Issue #4
4 min read

Anthropic and the Pentagon split

The /model picker now displays readable names and flags outdated pinned models, while session management adds rename/delete controls and /copy gains a full-response shortcut. Anthropic's Pentagon contract collapsed over AI safeguard disputes, prompting the White House to redirect agencies off Claude—yet the App Store briefly ranked Claude above ChatGPT.

How Talks Between Anthropic and the Defense Dept. Fell Apart — New York Times

On this page

Highlight

Model picker now shows readable names and your active model

The /model picker now displays human-readable labels like "Sonnet 4.5" instead of raw model ID strings, and flags when a newer version of your pinned model is available. The /model entry in the slash command menu also now shows which model is currently active, so you always know what you're running without opening the picker.

Worth Knowing

You can now rename and delete sessions from the sessions list

The sessions list now lets you rename and remove individual sessions directly, giving you basic housekeeping over your conversation history without leaving the UI.

New option to always copy full responses with /copy

If you find the /copy code block picker more friction than help, you can now select "Always copy full response" once and all future /copy invocations will skip the picker entirely and copy the complete response.

Smarter permission prefixes for chained shell commands

When you approve a compound command like cd /tmp && git fetch && git push, Claude Code now computes per-subcommand "always allow" prefixes instead of treating the whole string as a single rule, so future approvals are more granular and less likely to block legitimate variations.

New env var to disable claude.ai MCP servers

Set ENABLE_CLAUDEAI_MCP_SERVERS=false if you want to prevent claude.ai MCP servers from being surfaced in your environment — useful for controlled or enterprise setups where you manage MCP availability explicitly.

Managed settings now configurable via macOS plist or Windows Registry

Enterprise and IT administrators can now distribute Claude Code managed settings through native OS mechanisms — macOS plist files or the Windows Registry — rather than relying solely on config files. See the settings docs for details.

Under the Hood

Broad memory leak fixes across long-running and multi-agent sessions

A sweep of memory leaks was addressed this week: unbounded growth in the git root detection cache, JSON parsing cache, bash prefix cache, and MCP fetch caches are all fixed, and long-running multi-agent sessions now release completed subagent state and strip heavy progress payloads during compaction — all reducing memory pressure in extended sessions.

Config corruption and crash fixes on Windows and Linux

Several stability issues are resolved: concurrent Claude Code instances on Windows could corrupt your config file and wipe authentication — both cases are now fixed. Additional crashes affecting Windows ARM64, the WebAssembly interpreter on Linux and Windows x64, and BashTool on Windows with EINVAL errors are also patched.

From Anthropic

Anthropic Engineering · 8 min read

Quantifying infrastructure noise in agentic coding evals

Anthropic's engineering team shows that infrastructure configuration choices — not just model quality — can shift agentic coding benchmark scores by several percentage points, sometimes exceeding the gap between top-ranked models. If you run your own evals or interpret leaderboard results, this is worth reading before drawing conclusions about model comparisons.

In the News

The Anthropic-Pentagon dispute dominated headlines this week, but the product kept shipping — session management, permission handling, and memory stability all saw meaningful improvements.