Six releases landed this week, 2.1.221 through 2.1.226, across five days – after a week whose changelog stayed empty from Monday to Sunday. The change that will reach the most people didn’t come via the changelog at all: Anthropic announced it in a blog post on 7 August, and it takes effect on 14 August.
Auto mode becomes the default on 14 August
From 14 August, new Claude Code sessions on the Pro, Max and Team plans start in the Auto permission mode rather than Manual. Manual is the mode that halts before every file change and every command and waits for you to confirm. Auto has none of those prompts. Before anything runs, a separate model looks at it, and its job is to stop actions that are irreversible, destructive, or that reach outside your own environment. If your own rules say Claude Code has to ask before doing something, it still stops.
Anthropic attaches a caveat in the documentation, and it is worth repeating rather than skimming: auto mode cuts the number of prompts, but it is not a guarantee of safety, and it doesn’t take the place of reading what Claude Code actually did.
The case for the switch is a study of more than a thousand paying users. The reviewing model stopped 89 percent of dangerous commands. The people given those same commands to approve by hand turned them down in 13.6 percent of cases. Anthropic adds a second figure from outside the test itself: in real user sessions, manually approved actions ended in serious unintended harm more than twice as often as actions in auto mode.
That gap matches something we see in the people we train after a few days of work. A permission prompt does its job only for as long as it gets read. After that it turns into one more key between you and the result, your finger finds it on its own, and Claude Code goes ahead and does what it was going to do anyway. We suspect the change follows from how people actually work rather than from how they would work if every prompt were read to the end.
What this means for you depends on where you are. Been at this a while, and catching yourself approving steps without looking? The change moves in your direction and there is nothing for you to do. Just starting out? We’d stay on Manual for the first few weeks, and after 14 August that takes a deliberate move on your part, because newly opened sessions won’t return to it by themselves. It isn’t that auto mode is dangerous. It is that watching every command early on is how you learn what Claude Code reaches for, and in what order – and no amount of reading afterwards makes up for it. Switching permission modes is quick whenever you want it: Shift+Tab cycles the modes in the terminal, the desktop app has a dropdown by the input box, and the VS Code extension puts a mode indicator below it. If you have already set a default of your own, a one-time prompt asks whether to move to auto, and “no” is a real answer there. Our tip about settings goes through the modes properly, including how Don’t ask differs from Accept edits.
One related change is already live, as of 7 August. The reviewing model’s own work no longer counts against your plan’s usage limits on Pro, Max and Team. It used to, and that was a genuine obstacle – in auto mode you were spending your own limit on Claude Code checking itself, which made the extra safeguard cost more than working without it.
Focus view hides tool activity in VS Code
Version 2.1.221, out on 4 August, added Focus view to the VS Code extension: a toggle in the chat menu that folds everything Claude Code does along the way into one expandable summary per turn, reached with Ctrl+Alt+F or the “Claude Code: Toggle Focus view” command. Instead of the full record of every file read and every command run, you get a line telling you how much of it there was, plus an indicator for the tool currently working. It tackles the same problem we solved by switching Verbose output off in our tip about settings, and it tackles it from the better end – Verbose governs how much the tools hand back in the first place, while Focus view keeps all of it and merely folds it out of sight until you want a look. If you are finding your way around Claude Code, that beats switching information off, because the screen calms down and every step remains there to check.
Turn it on now and you get the fixed version straight away. Its first release also folded away the to-do list and the context of a question Claude Code was waiting on, which were the two things that should have stayed in plain sight; version 2.1.225 on 8 August sorted that out.
/review and /code-review are now one command
Three weeks ago we wrote that code review stopped invoking itself and now runs only when you ask for it. Version 2.1.223, out on 6 August, took the next step and made /review an alias for /code-review, so both names lead to the same review – either of the current changes in your folder, or of a pull request you name, called as /code-review <level> <PR number>.
The two changes fit together well. A review costs time and tokens, the decision to start one is yours, and now at least you don’t have to remember which of two names is the right one. Separate from both sits /code-review ultra, a deep review run in the cloud by many agents at once.
Claude Code sessions can message each other
Version 2.1.224, out on 7 August, added something that sounds like a team feature and is really for one person with several windows open. Claude Code sessions can now send messages to one another, across your own machines as well, and the ListAgents command shows which sessions are within reach. It works on macOS and on Linux. Alongside it came settings for what happens to such a message before it arrives – when the receiving session should check with you first, and how long the message waits before it expires.
Taken one at a time, this week’s changes are just conveniences. Together they shift where your attention goes: from “should I allow this command” to “is what came out of it the thing I asked for”. The second question is the harder one, nobody will answer it for you, and we’re going to spend the next few weeks in auto mode ourselves writing down what goes wrong along the way. An announcement won’t tell us how it behaves day to day.