Claude Code v2.1.328 adds auto mode on Bedrock and Foundry
The release turns on auto mode for Opus 4.6 and Sonnet 4.6 across four providers and makes interactive sessions start in it by default.
Auto mode widens across cloud providers
Claude Code v2.1.328 was released on 19 November, and its headline change is Claude Code auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway. The same entry adds a second, quieter detail: interactive sessions that have no permission mode set now start in auto mode.
The rest of the v2.1.328 release notes is a long list of smaller additions and fixes, mostly around hooks, self-hosted runners, cloud sessions and terminal behaviour.
Hooks, warnings and a resume switch
MCP tool hooks can now declare "onFailure": "block", the same option command and HTTP hooks already had. When a hook's call fails, the action it guards is blocked instead of going ahead. That gives a hook author a way to fail closed.
Three additions are about knowing when something is going wrong rather than changing behaviour. A warning appears when a message from a session that skips permission prompts may be held and dropped by a cloud session. A notice shows up in /status and claude doctor when an OpenTelemetry exporter has no endpoint, which the notes say means its data would likely be lost. And apps that embed Claude Code can enable the policy plugin's allowManagedModsOnly option, with the notes stating that an admin's managed settings still win.
There is also a new environment variable, CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME, which lets a host stop claude -p --resume and --continue from running a tool call that a hook deferred.
Self-hosted runner gets naming flags and two metrics
Three runner flags stand out. --repository-folder names the folder a session uses for a repository at --capacity 1, replacing the path under the base folder, which cannot be a symbolic link. --hook-label passes chosen session labels to the spawn-runner hook, and --metadata key=value labels a session for that same hook.
Two orchestrator metrics were added to /metrics: claude_code_self_hosted_orchestrator_session_queue_wait_by_reason_seconds, so the wait of a re-opened session can be told apart from a new one's, and claude_code_self_hosted_orchestrator_sessions_given_up_total, which counts by cause how often Anthropic gave up starting a runner for a session. Startup warnings now fire when a host's settings.json sets a default permission mode that sessions ignore, and the Claude apps gateway warns when a policy's maxEffortLevel is not a valid effort level.
Fixes aimed at long-running and cloud sessions
Several fixes matter for anyone running Claude Code without watching every turn. Auto mode was failing every turn outside Claude Desktop when CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS sat behind a gateway rejecting betas, on Bedrock, Agent Platform, Foundry or the Claude apps gateway. Effort could be switched off for a subagent's model for the whole session if a provider refused its progress summary's effort setting, a regression from 2.1.327. MCP tools whose input schema starts with anyOf, oneOf or allOf were going missing on third-party providers, and token counts on the spinner and for subagents were double-counting retried replies.
Other fixes cover Bash commands in zsh stopping at words that start with =, scheduled task files being overwritten or losing entries, resumed cloud sessions dropping compaction summaries, and a batch of Code Mode script bugs.
Limits and open questions
The notes do not explain what auto mode approves on its own, how it differs from bypass permissions, or why it was previously limited. They also do not say whether existing sessions are migrated or only new interactive ones start in it. The date appears as 19 Nov on the release page without a year.
What this means for building apps with AI
If you drive a coding agent through a gateway or a self-hosted runner, this release is mostly about predictability: hooks that fail closed, warnings when telemetry or messages may be lost, and metrics that separate a fresh session from a reopened one. For people building apps with AI rather than infrastructure, the practical takeaway is that agent configuration now has more knobs, which makes a clear brief matter more. Draftship's guide to writing effective prompts covers that side of the work; Draftship itself is not connected to this release.
Primary sources
- v2.1.328 · Claude Code releases, Nov 19, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.