Claude Code: watchdog auto-resumes stalled jobs after 19-hour limit

The setup exists because unattended sessions fail at the worst time. The developer's Claude Code runs in tmux on a home PC with two jobs: orders arriving by voice from a commuter, and a scheduled run every night at 2am. When the 5-hour or weekly quota runs out, Claude Code stops with an error, and with nobody at the screen the session could sit for hours after the limit lifted.

The interesting part is how carefully the watchdog avoids making things worse. It filters out errors that look similar but will not fix themselves, such as a context limit or a budget limit — waiting on those and resending would just repeat the same error. It reads the reset time from the error text, then double-checks the same usage figures shown by Claude Code's /usage and waits until every window is below 100%. Before sending, it re-reads the log and the screen in case the user typed something from their phone in the meantime, and it records the send before sending so a crash cannot cause a duplicate.

On the day it ran, the weekly limit hit right after the 2am scheduled job; the watchdog asked to continue 43 seconds after the limit reset at 21:00, and the stalled job was finished without anyone touching it. The developer notes the exact error wording may change between Claude Code versions, so the pattern matching may need updating — and when it does not match, the design falls back to doing nothing rather than something risky.