NachtSchicht: the morning brief is the product — merge without re-reading

Apr 30 Florian Riquelme: Claude Code night-shift queue with a trust protocol — worktree isolation, blast-radius verify, halt-on-ambiguity, immutable evidence; the brief is a contract, not a transcript dump.

Florian Riquelme’s Apr 30 NachtSchicht (night shift in German) names the overnight failure mode most “just leave Claude running” setups ignore: capability is fine; trust is not. Reading five thousand streamed tool calls to decide whether a PR is mergeable defeats the automation. The North Star is one question — can the developer merge without re-reading? — and the product is the morning brief, not the agent run that produced it.

Lifecycle is hard-protocol: queue → execute → verify → report. You nacht add / nacht import during the day; nacht run works the queue in isolated git worktrees; each task is verified against its blast radius and either commits cleanly or halts with a question for morning. Ambiguity does not improvise. Six non-negotiables: the brief is the product; rigor scales with blast radius; ambiguity halts; reversibility is architecture (irreversible ops need pre-authorization); evidence compounds into earned autonomy; UX collapses to one word (nacht alone routes you to the next safe move). The brief is a contract about what is true, backed by an immutable evidence trail — not a confidence narrative. Alpha, solo-first, dogfooded on itself (every NachtSchicht change queued through NachtSchicht); TypeScript strict with a large vitest suite; MIT.

Why it matters for code review

Overnight harnessed throughput only compounds when morning review is a short contract, not a transcript archaeology dig. Continuous factories need blast-radius-scaled verify gates and halt-on-ambiguity so unattended Claude Code hours produce mergeable inventory — and park the judgment calls that still need daylight — instead of manufacturing confident unverified diffs that recreate the review bottleneck the night was meant to shrink.