[BUG]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Title: [Bug] Cloud sessions can self-schedule silent hourly check-ins that drain the cloud sessions credit with no notice or cap

Summary
A Claude Code cloud session used the send-later / reminder tool to schedule a one-shot check-in on itself. The check-in prompt said "act on anything actionable, otherwise re-arm silently". Each run then booked the next one, which created an unbounded hourly loop. It ran 29 times in about 28 hours, overnight included. Almost every run found nothing to do, and every run spent the cloud sessions credit.

Impact
I claimed the $250 Max cloud sessions credit on 25 Sep 2026. Within about 24 hours, 132 of it was gone. A second session started the same pattern. There was no notification, no spend indicator and no cap. I only found out by listing my scheduled tasks.

Why each run was expensive
Each wake-up resumes a long-context persistent session. The runs were about 61 minutes apart, just past the 1-hour prompt-cache TTL, so it looks like every run re-read the full context uncached.

Steps to reproduce

1. In a cloud session, ask Claude to keep an eye on a PR (CI or review comments).
2. The session schedules a one-shot reminder on itself with a prompt to re-arm if there's nothing to do.
3. Each firing creates the next reminder about 1 hour later, with no end condition.

Expected

• Self-scheduled runs are visible and show their credit or usage cost.
• A limit on self-re-arming chains (a max count or an expiry), or user confirmation before a recurring pattern starts.
• Scheduling intervals that stay inside the cache TTL, or a warning when a session is idle-woken on a large context.

Environment
Claude Max, Claude Code cloud sessions (Sept 2026 GA). Support case: Fin conversation 215476119538305, escalated to human review. Session IDs available to Anthropic staff on request.

What Should Happen?

• Self-scheduled check-ins are visible to the user and show their credit/usage cost.
• Self-re-arming reminder chains have a limit (max count or expiry), or need user confirmation before a recurring pattern starts.
• Idle wake-ups of a large-context session are spaced within the prompt-cache TTL, or the user is warned.
• The cloud sessions credit is never drained by background runs the user doesn't know about.

Error Messages/Logs

Steps to Reproduce

1. On a Max plan with the cloud sessions credit claimed, start a Claude Code cloud session on a repo and open a PR.
2. Ask Claude to keep an eye on the PR (CI / review comments).
3. The session schedules a one-shot reminder on itself with a prompt like: "re-check CI and review comments; act on anything actionable, otherwise re-arm silently."
4. Each firing creates the next reminder ~61 minutes later, with no end condition.
5. Result: 29 unattended runs in ~28 hours (overnight included), each resuming the full long-context session. About 132 of the 250 credit was consumed in ~24 hours, with no notification.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.283 (Claude Code) — cloud sessions on claude.ai/code

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_