Claude Code 2.1.234: sessions resume on their own, permission prompts get hardened

Claude Code 2.1.234 shipped on August 17. There's one real behavior change worth knowing about, automatic session resume when your usage limit resets, plus a cluster of permission-prompt fixes worth reading if you run subagents or lean on multiple MCP servers. Here's what actually matters.

Your session resumes on its own when the usage limit resets

Until now, once you hit your claude.ai usage limit, the session just sat there until you resumed it yourself. Starting with 2.1.234, Claude Code resumes automatically as soon as the limit resets. Handy if you kick off a long task before stepping away, less handy if you were counting on that pause to check in before letting Claude keep going.

This is on by default. To go back to the old behavior, turn off "Continue automatically at usage limit" in /config.

A real batch of permission and secret-handling fixes

None of these look dramatic on their own, but together they're worth the update:

• Credential masking on relayed permission previews: masking applied to permission previews sent to a remote approver could hide the actual command, path, or destination being approved. Fixed, the approver now sees what they're actually approving.
• MCP diagnostics leaking resolved secrets: scope-conflict warnings now show the configured ${VAR} form instead of the resolved value, and connection-failure details only show the server's origin.
• Plugin marketplace allowlist bypass: an SCP-style git source could pass a strictKnownMarketplaces allowlist even when its real host differed from the one git actually connected to. Closed.
• NTLM credential leak on Windows: a path prefixed with an NT device namespace (\??\) could dodge the usual path validation on remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads.
• Dropped permission answers: a session-scoped answer, including a denial, could get lost when it was answering a permission prompt triggered by a background subagent.
• Stale answer on a re-prompt: if an IDE diff tab closed while a new permission prompt appeared, that new prompt could inherit the previous answer instead of waiting for a real one.

None of this needs action beyond updating, but if you approve permission previews remotely for what an agent is doing, that first fix alone is worth the upgrade.

Two habits that quietly changed

Esc no longer clears a text selection in fullscreen mode: it still interrupts or dismisses as usual, but a highlighted selection now stays highlighted instead of disappearing. A new selection:clear keybinding action takes over if you want a dedicated key for that.

The "Default teammate model" setting is gone from /config. An agent-team teammate now uses the leader's model by default unless the spawn names one explicitly. If you had that setting configured, it no longer applies.

Everything else, worth a skim

• Compaction recovery: auto mode in very long sessions no longer keeps re-asking for network access on a sandboxed command it already approved before compaction.
• claude setup-token now rejects unexpected extra arguments instead of silently ignoring them.
• /permissions and /add-dir can now open while Claude is working, and rule changes apply for the rest of the current turn.
• /goal clears itself with a notice when a turn dies on an unrecoverable error, and now checks in when a background task keeps it waiting more than 30 minutes (CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0 to opt out).
• The claude-api skill now costs roughly 25,000 tokens to load instead of 200,000-plus, reference docs load on demand.
• GitLab: a merge request badge now shows in the footer and statusline for repos with a GitLab remote and an authenticated glab CLI.
• Remote Control: signing this computer into a different claude.ai account now stops the running session within seconds with a reason shown, instead of a silent 404 hours later.
• Auto-generated session titles are shorter and more specific now ("Login button bug" instead of a full restatement of your request).

Should you update

Yes. The permission and credential-masking fixes alone justify it, especially if you approve actions remotely or run background subagents. Just double-check whether automatic session resume at the usage limit works for you, if not, turn it off in /config before you step away.

Source: official Claude Code changelog.