Project Purge, Gateway Models, and a $900B Valuation

I'm Shannon, and this is the Claude Notes Brief -- Claude Code updates and Anthropic news for the week of May fourth. A new command for clearing project state. OAuth login now works in WSL2, SSH, and containers. And Anthropic is reportedly weighing funding at a nine hundred billion dollar valuation.

Let's start with Claude Code -- the headline this week is a new way to wipe everything Claude knows about a project, which fills a gap that's been awkward for a while. The new project purge command deletes all Claude Code state for a given project: transcripts, tasks, file history, and the config entry itself. It supports a dry-run mode, an interactive mode, and a yes-flag for scripting, so you have a clean way to remove residual state before you archive a repository or hand it off to someone else. That focus on cleanup pairs nicely with a friction point that just got resolved on the authentication side.

OAuth login now works in WSL2, over SSH, and inside containers. Previously, if the browser callback couldn't reach localhost, login would just hang. Now you can paste the OAuth code directly into the terminal, which is a small change with an outsized impact for anyone who works in remote or sandboxed environments. Staying on the configuration theme, the model picker is now smarter about gateways.

If you point your base URL at an Anthropic-compatible gateway, the picker pulls its list from that gateway's models endpoint. So if your team runs Claude Code through an internal proxy, the available models just show up -- no manual configuration. There's also a workflow improvement worth mentioning. You can now paste a pull request URL from GitHub, GitHub Enterprise, GitLab, or Bitbucket into the resume search, and Claude Code will find the session that created it.

It's a faster way to pick up where you left off on a specific PR without scrolling through history. And for Windows developers: when the PowerShell tool is enabled, Claude now treats PowerShell as the primary shell instead of defaulting to Bash. Detection also expanded to cover PowerShell 7 installed through the Microsoft Store, through MSI without a PATH entry, or as a .NET global tool.

Moving under the hood -- this was a big week for memory and session reliability, and these are the changes you'll probably feel most in long sessions. On the memory side, the team fixed several multi-gigabyte leaks. Unbounded memory growth when processing many images is resolved. The usage command was leaking up to two gigabytes on machines with large transcript histories -- that's fixed.

And long-running tools that fail to emit progress events no longer leak. Pasting an oversized image also won't break your session anymore, since images are now downscaled on paste. Session reliability got similar attention. Stream idle timeouts after a Mac goes to sleep, or during long model thinking pauses, are fixed.

The resume flag now skips corrupt transcript lines instead of failing outright. And the Bash tool no longer becomes permanently unusable if the startup directory gets deleted or moved mid-session. On the reading list, Anthropic Engineering published two posts worth your time. One covers prompt caching lessons from building Claude Code -- directly relevant if you're tuning your own agent loops.

The other, from Brendan MacLean, walks through onboarding Claude Code to a seven hundred thousand line codebase using the same playbook used for new engineers. We'll link both in the show notes.

Now to the broader news, and there's a lot moving this week. The New York Times is reporting that the Pentagon signed agreements with seven major AI companies to expand classified work -- and Anthropic was not among them. The ongoing dispute over access to Anthropic's Mythos model continues to shape which models get deployed in U.S. defense contexts. That story connects directly to a Bloomberg piece on why Mythos is drawing scrutiny from governments and central banks around the world.

The Bundesbank has called for EU access, and the article is useful background on the policy fights influencing Anthropic's product decisions right now. On the financial side, Bloomberg is also reporting that Anthropic is weighing a new funding round that would value the company at over nine hundred billion dollars, topping OpenAI. It's relevant context for the scale of capital being committed to the infrastructure Claude Code runs on. And one more story worth flagging from The Guardian: a Claude-powered agent reportedly deleted a firm's entire database, then explained in detail why it shouldn't have.

A pointed reminder to keep destructive permissions tightly scoped when you're running agents against real infrastructure. That's it for the brief. I'm Shannon, and we'll see you next week.