Claude Code's "Continue automatically at usage limit" is on by default, and an abandoned session can spend your next 5-hour window

Claude Code has a setting most users never touch because it is on by default: Continue automatically at usage limit (autoContinueAtUsageLimit). When a claude.ai usage limit stops a session mid-task, Claude Code waits in the open session and, at the reset, prompts Claude to pick the task up again. A bug report filed on November 13, anthropics/claude-code#109455 by GitHub user Ssmoimo, shows what that default costs when the stopped session is one you had already given up on.

What happened in #109455

The reporter's Claude Fable 5.2 session hit the usage limit mid-task. They moved the work to a fresh Claude Opus 5.5 session and treated the old one as finished, but left its window open. At the reset (November 12, 21:21 US Central), Claude Code sent the old session its continuation prompt, "Your claude.ai usage limit has reset. Continue the task you were working on", and it went back to work unattended.

By the reporter's account, the revived session ran about nine turns at roughly 360k tokens of context. Its first turn had to re-cache 328k tokens, because the one-hour prompt cache had long expired during a five-hour wait. In total it used about 6.5 million Fable tokens, roughly 90% of the 31% of the new window that /usage showed as spent. That is about 28% of a fresh 5-hour window gone while the session the user was actually working in cost $0.70. There was no prompt and no opt-in, and /usage did not say which session had used the allowance. The zombie session also repeated a file restore the active session was doing, and the two overwrote each other. The report is on Claude Code 2.1.323 on Windows; the issue is open with no response from Anthropic so far.

This is the feature working as documented

The setting arrived in Claude Code 2.1.234 on August 17: "Claude Code now continues your session automatically when a claude.ai usage limit resets; turn it off in /config". The interactive-mode docs say automatic continue "is on by default in interactive sessions signed in with a claude.ai subscription", and that at the reset Claude Code "prompts Claude to pick the task up where it stopped. It doesn't resend your last message."

The more useful part of the docs is the list of things that end the wait without continuing: sending a prompt, exiting Claude Code, /login to another account, clearing or rewinding, /resume-ing another session in that window, /teleport, /restart, or moving the session to Claude Desktop, the background or the cloud. Opening a new session in another terminal is not on that list. Neither is "I moved on." The handoff in #109455, a new session started elsewhere while the old window stayed open, leaves the old wait armed, which is exactly what happened.

There are guardrails, and they explain why this burned "only" nine turns. Automatic continue stops after three continuations (a fix for going past three landed in 2.1.314, per the changelog), and Claude Code won't start a wait on its own when the reset is more than 24 hours away, so a weekly limit that resets days out doesn't trigger it. It also doesn't arm in Remote Control or agent-team teammate sessions, background sessions or -p runs, or when you are billed per request through an API key or a cloud provider. Permission prompts still apply in the continued task, so a session that needs approval for its next command will stall rather than run. Sessions in auto mode or with broad allow rules have less to stop them.

One gap fits the report. Version 2.1.304 added a warning on Pro and Max when "resuming an inactive conversation would use a noticeable share of your 5-hour usage limit". That warning is for a resume you start yourself. Neither the changelog nor the docs say it applies to the automatic continuation, and the reporter saw no prompt before the 328k re-cache.

Not the only rough edge

Searching the claude-code tracker for the continuation text turns up a small cluster of related open reports:

• #109325 (alex-feel, Nov 12): the continuation prompt fired about six minutes after the session had already resumed by itself, starting a redundant turn.
• #95935 (mfcunningham, Sep 21): in the transcript, the automatic "please continue" message looks exactly like one the user typed, with no marker that the platform sent it.
• #92004 (MisterTuur, Sep 4): with about 11 sessions open in the macOS app, only one offered to continue at the reset and the rest stayed parked. That is the opposite complaint, and a reminder that people running many parallel sessions want this feature.
• #100650 (dowoge, Oct 9): auto-continue never armed with Remote Control connected. The docs now list Remote Control sessions as ones where Claude Code doesn't start the wait on its own.

None of this makes automatic continue a bad idea. For a long refactor you left running overnight it is the point. The trouble is that the setting can't tell a paused session from an abandoned one, and on a subscription the cost of guessing wrong comes out of the next window rather than out of a bill you would notice.

What to do

If you hand work off to a new session, /exit the old one, or press Esc at its empty prompt while the "Continuing automatically at …" line is showing (the docs say a cancel holds for that reset window). To turn the behavior off entirely, use /config or put this in your user settings:

{
  "autoContinueAtUsageLimit": false
}

With it off, Claude Code opens the usage-limit menu instead, where "Wait here, then continue automatically" is still one choice away. One quirk from the docs: /config autoContinueAtUsageLimit=false works, but the key=value form can't turn it back on, "because the setting grants unattended execution." Anthropic evidently knows what the setting is.

If an abandoned session has already eaten a window, Anthropic's one-time free full limit reset for Pro, Max, Ultra and Team accounts is still available until December 10, as we reported. Spending it to undo your own idle session would be a slightly bitter use of it.