The release cadence for Claude Code stays high. After 2.1.224 with its self-hosted environments, two more versions land right behind it — and 2.1.225 carries more than the plain version number suggests.
Spend limits that actually tell you something
The highlight of 2.1.225: gateway spend limits are now baked into the usage warning. When you hit your limit, the message no longer just says a dry “limit reached” — it names the actual cap, when it resets, and the message from your operator. The catch is that the gateway itself has to be running 2.1.225. For teams that route Claude Code through a central gateway, that’s a real quality-of-life win.
There’s also a workspace trust prompt for claude agents. Open an untrusted directory with it, and Claude Code now asks first — exactly like the regular claude command has done for a while. A small but consistent safeguard.
The fixes that save your day
The rest of 2.1.225 is a long list of repairs, and a few of them probably rescued someone’s afternoon:
- A transient 401 replaced a long-lived
CLAUDE_CODE_OAUTH_TOKENwith a short-lived login token — breaking headless sessions until a restart. Fixed. - MCP OAuth servers on macOS sometimes failed with a burst of 401s, as if never authenticated — caused by a keychain read timing out. Fixed.
- Auto Mode was wrongly counting a safety-filter refusal of its own permission check toward the block limit. The action stays denied, but Claude is now told to move on instead of stubbornly retrying.
- Cross-session messages sat parked in headless sessions with no notice and no expiry.
- And conversation history broke on a Remote Control resume after very large histories had been compacted.
On top of that, SendMessage can now start a conversation with your Remote Control sessions on other machines directly by name — instead of waiting for the other side to message you first.
2.1.226, by contrast, is a classic cleanup release: “bug fixes and reliability improvements,” and that’s all the changelog says.
My take
Releases like this are unspectacular — and that’s exactly the point. No big feature fireworks, just hundreds of small things you only notice when they’re missing. The trust prompt for claude agents and the gateway limits show where this is heading: Claude Code is turning into infrastructure that teams want to control and secure centrally. And anyone who’s ever lost a headless session to a token error knows to appreciate these quiet fixes.