OvernightAgent: four gates between every step and 'done'

Souler Ou’s OvernightAgent (oa, npm @soulerou/oa-cli; repo Apr 20) is the overnight supervisor for people who already have a coding agent and keep losing the night to silent drift. You intake tasks, seal a plan with a wall-clock budget, and oa run --detach. Each task gets its own git worktree; your checkout stays clean. After every step the supervisor runs a four-gate verify pipeline — structured oa-status tail → commit-since-step-start → your verify command → an AI reviewer — and only then advances. Failures open a fix-loop with reviewer findings injected into the next prompt; exhaustion can defer leftovers into FINDINGS.md instead of pretending the night succeeded.

The morning artifact is the review surface: append-only events.jsonl (dozens of typed kinds) plus auto-rendered SUMMARY.md — outcomes, attempt tables, open P0/P1 issues, links into per-attempt prompts and logs. Crash recovery is explicit (oa rerun rewinds in-flight worktrees and re-enters). Hardening that matters for continuous factories: Claude compact-recovery hook that re-injects task context after auto-compaction, stall soft/hard thresholds, plan-level error budget, rate-limit backoff that retries the same prompt without burning verify attempts, optional macOS sandbox-exec, and live heartbeats so “slow tokens” is distinguishable from “wedged.” Adapters are swappable (claude / codex / opencode) behind one supervisor. Honest v0 limits: sequential tasks only; parallel schema accepted but not yet concurrent.

Why it matters for code review

Always-on harnessed momentum beats a heroic Claude Code burst only when each overnight step pays a verification tax before the next unit starts. Review oa the way you review a merge queue: the four gates and the morning SUMMARY are the merge criteria; the agent is just the worker behind them. Continuous factories need that ledger — promote what survived gates, reopen deferred findings, and refuse to treat chat confidence as done.