Claude Code Update: Recover Sessions and Cut CPU Costs

If you have ever run /clear in Claude Code and immediately wished you had not, last week's update is the one to install.

Anthropic shipped Claude Code v2.1.191 on 24/06/2026, and the headline addition is /rewind, a new slash command that recovers your session state from just before a /clear command was run. For anyone using Claude Code on longer agentic tasks, this closes a genuine gap in the tool.

Until this release, Claude Code's checkpoint system let you roll back code and conversation states by pressing Escape twice. That covered most mistakes but hit a hard limit at /clear, which wiped the in-flight thread entirely. If you cleared too early, you were left choosing between /resume, which works on saved named sessions rather than live ones, or starting over from scratch.

The /rewind command fills that gap. It opens a checkpoint menu showing the state of your code and conversation from just before the clear happened. You can restore either or both independently, so if the code edits were fine but you want the context back, you can take just that. The existing Esc-twice shortcut still handles in-flight rollbacks, /resume continues saved sessions, and /branch lets you fork the conversation at any earlier point. Together they give teams working on long, multi-file tasks a much more resilient safety net.

Performance gains worth noting

Two other changes in v2.1.191 matter for teams running Claude Code at any real scale.

Anthropic reduced streaming CPU usage by roughly 37% by coalescing text updates to 100ms intervals. If you run Claude Code on a laptop or a constrained cloud instance, that reduction translates directly into less heat, less fan noise, and lower compute costs during long agent sessions. The update also reduces memory growth from terminal output caching, which previously became noticeable in sessions lasting more than an hour or two.

Background agent stops now behave correctly. In previous versions, halting an agent from the tasks panel could silently undo itself, occasionally causing the agent to restart and produce duplicate edits or repeated bash commands. From v2.1.191, a stop is permanent. That matters significantly if you have automated overnight pipelines where an agent restarting unchecked could touch files you did not intend.

Hooks and MCP reliability

There was also a long-standing bug that will be familiar to anyone who configured Claude Code hooks: comma-separated matchers in settings.json, for example "Bash,PowerShell", silently never fired. That is now fixed. If you copied hook configurations from community guides and wondered why nothing ran, update first, then audit your settings file and test again.

On the MCP side, the release adds retry logic with backoff for transient connection errors during tool, prompt, and resource list calls, plus better OAuth handling in headless environments. Both changes make Claude Code more stable on slow or patchy networks, which matters for UK teams spread across multiple offices or connecting over mobile.

What this means for UK businesses

Claude Code has become a serious production tool for software teams, not just a developer experiment. Teams using it for automated refactoring, test generation, technical documentation, and multi-step deployments will now lose less work when sessions go wrong, spend less on compute, and get more predictable behaviour from background agents.

Applying the update is straightforward. Run your usual npm or homebrew upgrade, confirm with claude --version that you are on 2.1.191 or later, and your existing sessions benefit immediately. There is nothing to reconfigure.

If you have been holding back on adopting Claude Code because of the risk of losing context during complex sessions, this release directly addresses that concern. The /rewind feature alone removes one of the most frustrating failure modes in the tool, and the reliability fixes around agents and MCP connections make the whole experience noticeably steadier.

At Adevious AI, Claude Code is part of our core build stack. We can help UK businesses introduce it safely, configure hooks and MCP connections for their specific workflows, and keep up with the pace of updates coming from Anthropic. If you would like to talk about whether Claude Code could save your team time and reduce your development overhead, get in touch with us at adevious.co.uk.