Night Shift

Night Shift
Step away for the night.
Claude Code keeps grinding.
Wake up to well planned, committed, tested, reviewed work.

Night Shift is a Claude Code skill that turns a normal Claude Code session into an autonomous overnight development agent. You approve one thing — an objective — then walk away. The agent plans, implements, and commits work against a feature branch for up to 8 hours.

When OpenAI Codex is available, every commit is gated by an independent review from a second, separately-prompted LLM. When it isn't, the agent falls back to a clearly-marked self-review and keeps working — you lose the dual-LLM guarantee but not the loop.

Why

Long-horizon autonomous coding agents fail in predictable ways: they hallucinate progress, paper over broken tests, escalate scope into unrelated refactors, sometimes push to main. Most "run it overnight" setups are one LLM grading its own homework.

Night Shift makes three structural bets:

1. Two independent LLMs review each other, never the same one grading itself. Claude writes; Codex (when available) adversarially reviews. A task cannot be committed until a real review artifact exists on disk — a compacted, forgetful agent cannot quietly skip the gate.
2. Small, reversible units. Every task is one commit. A failing review or test reverts only that task — not the whole session. Every commit on the branch passes tests.
3. The agent can't decide it's done. Ending on "we're done" requires both the agent proposing it *and* Codex agreeing. Without Codex, the shift runs to the 8-hour hard cap.

How it works

Each task runs: Claude plans → Claude implements → Codex reviews the code (loops until clean, hard-stops and reverts at 10 rounds) → tests must pass → file-gated commit. The pre-commit gate refuses to stage a task whose…