Field note

Galea: overnight Claude reclaims the evaporating usage window

Jun 15: 5h Claude windows reset while you sleep — caffeinate + claude -p queue; FOMO as wasted capacity; productive complication: weekly ceiling still taxes daytime.

Source: Run Claude Code Overnight While You Sleep

Jean Galea’s Jun 15 field note reframes overnight FOMO as an accounting problem, not a model problem. Claude Code subscription metering runs on a rolling ~five-hour window (plus a weekly ceiling). Sleep seven or eight hours and that window resets once or twice while you use none of it — capacity that isn’t banked evaporates. His answer: hand Claude a queue before bed, keep the Mac awake with caffeinate -i, run claude -p headless jobs sequentially, review artifacts over coffee.

The harness is intentionally small. Each job gets --model opus, an --allowedTools allowlist (or a trusted worktree + careful bypass), --max-budget-usd, and a log file. Prefer branches/worktrees over the live working tree; never grant unattended push/deploy/send/delete. Jobs that pay off share a shape: tightly scoped, morning-judgeable in a minute — research brief to a file, tests on a branch, dependency audit without code changes, a first draft you’ll rewrite. Bigger jobs get a handoff prompt: checkable done-state, context, blocker rule (mock/assume and keep going — ranked by cost-of-being-wrong), and evidence that works counts, narration doesn’t.

The productive complication for overnight / always-on factories: reclaiming idle windows is real, but the weekly ceiling is still finite. Marathon nights can throttle the daytime when you actually need interactive capacity; on API billing there is no “free” idle window at all — every overnight token is money. Lid-closed Apple Silicon still sleeps despite caffeinate, so the honest laptop recipe is lid open + power, or move the run off-machine. Morning review is mandatory; some nights produce confident nonsense.

Why it matters for code review

FOMO that agents aren’t working overnight is often unused quota dying on the table — Galea makes that visible. Continuous factories still need Mouse-style admissible evidence and a hard merge cord; this note is the capacity half of the stack, not the verify half. Pair with Anthropic Auto mode (continuity without permission babysitting) and InterWorks’ always-on upkeep fleet: reclaim the night only for reversible, evidence-backed jobs, and treat the weekly ceiling as a real budget so overnight tortoise motion doesn’t starve the daytime review queue.