Nightshift: overnight work only counts behind seven verification stages
Apr 6 Orbit/Sharma: open-source overnight orchestrator — hardening + feature loops, worktree isolation, shift logs; runner enforces policy (blocked paths, verify command, anti-tunnel-vision), agent supplies intelligence.
Source: Introducing Nightshift: An Autonomous Overnight Agent That Actually Works
Pranit Sharma’s Apr 6 Orbit launch note frames overnight agents as an orchestration failure, not a model failure: interactive Claude Code / Codex / Cursor work; leave them unattended and you get tunnel vision, lockfile edits, unauditable diffs, and “fixes” that no longer build. Nightshift is an open-source Python control plane (Codex or Claude adapters) that runs for hours with enforcement in the runner, not prompt discipline.
Two loops share one daemon. Hardening reads repo instructions and the shift log, picks a discovery strategy (security, resilience, coverage, a11y, hygiene, performance, polish), commits small fixes, and logs larger issues for humans. Feature builder takes a natural-language goal, profiles the repo, decomposes into waves, and resumes across sessions. Both pass the same gates. After every cycle the runner runs seven stages: commit + shift-log credit, blocked-path/lockfile rejection, configured verify command (pnpm build / cargo check / inferred), zero file deletes, anti-tunnel-vision balance across categories/paths, breadth of exploration, and flagging of prompt/control-file self-edits. Failures revert or halt. Morning artifacts are the review surface: human-readable shift log, machine-readable state.json, raw runner log, and an isolated nightshift/YYYY-MM-DD branch with atomic prefixed commits — working directory untouched.
Daemon roles (builder, reviewer, overseer, strategist, achiever), handoffs, and learnings files accumulate between sessions; security pretends the 8-hour unattended case is hostile (blocked globs, self-mod guard, watchdog, red-team preflight). Receipts claimed: 935 tests, 80 merged PRs, 28 modules, MIT.
Why it matters for code review
Always-on factories compound only when morning inventory arrives already gated — verify command green, blocked paths respected, tunnel vision steered, and an auditable shift log that separates claimed work from credited work. Continuous harnessed momentum looks like Nightshift’s split: the agent provides intelligence; the runner owns “done.” FOMO that nothing ran overnight is the wrong fear if what ran produced an unauditable worktree; the scarce morning job is reading the shift log and promoting reversible commits, not reconstructing what the model believed it did.