Night Shift: overnight Claude with Codex that must agree you're done
ppuliu skill: up to 8h on a feature branch; file-gated Codex review per commit; dual consensus to end; never push/PR — self-review only when Codex absent, explicitly flagged.
Source: Night Shift — autonomous overnight development agent for Claude Code
Night Shift is the tortoise overnight pattern with the review cord welded shut. Approve one objective, walk away; Claude Code plans/implements/commits on a feature branch for up to 8 hours. Structural bets (stated plainly): (1) two independent LLMs — Claude writes, Codex adversarially reviews; a compacted agent cannot skip the gate because commit requires a real code-review.txt on disk matching recorded verdict; (2) small reversible units — one commit per task, failing review/tests roll back only that task; (3) the agent cannot declare done alone — early exit needs Claude proposing "we're done" and Codex agreeing on a re-review of the objective/KR history; without Codex the shift runs to the hard cap with every review marked CODEX UNAVAILABLE.
Loop per task: plan → implement → Codex review (fix/revert hard-stop at 10 rounds) → tests must pass → file-gated commit. Morning artifact is .night-shift/runs/<id>/handoff.md (what shipped, commit hashes, review rounds, unattended decisions). Safety table covers the usual overnight failure modes: no git push, no PR open, drift check stops on external HEAD change, scope gated per key result against the objective. Needs --dangerously-skip-permissions in a trusted directory — the FOMO half without the merge half.
Why it matters for code review
This is Mouse's stop-at-merge spirit with an explicit dual-model review ledger. Continuous overnight motion only compounds when "done" is consensus + files on disk, not a single model's end-of-session claim. Use it as the reference shape for tortoise factories: independent reviewer process, admissible review artifacts, hard wall-clock cap, human owns push/merge in the morning.