Four things changed in Claude Code this week and three activated without asking you
There’s a specific hour that disappears from your week and doesn’t show up in any metric. You pass something big to Claude Code, your usage window gets consumed at 14:40, and the session stops mid-task. The limit resets at 15:00. You return to your desk at 16:15. There was no one to press Enter, so that hour went unused while everything needed to continue was already available.
Two releases landed on August 17 and 18 — 2.1.234 and 2.1.235. Between them are about eighty lines, and almost all of them are fixes. Four items change how the terminal behaves day to day, and three of them are already active on your machine, whether you asked for them or not.
What activated on its own: the session waits for the limit to reset
The changelog line is exactly this:
“Claude Code now continues your session automatically when a claude.ai usage limit resets”
That’s it. The session stays there, waits for the reset, and picks up the task on its own.
It comes turned on. What you do in /config is turn it off — the row is called “Continue automatically at usage limit”.
Two things worth knowing before you leave it alone:
It resumes without supervision, with the permissions the session had. If you started with --dangerously-skip-permissions or left auto mode running, that’s the state it returns to. The reset doesn’t ask you again. If you’re used to launching long sessions and walking away, this is the setting worth spending thirty seconds on, not the one to ignore.
Cowork and Desktop are another surface. They launched their own version of this, also turned on by default, but the toggle lives in the limit banner and applies to your whole account — not in /config. If you turn it off in the terminal and expect the desktop app to follow suit, it won’t.
And an honest note about the state of the feature: the Cowork changelog already registers a fix for it, published after 2.1.234. Automatic continuation was firing against the imported session confirmation prompt or against an expired sign-in state, instead of waiting correctly. Version one shipped with rough edges. They’re being smoothed out in public, which is the normal way for a feature like this, but it means “enable and forget” is a decision for next month, not today.
What activated on its own: the GitLab merge request badge
The footer and statusline now carry an MR badge. If the repository has a GitLab remote and glab is authenticated, you get MR !N with its status: draft, pending, or green.
There’s no setting to enable. It appears when both conditions are met, which is exactly why it counts as something that activated on its own.
Step 1 — install glab.
# macOS
brew install glab
# Debian/Ubuntu
sudo apt install glab
Step 2 — authenticate.
glab auth login
For GitLab self-managed, point it to your instance during the prompts instead of gitlab.com.
Step 3 — confirm it stuck.
glab auth status
Step 4 — verify the remote is really GitLab.
git remote -v
If origin points elsewhere, the badge won’t render no matter how well glab is configured.
Step 5 — start a session in that repository. The badge appears in the footer and statusline as soon as there’s an open MR on the branch.
This is the third GitLab item in as many releases. 2.1.233 added merge request URL support to the --worktree flag and the agents view, along with token redaction for GitLab token families. The pattern is pretty clear: GitHub parity debt is being paid down.
What you activate: spellcheck
This one, from 2.1.235, is genuinely opt-in. Misspelled words get underlined in the prompt input as you type.
It doesn’t ship with a dictionary included. It uses the aspell, hunspell, or ispell you already have installed — so if none is on the machine, enabling the setting doesn’t do anything visible.
# macOS
brew install hunspell
# Debian/Ubuntu
sudo apt install hunspell hunspell-en
Then enable the spellcheck setting. Open /config and search for the row instead of writing the JSON by hand — the changelog names the setting but doesn’t publish the exact spelling of its value, and /config will show you the real one.
Small feature. Matters more than it sounds if you write prompts in a second language and prefer to catch the typo before it becomes an ambiguous instruction.
The one nobody put in the headline
Buried in the same release: /permissions and /add-dir can now be opened while Claude is working, and rule changes apply to the rest of the current turn. In the fullscreen TUI, /autocompact, /theme, /help, /config, and /advisor also open mid-turn.
This is the item that changes the actual workflow. The old play was interrupt the turn, fix the permission, and re-explain what you wanted. Now you adjust the rule and the turn continues with it in place.
In the same category: the built-in claude-api skill went from over 200,000 tokens of context to about 25,000, loading its reference docs on demand. If that skill was in your setup, you recovered a good chunk of a context window without doing anything.
Why update even if none of this is for you
Two items you’re never going to see and that you should still want.
Path hardening on Windows. 2.1.233 closed a vector for NTLM credential leakage where Windows paths written with the NT device prefix \??\ were slipping through UNC path validation. 2.1.234 extends that rejection to file accesses that happen before approval: remote reads, session restore, CLAUDE.md includes, workflow scripts, and uploads. Worth being precise here, because it’s been reported imprecisely — the vector closed in .233. What .234 does is cover the ground that was left.
The prompt cache fix. In 2.1.235, a language server disconnecting or reconnecting mid-session no longer invalidates the entire prompt cache. If you work in a repository where the LSP crashes once in a while, you were paying to rebuild the entire cached prefix every time it happened. That’s silent money, and it’s done.
What to do this afternoon
Update. Then spend two minutes on exactly one decision: whether you want your sessions to resume without you when the window opens again. Everything else here is pure gain — but that changes what your machine does while you’re not looking at it, and it changed by default.
Do you want your agent to pick up the task on its own at 3 a.m., or is that a line you prefer to keep?