Pentagon Fallout, Model Picker Overhaul, and Memory Leak Fixes

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of March second. The model picker gets a readability overhaul. A sweep of memory leak fixes targets long-running sessions. And the Anthropic-Pentagon standoff reshapes the company's week.

Ten releases for Claude Code this week, headlined by a model picker that finally speaks your language -- showing readable names like Sonnet four point five instead of raw model ID strings. The picker also flags when a newer version of your pinned model is available, so you're not unknowingly running something outdated. And the model entry in the command menu now displays which model is currently active, so you can check at a glance without opening anything. It's a small change on paper, but if you switch models often, it removes a layer of friction you probably didn't realize you were working around.

Session management got some attention too. You can now rename and delete individual sessions directly from the sessions list -- basic housekeeping that keeps your conversation history from turning into a wall of untitled entries. And on a related note, copying responses is now smoother. If you've found the code block picker that appears when you copy a response more annoying than helpful, there's now an option to always copy the full response.

Set it once and every future copy skips the picker entirely. Permissions handling also got smarter this week. When you approve a compound shell command -- something like changing directories, then fetching, then pushing with git -- Claude Code now computes separate approval prefixes for each subcommand instead of treating the entire chain as one rule. That means future approvals are more granular and less likely to block legitimate variations of commands you've already approved.

And for enterprise and IT administrators, managed settings can now be distributed through native OS mechanisms -- macOS plist files or the Windows Registry -- rather than relying solely on config files. That's a meaningful step for organizations that manage developer tooling at scale.

Under the hood, the theme this week is stability across the board. A broad sweep of memory leaks was addressed, targeting unbounded growth in several internal caches -- git root detection, JSON parsing, bash prefixes, and MCP fetch caches all had fixes. Long-running multi-agent sessions now release completed subagent state and strip heavy progress payloads during compaction, which should meaningfully reduce memory pressure if you tend to keep sessions open for hours. That stability push extends to Windows and Linux too, where several crash and corruption issues were resolved.

Running multiple Claude Code instances on Windows could corrupt your config file and wipe authentication -- both cases are now patched. Additional crashes affecting Windows on ARM sixty-four, the WebAssembly interpreter on Linux and Windows, and shell tool errors on Windows are also fixed.

Beyond the product, this was an extraordinary week for Anthropic as a company. The New York Times published a detailed account of how contract negotiations between Anthropic and the Pentagon collapsed over AI safeguards. The Defense Department wanted safety constraints removed; Anthropic refused. The dispute escalated to a government blacklist and Anthropic filing suit -- and as the BBC reports, the White House followed up by directing federal agencies to stop using Claude entirely.

That order has direct implications for enterprise customers operating in or adjacent to the federal government. But here's where the story takes an unexpected turn. CNBC reports that in the days following the Pentagon standoff, Claude surged past ChatGPT to become the number one free app on the US App Store -- an unusual case where a policy dispute drove a measurable spike in consumer adoption. The Wall Street Journal frames all of this as a broader question about who controls AI development and on what terms.

We'll link all four pieces in the show notes. On the engineering side, Anthropic's team published a blog post on quantifying infrastructure noise in agentic coding evals. The key finding is that infrastructure configuration choices -- not just model quality -- can shift benchmark scores by several percentage points, sometimes exceeding the gap between top-ranked models. If you run your own evals or interpret leaderboard results, it's well worth a look before drawing conclusions.

That's it for the brief. I'm Shannon, and we'll see you next week.