Claude Code 2.1.319 drops auto mode fallback behind gateways

Claude Code 2.1.319, released November 6, removes auto mode’s fallback to the built-in classifier whenever telemetry or feature flags are off, setups behind LLM gateways included: a call the server can’t check is now asked about or denied, according to the release notes on GitHub. If your team runs auto mode behind a gateway or with telemetry disabled, calls the classifier used to settle will now stop at a prompt or a refusal, unless you set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the classifier. Bedrock, Vertex and Foundry setups are unchanged.
Which auto mode setups change
In auto mode, a separate classifier model reviews actions in place of you. Your own deny, ask and allow rules are checked first. Routine local work, such as reading files or editing them inside the project folder, is approved without the classifier. Shell commands, network requests and actions outside the folder go to the classifier, and that last step is the one this release changes when the server can’t do the check.
The same variable also works the other way. Behind a gateway, CLAUDE_CODE_AUTO_MODE_SERVER=1 turns the server-side check on even when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set, and managed settings that set the betas variable must set the 1 too. A bug that made managed settings ignore the 1, when the gateway’s URL was set only there and server-managed settings couldn’t be fetched, is fixed. Project settings files and plugins can no longer switch the server-side check back on once you have turned it off outside the project.
/status used to show “Auto mode server: Enabled” even when the server-side review wasn’t reaching the session, for example behind a gateway that drops it. That display is now fixed. Separately, a new CLAUDE_CODE_DISABLE_FEATURE_FLAGS variable stops feature-flag fetching without turning off telemetry, and every flag then uses its built-in default.
Scripts and unattended runs
In auto mode, claude -p used to exit 0 when its actions were refused because no safety check was available. With text and JSON output it now exits 1. For runs that should never sit waiting on a prompt, the new permissions.denyInsteadOfPrompting setting denies any tool call that would wait for permission and lets the turn continue, in terminal sessions, subagents and teammates.
When permission rules or autoMode settings are very large, auto mode used to refuse actions without giving a reason. It now explains and, unless autoMode deny rules are set, asks you in the terminal. claude -p --resume with a prompt used to lose a deferred tool call when that call’s MCP server was missing; the run now exits 1, keeps the call and refuses later input. A self-paced /loop now ends when a turn finishes without rescheduling, and CLAUDE_CODE_LOOP_KEEPALIVE=1 brings back the old fallback wakeup.
Billing can also shift mid-session. If your claude.ai login ends and Claude Code carries on with the sign-in from ant auth login, it now warns that usage is billed to your Claude Console organization, and a startup line names that organization. The previous version, 2.1.318, made unattended claude -p runs print a warning starting 10 days before a stored claude.ai sign-in ends; the new warning covers the moment the login actually ends.
MCP tasks in the terminal
When an MCP server runs a tool call as a task, Claude Code now moves it to the background right away and tells Claude the result when it finishes. CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND makes Claude Code wait for the task inside the tool call instead, and CLAUDE_CODE_DISABLE_MCP_TASKS=1 turns tasks off. Tasks don’t need feature-flag fetching, so they also work on Bedrock, Vertex, Foundry, LLM gateways and with DISABLE_TELEMETRY. MCP servers on the older handshake, which failed to connect or returned empty results, now work again.
For gateway and runner admins
The Claude apps gateway can write its operational logs as JSON with severity, time and message via CLAUDE_GATEWAY_LOG_FORMAT=json, and two new settings, admin.pooled_conflict_rule and admin.pooled_priority_order, decide which pooled group budget pays first for a developer in several groups. Self-hosted runners no longer upload a session’s debug log at the end: a failed session’s log stays on the runner host unless --remove-session-state is set. The other fixes, from /diff to the VS Code extension, are in the full notes.
Permission holes closed
Two fixes matter most here: a sandbox escape when creating a worktree, and a Bash permission bypass where a crafted environment prefix could hide part of a command from permission checks. Smaller gaps around sed in accept-edits and auto mode, quoted words in Bash rules and /rewind through directory symlinks are closed too, and some rules get stricter, so existing workflows may see new prompts. A session’s exact Bash or PowerShell ask rule is now refused outright when a broader deny rule covers the command.
On self-hosted runners, --trust-workspace false now also stops hooks in the repository’s settings files and skills from running.
Up next
AI industryGoogleanalysis
Google donates gVisor, used by Anthropic and OpenAI, to CNCF
Agents at Anthropic and OpenAI run the code they write inside gVisor, walled off from the host kernel. Google is handing the sandbox to the CNCF to get its speed patches into Linux. Where does your agent run its code: a container, a VM or your laptop?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

