are off, so unreachable server-side checks are now sent to the user for approval or denied.
Type: Breaking change · Applies to: Claude Code 2.1.319, CLAUDE_CODE_AUTO_MODE_SERVER, LLM gateway
Why it matters
Organizations routing Claude Code through an LLM gateway may see more prompts or denials in auto mode after upgrading, which could slow automated work.
What you can do
Teams using auto mode behind an LLM gateway should review their CLAUDE_CODE_AUTO_MODE_SERVER setting before upgrading to 2.1.319.
Original source
Where we saw it
Claude Code changelog
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works