Claude Code 2.1.234 auto-resumes after usage-limit reset

Anthropic shipped Claude Code 2.1.234 on August 17, 2026. The change operators will feel first is simple: a long session no longer has to die at the claude.ai usage cap. The CLI can wait the limit out and continue the same session when the window resets.

That is the overnight story. Leave work running, hit the cap, and Claude Code picks the thread back up without you sitting on the terminal. If you do not want unattended restarts, the same release documents an off switch in /config.

The GitHub tag and the official changelog page match. They document auto-continue, a GitLab merge-request badge, a remaining NT-namespace path reject, and Remote Control permission and effort sync. The rest of the ship is a dense reliability pass — permission answers that no longer vanish, a fullscreen-renderer restart that keeps your flags, and a handful of Remote Control file and account-switch fixes.

Auto-continue when the claude.ai cap resets

The notes are explicit. Claude Code now continues your session automatically when a claude.ai usage limit resets. You turn the behavior off in /config under Continue automatically at usage limit.

That is the documented control. The changelog does not describe a separate command, a timer, or a prompt you have to answer at reset. The session waits; when the limit lifts, it continues. Operators who want that for overnight jobs leave the setting on. Operators who do not want the agent to wake itself up unattended open /config and disable it before they walk away.

/config is easier to reach in this build than it used to be. The fullscreen TUI can open /config mid-turn, along with /add-dir, /autocompact, /theme, /help, and /advisor. /permissions can be opened while Claude is working, and rule changes apply to the rest of the current turn. You do not have to drain the session to flip auto-continue or tighten a rule.

A related resume fix matters if you log in with an OAuth token in the environment. After /login while CLAUDE_CODE_OAUTH_TOKEN is set, the stale-token reminder no longer leaks into Claude’s automatically resumed turn. It now appears only to you.

/goal also got cleaner failure behavior in the same release. When a turn dies on an unrecoverable error — revoked auth, an exhausted credit balance, or a context overflow — /goal clears itself with a notice instead of staying armed. Background tasks that keep a goal waiting for 30 or more minutes now get a check-in rather than an indefinite wait; set CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0 to opt out of that check-in.

A GitLab MR badge in the footer and statusline

Repos with a GitLab remote and an authenticated glab CLI now show a merge-request badge in the footer and statusline. The badge is !N, with draft, pending, and green states.

This is the always-on status version of work that started in 2.1.233, which added GitLab merge-request URL support to the --worktree flag and the claude agents view, where MRs already displayed as !N. 2.1.234 puts that state on the chrome you already watch. If the remote is GitLab and glab is signed in, you get the current MR without leaving the session.

If the badge is missing, the notes point at two conditions only: a GitLab remote, and an authenticated glab CLI. There is no extra Claude Code setting listed for it.

Remaining NT-namespace path rejects

2.1.233 closed an NTLM credential-leak vector by stopping Windows paths spelled with the NT \??\ device prefix from bypassing UNC path validation. 2.1.234 hardens the leftover pre-approval file accesses against the same vector.

Remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads now reject Windows NT-namespace (\??\) paths. Those are the remaining reads that could run before an approval prompt. They now fail closed on that prefix.

If Windows tooling still hands Claude Code \??\ paths for a restore, an include, a workflow script, an upload, or a remote read, expect those calls to be rejected. That is the intended security change, not a regression in ordinary C:\ paths.

The same release also tightens what permission previews are allowed to show and hide. Relayed previews go only to channel servers admitted by the inbound trust gate, and a server’s explicit permission-capability opt-out is honored. Credential masking on those previews can no longer hide commands, paths, or destinations from the approver. Oversized private-key blocks redact under full-strength redaction, and provider API tokens that mask on previews now mask even when they sit directly in front of a shell delimiter.

MCP diagnostics stopped printing resolved secrets as well. Scope-conflict warnings now show the configured ${VAR} form, and connection-failure details show only the server origin.

Remote Control syncs permission mode and effort

Remote Control sessions started from Claude Code Desktop or VS Code now keep phones and claude.ai/code updated on the session’s permission mode as it changes. claude.ai/code also stays current on the model. Effort picks made on a phone or on claude.ai/code now apply to terminal- and Desktop/VS Code-hosted sessions, and the session publishes its effort level to connected clients.

In practice, the host and the remotes are no longer free to drift. Change permission mode on the desktop or in VS Code, and the phone and claude.ai/code see it. Change effort on the phone or on the web, and the terminal or desktop host applies it.

Account and organization switches are no longer a slow, confusing failure. Signing the host computer in to a different claude.ai account or organization now stops the running session within seconds and says why, instead of a misleading HTTP 404 hours later.

Files sent to the user during Remote Control sessions hosted by Claude Code Desktop or VS Code now upload, so they open on phone and web instead of showing an empty card. That closes a hole where a share looked successful and then rendered as nothing on the remote client.

Reliability fixes that matter on long sessions

Auto mode in very long sessions no longer repeatedly re-checks and denies sandboxed commands’ network access after the conversation has been compacted. Session-scoped permission answers, including denies, are no longer dropped when you answer background subagent tool permission prompts.

Accepting the “Try the new fullscreen renderer?” prompt no longer restarts the session without its permission mode (for example --dangerously-skip-permissions), tool allow/deny rules, model, or effort flags. /tui no longer drops launch --allowed-tools / --disallowed-tools rules on restart. If the session has restrictions a restart cannot carry over, /tui declines to switch and says why.

A few other notes from the same tag are worth a glance. Hosts that give each session its own config directory can set CLAUDE_CODE_PROJECT_DIR_NAME to choose a short name for the per-project transcript directory. The built-in claude-api skill drops from roughly 200k-plus tokens of context to about 25k by loading reference docs on demand. Auto-generated session titles now read as short, specific names rather than sentences that restate the request. On Windows, startup no longer stalls on repeated rename retries when ~/.claude.json is read-only.

The official changelog page is generated from CHANGELOG.md on GitHub and matches the v2.1.234 release notes. Run claude --version to confirm what is installed.

Next step: Update to 2.1.234, then open /config and set Continue automatically at usage limit to match how you want overnight sessions to behave. If you work in GitLab repos, authenticate glab and check the footer for the !N badge.