Coding with AI

Claude Code Loops Now Last Longer, With a Catch

Claude Code 2.1.330 extends new recurring tasks to 30 days, but older tasks keep seven days and Anthropic's scheduling guide still disagrees.

Illustrative glass calendar pages receding beside a closed laptop, with a glowing marker at the far end.

Anthropic’s November 20, 2026 Claude Code 2.1.330 release notes say that newly created recurring tasks and /loop schedules can last up to 30 days instead of seven, provided their Claude Code session keeps running. Tasks created before this release still expire after seven days. This changes the lifetime of a session schedule, not the requirement to keep that session available. Anthropic’s general scheduling guide still says recurring tasks expire after seven days, so the release and guide currently disagree. We have not tested the behavior in a running client.

What changed, and what did not

The release draws its boundary at task creation. Upgrading the client does not turn a schedule you created earlier into a 30-day schedule, according to Anthropic. A new schedule made with the updated client is the one covered by the longer lifetime. The notes also say Claude Code will tell you when these tasks end, though they do not specify in that line how the expiry is presented in every interface.

Schedule or source Expiry stated by Anthropic What to do
Recurring task or /loop made before 2.1.330 Seven days Check its listed end time; recreate it if you need a later expiry.
New recurring task or /loop made with 2.1.330 Up to 30 days while its session keeps running Confirm the client version and the task’s displayed end time.
Current general scheduling guide Seven days Treat this as a documentation conflict, not proof that the release change is absent.

The table follows the release notes for the version-specific change and the scheduling guide for its conflicting general statement. "Up to" matters: a longer expiry is a ceiling, not a promise that a prompt will run 30 days without interruption. The notes do not report a new price, model or plan for this feature. Nor do they measure completion rates or savings.

Thirty days still depends on the session

Anthropic describes /loop as session-scoped scheduling. Its guide says a scheduled prompt runs between turns while Claude Code is running and idle. Closing the terminal or ending the session stops it firing. Backgrounding can carry a loop into a background session, but that process still has to run. If a scheduled time passes while Claude is busy, the task fires once when Claude becomes idle; it does not replay every missed interval.

That distinction is useful when choosing where to put recurring work. An illustrative example: a developer could ask a running session to check a deployment every few minutes. A longer expiry reduces the need to recreate that schedule solely because seven days elapsed. It does not keep a closed session alive, recover every missed check or guarantee that a deployment alert arrives. Those are separate operational requirements.

The guide compares /loop with cloud and Desktop scheduled tasks. Cloud tasks do not require your machine or an open session. Desktop tasks require your machine but not an open Claude Code session. /loop requires both the machine and session to remain available. For work that must continue without the original session, the official comparison points to cloud or Desktop scheduling; GitHub Actions is another option the guide lists. Choose based on the required runtime, not the advertised 30-day expiry alone.

Resuming is more nuanced than leaving a session running. Anthropic says --resume or --continue restores unexpired CronCreate tasks, with exceptions, but a self-paced /loop is not restored. A recurring task does not make up runs missed while the session was stopped. If your work relies on a particular next run, inspect the restored schedule rather than assuming that the calendar interval continued uninterrupted. The limitations section gives the exceptions.

Check the schedule you actually have

Before relying on the longer lifetime, make a read-only inventory. This is a proposed verification checklist, not a test we performed:

  1. Record the Claude Code version in the environment that owns the schedule. The release claim is specific to 2.1.330.
  2. List the task ID and when each recurring task was created. A pre-release task keeps the seven-day limit stated in the release.
  3. Note the cadence, next run and displayed end time for each task. The guide describes the scheduled-task list as showing IDs, schedules and next run times; compare the end information your client shows with the release’s promise to say when tasks end.
  4. Record what the task is meant to produce and where that output will be checked. A scheduled prompt firing is not the same as the intended work succeeding.
  5. If a task must survive a closed session, compare the cloud and Desktop options before using /loop for it.

Do not cancel or recreate an existing schedule just to verify this announcement. If its expiry matters, check the active task first and change it only after you have decided how to preserve its intended work. The version and displayed expiry should carry more weight than an assumption based on either of Anthropic’s currently conflicting pages.

The same distinction between a scheduled action and a successful outcome matters for unattended coding work. Our guide to Claude Code unattended failures covers what to check when a job reaches a denied action. For tasks using MCP tools, our MCP result identity guide covers a separate verification step. These are related workflow checks, not evidence about the new expiry. If you are separately choosing a subscription, the AI plan finder helps compare plans; the expiry announcement itself does not change a plan. The AI model tracker follows model releases; this announcement concerns the Claude Code client instead.

Watch for copied loops in VS Code forks

The same 2.1.330 notes say Anthropic fixed a VS Code issue in which a forked conversation could run the original conversation’s /loop schedules a second time. This is a specific fix for conversation forks in VS Code. It does not establish that every scheduler or every earlier duplicated run was affected.

If you use the VS Code extension’s conversation fork, review the task lists in the original and forked conversations after updating. That is a sensible inventory check, not a reproduction result from our team. A copied schedule might create extra work even when the expiry is correct, so check both the task’s owner and its next run before relying on it.

Sources

All pages read on November 22, 2026.

Free toolIs a monthly AI plan cheaper than paying per use?Compare for my usage
DailyTech Editorial Team

We explain AI with numbers you can check. Every price, limit and benchmark links to its source and the date we read it. How we work · Report an error