Claude Code v2.1.152 Ships Auto Mode to All Users, Adds Skill Controls and Message Hooks

Claude Code shipped version 2.1.152 on May 27, its most feature-dense release since the Code with Claude developer conference in San Francisco earlier this month. The headline change is a quiet but consequential policy shift: auto mode, Anthropic's hands-off coding mode that lets the tool execute multi-step tasks without manual confirmation, is now active for all subscribers by default. No opt-in screen, no setting to locate. If you open Claude Code today, auto mode is running.

Auto Mode Leaves Preview

When Anthropic introduced auto mode in March, it launched with an explicit opt-in. The friction was deliberate: users needed to understand what they were enabling before the tool started taking actions on their behalf. The mode works by running a second Claude instance alongside the primary model. Before each action, whether writing a file, running a bash command, or spawning a background agent, a Sonnet 4.6 safety classifier scores the operation's risk. Low-blast-radius work moves forward without interruption. Higher-risk operations are challenged or blocked outright.

The 2.1.152 release does not change that architecture. The classifier still runs, the approval gates are still in place. What changed is Anthropic's confidence in the system. By removing the opt-in requirement, the company is signaling that auto mode has cleared its internal bar for safe default behavior across the full subscriber base, from individual developers on Pro to enterprise teams running large custom configurations. The decision tracks the broader arc of Claude Code's expansion into agent-oriented workflows.

Smarter Skills With Built-In Guardrails

The update also advances the skill system Anthropic introduced at Code with Claude. Skills are reusable, shareable task packages that extend Claude Code's default behavior. A security audit skill might layer in specific analysis tooling. A documentation skill might constrain the model to reading and summarizing, never writing files. Starting with 2.1.152, skills can declare disallowed-tools in their frontmatter, which removes the listed tools from the model's available palette for the duration of any session where that skill is active.

For enterprise teams distributing skill packages to large developer pools, this changes the risk calculus significantly. An organization can now ship a Claude Code configuration scoped to exactly the actions its security policy permits, with the restrictions encoded in the skill itself rather than dependent on each user managing individual settings. A companion command, /reload-skills, lets teams push updates to skill directories without requiring a full restart. SessionStart hooks can also return reloadSkills: true to trigger a refresh automatically at the beginning of each new session.

"Skills and slash commands can now set disallowed-tools in frontmatter to remove tools from the model for that session, and you can reload skills with /reload-skills without restarting." Anthropic Claude Code changelog, May 27, 2026

New Hooks and the Review-to-Fix Pipeline

The MessageDisplay hook event gives operators a new interception point in the Claude Code pipeline. When registered, this hook can inspect, transform, or suppress the text Claude would otherwise send to the user. For enterprise deployments that need response filtering before output reaches end users, this is a meaningful addition to the hook system already covering session start, tool use, and output handling. The pluginSuggestionMarketplaces managed setting, also added in this release, allows enterprise administrators to allowlist specific plugin sources for their organizations, so developers see only vetted extensions.

The code review workflow got a practical upgrade. /code-review --fix now applies its findings directly to the working tree, replacing the previous gap between identifying an issue and manually applying a fix. /simplify is folded in as an alias for the same command. Vim mode users gain / in NORMAL mode for reverse history search. The /usage command now streams large session files and breaks down token consumption by category, giving teams a cleaner view of where compute is going across skills, subagents, plugins, and MCP servers. Thinking summaries now render as markdown and remain visible for at least three seconds, capped at ten lines.

A More Autonomous Claude Code

Version 2.1.152 is the clearest evidence yet that Anthropic is building Claude Code as an autonomous coding agent first and an interactive assistant second. Auto mode at general availability is the centerpiece, but the skill guardrails and message hooks carry the longer story: Anthropic is constructing the infrastructure that makes enterprise-scale autonomy operationally viable. Teams can now define precisely what Claude Code can and cannot do on their systems, encode those definitions as distributable skills, and trust the defaults will hold at the model level.

Fourteen Claude Code updates shipped in May alone. If 2.1.152 is any signal, the direction is toward a system that handles more repetitive, bounded software development work on its own, flagging a human only when something genuinely requires a decision beyond its sanctioned scope. What started as an autonomous coding assistant is becoming an enterprise operations layer.