Anthropic Dispatch: Claude Code 2.1.152 Ships Today and Skills Just Got Teeth

The One Where Code Review Learns to Fix Things

May twenty-seventh, twenty twenty-six. Claude Code two point one point one fifty-two just dropped, and the headliner is a command that quietly became something much more useful. Slash code-review now takes a dash dash fix flag. What that means is the review does not just tell you what is wrong anymore. It applies the findings to your working tree. It surfaces reuse opportunities, simplification, and efficiency suggestions, and then it actually makes the changes. Slash simplify now just invokes slash code-review dash dash fix under the hood.

Think about what this means for a daily workflow. You finish a feature. You run slash code-review dash dash fix. It reads your changes, identifies the patterns you missed, the duplication you introduced because you were moving fast, the thing that could be three lines instead of twelve, and it rewrites them. You review the diff. Accept or reject. Done. That loop used to be a separate conversation where you asked Claude to look at your code and then manually applied its suggestions. Now it is one command.

Skills Get a Kill Switch

The second big change is that skills and slash commands can now set disallowed-tools in their frontmatter to remove tools from the model while the skill is active. This is subtle but powerful. If you have a skill that should only read files and never write them, you can now enforce that at the skill level. The model literally cannot access the tools you block.

There is also a new slash reload-skills command that re-scans skill directories without restarting the session. And SessionStart hooks can return reloadSkills true to trigger the same thing automatically. So if your hook installs a skill on session start, it is available immediately instead of requiring a restart. This is the kind of infrastructure that matters when you are building automated pipelines. Skills that install themselves, configure their own tool access, and start working without human intervention.

Auto Mode Drops the Consent Dialog

Here is one that changes the default experience. Auto mode no longer requires opt-in consent. That dialog box that appeared every time asking if you wanted to enable auto mode is gone. If you have been running Claude Code daily, you have clicked through that consent dialog hundreds of times. It was there for good reason during the early days, but at this point anyone using auto mode knows what it does. Removing the friction is the right call.

Hooks Can Now Transform What You See

A new MessageDisplay hook event lets hooks transform or hide assistant message text as it is displayed. This is a display-layer hook, not a content hook. The model's actual output is unchanged, but what renders in your terminal can be filtered, reformatted, or suppressed entirely by a hook.

The use cases are interesting. You could build a hook that strips verbose reasoning from display while keeping it in the transcript. You could highlight specific patterns in the output. You could suppress internal tool-coordination messages that are useful for the model but noise for the human. Combined with the existing hook system, this means the entire Claude Code interface is now programmable at every stage: input, processing, and output.

The Vim Mode Easter Egg

If you use vim keybindings, slash in normal mode now opens reverse history search, matching how bash and zsh vi-mode work. This is the kind of detail that tells you someone on the Claude Code team actually uses vim bindings daily and got annoyed that it did not work the way their muscle memory expected.

The Bug Fix Graveyard

The bug fixes in this release are extensive and some of them are genuinely interesting. Terminal styling degrading in very long sessions got fixed by recycling the renderer's style pool. That explains the visual corruption people reported during marathon coding sessions. Plugin MCP servers with the same command but different environment variables were being incorrectly deduplicated, which means some people were running the wrong server without knowing it. And there is a fix for sessions getting stuck after a model or login switch left stale thinking-block signatures in history.

That last one is worth dwelling on. If you switched models mid-session and things got weird afterward, stale thinking blocks from the previous model were polluting your conversation history. The fix strips them proactively and adds a retry safety net. This is the kind of bug that makes you question everything that happened in a session without realizing why.

What This Means for Headless Automation

If you are thinking about running Claude Code in automated pipelines, this release moves several pieces into place. Skills that control their own tool access. Hooks that install skills dynamically. Hooks that transform display output. Auto mode without consent gates. The fallback model switch that keeps your session alive instead of failing when the primary model is unavailable. Each of these is a small thing. Together they describe a system that can be configured to run unattended with guardrails you define, not guardrails you click through.