nightcrew: BACKLOG in, typed failures out — verify before any land
Jul 2 nightcrew: overnight control plane around Codex — plan from markdown backlog, worktree isolation, deterministic verify, independent merge review, streak guards, morning report.
nightcrew (npm; Jul 2) frames the overnight problem the same way Orbit and AgentRelay do — interactive CLIs fail unattended in known ways — then turns each failure into a typed, recoverable event instead of a morning surprise. You write a markdown BACKLOG and go to bed. The control plane (Codex via the official SDK today) authors a bounded plan, executes in an isolated nightcrew/<plan-id> worktree, runs your deterministic verify profile, and only then asks an independent merge reviewer on a fresh session to judge intent, scope, and honesty. Your checkout stays clean; protected-path edits revert before commit.
The failure table is the product: off-task wander → plan review on bounded seams; “done” without proof → verify profiles gate landing and downgrade to repair; self-approving slop → separate merge review; infinite retry → failure/no-commit/control-only streak guards; quota wall → quota_exhausted schedules a window-aware resume. Loop operations are explicit: plan → execute → verify → repair → periodic garden hygiene. nightcrew run for one supervised iteration, loop for a bounded unattended session, crew start for the daemon across projects and schedule windows; morning is nightcrew report — what landed, what failed, what it cost, which decisions wait. .nightcrew/ is committed; runtime state is gitignored — the repo is the database.
Why it matters for code review
Overnight harnessed throughput compounds when landing is verify-then-independent-review, not agent self-certification, and when spirals stop with typed reasons instead of silent token burn. Continuous factories need that morning ledger: promote what survived gates, reopen what repair couldn’t finish, and treat the BACKLOG-plus-report pair as the review surface — not a mystery worktree of claimed greens.