Nightwatch: Captain plans by day, fleet ships overnight

Vincent Vankrunkelsven’s Aug 11 Conveo post is the tortoise overnight factory written as an operating skill, not a demo session. Afternoon: a human Captain plans with a laptop-side Helm agent that fans out for codebase context, drafts a spec, and lands a Linear project with an explicit dependency graph. Night: the Helm hands the plan to a persistent remote Field Orchestrator on an always-on EC2 box; that orchestrator self-plans build order and spawns workers — one ticket each, own git worktree, isolated app instance — while the Captain sleeps. Morning: one master PR, auto-generated test plan, and an in-app reviewer spun up on the laptop for UX QA. Their template-preview example: 7.5 hours, 3 agents, one neat PR driven green with screenshots.

The verification contract is the reusable part. Every ticket must define the cheapest closed verification loop the worker can run itself — functional (unit test or purpose-built script, no human) plus non-functional (CI + specialized AI reviewers on the PR). Bake both into acceptance criteria: “build X; verify with pnpm tsx scripts/verify-x.ts showing Y” beats “build X.” Worker PRs run until validation passes; the orchestrator merges them into a single master PR and drives that green too. Roles are explicit so overnight authority does not stall: Captain owns scope/merge; Helm co-authorizes then goes dormant; Field Orchestrator is persistent and co-equal; workers listen to both.

Keeping it alive is the other half of the receipt. Durable agent mesh (Redis Streams, not pub/sub) so mail waits for sleeping agents; monitors that wake Claude for one turn per inbound message; a 15-minute orchestrator tick that checks commits and file activity, not just chat — because a worker that acks “got it” and sleeps forever never emits the event that would timeout it. Hard-won overnight failure modes: first-run trust prompts that stall headless agents; workspaces silently branching from main when the remote never fetched the base; respawning into a fresh workspace and wiping the old agent’s files; Node heap / io_uring defaults that kill sessions silently. They treat first morning UX as usually rough on purpose, then feed night logs back into the skill.

Why it matters for code review

This is continuous harnessed throughput, not peak Claude Code mastery. Review relocates into per-ticket verify scripts + PR validation agents during the night, with the human morning pass scoped to UX and judgment — not reading 8k lines of generated React. FOMO that agents aren’t working while you sleep is answered with a fleet and a mesh clock; without both you get ack-and-nap workers and a false sense of overnight progress. Pair with dual-LLM stop-at-merge patterns (Night Shift) when the factory’s “done” still needs an independent reviewer that cannot be skipped by compaction.