Failing Closed with Agent SDK Hooks in Claude Code 2.1.317

Explore the updates in Claude Code versions 2.1.321 and 2.1.322, featuring the introduction of Claude Fable 5.2 and Fast mode support on Google Cloud Vertex AI. Learn how to configure your environment, manage separate prompt caches to avoid unexpected token costs, and leverage new skill auto-reloading.

Learn how to use the new --settings flag added to claude remote-control in version 2.1.321 to configure custom profiles and safeguards for remote sessions. We also cover essential caveats, model updates, and critical sandbox and MCP OAuth security patches.

Explore the new /claude-api latency-optimize command in Claude Code 2.1.319, which tests faster API settings against your evals to reduce latency bottlenecks.

In Claude Code 2.1.319, auto mode removes its automatic client-side fallback when server safety checks fail, meaning headless pipelines behind gateways can now fail closed with exit code one. Learn how to diagnose your setup using /status, configure the CLAUDE_CODE_AUTO_MODE_SERVER environment variable, and prepare your CI workflows.

Discover how Claude Code 2.1.319 prevents unattended agents from stalling with the new permissions.denyInsteadOfPrompting setting. Learn how to configure it properly alongside explicit allow rules, and catch up on smaller quality-of-life additions like background MCP tasks.

Learn how Claude Code 2.1.318 expands deniedMcpServers URL matching to block Anthropic-provided servers and cloud connectors. We cover configuring organizational denylists, verifying policies with claude mcp list, and several smaller MCP quality-of-life fixes.

Discover how Claude Code 2.1.318 lets you pass a custom effort parameter to $.model.fork, cutting down latency and token burn on lightweight mod tasks like handoff briefs and summaries. We also break down other key updates in this release, including cross-account IAM role assumption on the apps gateway and a fix for false-success exits in non-interactive mode.

Learn how Claude Code 2.1.318 prevents unattended CI and headless runs from failing silently due to expired logins. We cover new expiration fields in claude auth status, warnings in claude -p, and switching to setup-token for permanent machine credentials.

Learn how Claude Code version 2.1.317 introduces onFailure: "block" for programmatic Agent SDK hook callbacks, preventing unattended agents from failing open on unexpected errors. We also cover protections against account email leakage in tools, isolated transcripts for nested headless sessions, and full-file re-reading behavior.

Learn how to use Claude Code's blockWritesOutsideWorkingDirectories setting to lock down file write boundaries during unattended or headless runs across every permission mode.