Rewind Crosses /clear, Mythos Returns to US Orgs

I'm Shannon, and this is the Claude Notes Brief -- Claude Code updates and Anthropic news for the week of June twenty-ninth. Rewind now recovers context from before you cleared the conversation. Bash mode gets a couple of useful upgrades. And Mythos returns to a vetted list of US organizations.

Let's start with Claude Code -- the headline this week is that rewind now works across clear, which fundamentally changes the tradeoff you've been making mid-session. Previously, if you cleared the conversation to refocus Claude or free up the context window, you gave up the ability to step back into anything you'd done before. That's gone. You can now rewind through a clear and recover the earlier context, which means clearing is no longer a one-way door.

If you've been using clear defensively to stretch context, the team is suggesting you try rewind first. That same theme of making Claude more responsive to how you actually work shows up in bash mode, which picked up two changes. When you run a bash command with the bang prefix, Claude now automatically responds to the output instead of just quietly adding it to context. If you preferred the old behavior of using bang purely to feed context without a response, there's a setting to turn the new behavior off.

Alongside that, bash mode now offers live file path completion as you type, matching what you already get in the main prompt. Useful if you frequently chain commands against paths you don't want to retype. On the MCP side, you can now log in and out of MCP servers directly from the command line without opening the interactive menu, and there's a no-browser flag that redirects OAuth through standard input. That last piece matters if you manage MCP servers over SSH or in scripts where a browser-based flow just isn't going to work.

Background subagents also got a meaningful behavior change. They no longer auto-deny tool requests that need permission. Instead, the prompt surfaces in your main session showing exactly which agent is asking, and pressing escape denies just that single tool call rather than killing the agent entirely. And on the security side, there's a new sandbox setting that blocks sandboxed commands from reading credential files and secret environment variables.

Separately, if your organization restricts which models you can use, those restrictions now appear directly in the model picker with a clear message when you select one that's blocked.

Moving under the hood -- performance and stability work this week centered on long-running sessions. Streaming responses now batch text updates every hundred milliseconds, which cuts CPU usage by roughly thirty-seven percent during a stream. The terminal output cache was also reworked so it no longer drives memory growth over long sessions. If you keep Claude Code open all day, both of these compound.

On the stability side, hook matchers got a round of fixes. Hyphenated names like code-reviewer now match exactly instead of partially, and comma-separated matchers actually fire on each entry. Stopped background agents now stay stopped, and the Visual Studio Code extension no longer hangs when you resume a large session.

On to the broader news. Wired is reporting that the Trump administration has agreed to let Anthropic release Mythos to a vetted list of US companies and agencies, resolving the negotiations we covered last week. If your team's model access was affected by the earlier restrictions, this is the resolution to track. That story connects directly to a piece in the Wall Street Journal arguing that Chinese labs have closed the gap with Anthropic on cybersecurity-relevant capabilities, including the ones that triggered the Mythos restrictions in the first place.

It's useful framing for understanding why the export and access debates are escalating right now. And in a related thread, the Journal is also reporting that Anthropic has accused Alibaba of running a sustained campaign to extract Claude's capabilities through distillation. The dispute is reshaping how Anthropic talks about model access controls and downstream policy for API customers. On the engineering blog, Anthropic published a piece on building effective human-agent teams, with patterns drawn from real workflows -- relevant if you've been experimenting with the agent teams features that have landed recently.

The team also introduced Claude Tag, a Slack-based way to bring Claude into shared channels with a new agent identity access model. We'll link both in the show notes. That's it for the brief. I'm Shannon, and we'll see you next week.