nightshift

When an agent runs unattended for hours, the dangerous failures are the quiet ones: the objective drifts, the budget is silently exceeded, a step fails and nobody notices. No single-point tool is watching for those. nightshift fills that gap — it is not a scheduler, not a sandbox, but a supervisory layer that stays awake while you are not.

Before you walk away you sign a manifest: what the goal is, which shell commands prove it was met, which commands must never run, and what it is allowed to spend. During the run, every mutating command is wrapped in a checkpoint that rolls itself back when verification fails, and before each execution the gate re-evaluates forbidden patterns, irreversibility class, curfew, and class policy — a refusal means the command did not run and nothing was changed. In the morning you do not get a summary written from memory; you get an audit rendered line by line from the run journal: what changed, what was verified, what was not, what the gate refused, and the exact command to undo the night. The verdict is one word: verified, failed, incomplete, recovered, nothing-done, refused, or unclear.

Built for developers and small teams who put agents on long unattended shifts — batch refactors, nightly data cleaning, automated release prep. It assumes you already have point tools like blastradius, rewind, and ctx-budget in place; nightshift stitches them into a continuous supervisory chain. It does not decide what to do, does not schedule, does not isolate. It does one thing: hold the run to the contract you signed and honestly audit the gap.