Claude Code auto mode stops falling back to its local classifier
Version 2.1.319 removes a silent fallback: when the server cannot check a tool call, auto mode now asks or denies rather than deciding on its own.
Anthropic shipped Claude Code 2.1.319 on 6 November, changing how auto mode behaves when it cannot reach Anthropic’s server to classify a tool call. Auto mode lets Claude Code approve or deny tool calls on its own, which is what makes unattended and headless runs practical. Until now, if telemetry or feature flags were off, or you were running behind an LLM gateway, a call the server couldn’t check quietly fell back to a built-in local classifier that decided for you.
That fallback is gone. The same calls are now asked about or denied outright, with no automatic approval. Setting CLAUDE_CODE_AUTO_MODE_SERVER=0 restores the old behaviour. Claude Code on Bedrock, Vertex or Foundry is unaffected. The same release adds a permissions.denyInsteadOfPrompting setting, and moves MCP tool calls flagged as tasks to the background instead of blocking the session while they run.
Why it matters: if you run Claude Code unattended, in CI, or behind a corporate LLM gateway with telemetry disabled, auto mode may now interrupt or block calls that previously went through silently. Check your pipeline before the next scheduled run, and set CLAUDE_CODE_AUTO_MODE_SERVER=0 if you need the old behaviour back.