Anthropic Releases Claude Code v2.1.328 with Auto Mode for Opus/Sonnet 4.6 and MCP Hook Blocking
What changed
Anthropic released Claude Code v2.1.328, bringing model execution updates, stricter tool hook behavior, and infrastructure metrics for self-hosted runners.
Key additions and fixes in this release include:
- Auto Mode on Cloud Gateways: Added auto mode support for Claude Opus 4.6 and Sonnet 4.6 across Amazon Bedrock, Google Cloud Agent Platform, Microsoft Foundry, and the Claude apps gateway. Interactive sessions configured without an explicit permission mode now start in auto mode by default.
- MCP Hook Controls: Added the
"onFailure": "block"parameter tomcp_toolhooks. Similar to existing command and HTTP hooks, any failing MCP hook execution now blocks the underlying action. This expands on the control policies updated in Claude Code v2.1.295. - Deferred Tool Execution Control: Introduced
CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME, an environment variable allowing host environments to preventclaude -p --resumeand--continuecommands from running tool calls deferred by a hook. - Self-Hosted Runner Updates: Added
--repository-folderto designate repository names under--capacity 1setups (resolving symbolic link path issues), added--hook-labelfor orchestrators to pass labels tospawn-runnerhooks, and added--metadata key=valueflags for session labeling underclaude --environment. - Telemetry & Metrics: Added two new metrics to the orchestrator’s
/metricsendpoint:claude_code_self_hosted_orchestrator_session_queue_wait_by_reason_secondsto differentiate queue waits between new and re-opened sessions, andclaude_code_self_hosted_orchestrator_sessions_given_up_totalto track failed runner startup attempts by root cause. - Command & Parsing Fixes: Resolved a Zsh error where commands containing leading equal signs (such as
echo ====or[ a == b ]) stopped execution due to path expansion. Fixed automatic context compaction so that it rebuilds system prompts dynamically rather than reusing stale prompts. - File Handling Fixes: Resolved an issue where unreadable scheduled task files caused project task deletion upon saving. Fixed token counting errors on subagents and status spinners that previously double-counted retried responses or stalled during model reasoning.
Why it matters
For enterprise teams using third-party cloud infrastructure like Amazon Bedrock, Google Cloud, or Azure/Foundry, auto mode parity ensures consistent autonomous execution without needing direct API connections to Anthropic.
The addition of blocking behavior for MCP tool hooks closes a reliability gap in agent workflows. Previously, an MCP hook failure might allow an agent to proceed unhindered; setting "onFailure": "block" forces execution pipelines to halt immediately when dependent tools fail.
Finally, the self-hosted runner metrics provide granular data for engineering teams managing agent clusters at scale. Tracking startup abandonments and queue wait reasons makes it easier to diagnose capacity bottlenecks under heavy multi-session workloads.
What to do next
- Update local CLI installations to v2.1.328 via npm or your configured package distributor.
- Update
mcp_toolconfigurations in.claude/settings.jsonto set"onFailure": "block"for critical tool dependencies. - Scrape the new
claude_code_self_hosted_orchestrator_*Prometheus metrics from self-hosted orchestrator instances to set up alerts for runner spin-up failures. - If running self-hosted runners using symbolic links under
--capacity 1, update runner deployment scripts to pass the--repository-folderflag.
Limitations
When running sessions through third-party cloud providers, large tool outputs can still hit prompt length limits. In v2.1.328, Claude Code handles this by discarding the largest tool results rather than crashing the session, which prevents permanent session failure but may truncate context provided to the model.
Sources
Share Article
If this article helped you, please share it with others!