Nocturne

Nocturne is a Claude Code skill that turns a repo and a checkbox backlog into an autonomous dev loop. A design-layer Claude reads the repo, authors a harness tailored to it, and launches a headless Claude loop that implements, verifies, and commits each task on a dedicated branch.

The premise is trust but verify. A task counts as done only when the quality gate passes an independent re-run, a real new commit exists, and the worktree is clean. Until the harness observes that, the task sits in a done/not-done superposition.

The quality gate is derived from the repo itself, and an explicit done-statement outranks CI config, package scripts, and language defaults, so the loop verifies against what the owner meant.

Green is not enough on its own. An anti-gaming diff guard rejects attempts that weakened checks or deleted assertions, acceptance criteria are parsed from the backlog and enforced in the diff, and protected paths are checked after every commit. Budget caps bound dollars, wall clock, and per-task time. Commits are atomic, one per task, and failed attempts are discarded.

Models are tiered per task with automatic escalation on retry, and rate limits pause and resume the run instead of burning attempts. The harness is Python 3.10+ standard library only, with 323 passing tests, and runs on Windows, Linux, and macOS.

The Nocturne card on this site was built end to end by a Nocturne run, and the 3/3 above is that run's scoreboard. The backlog had three tickets: the project data, the homepage wiring, and the SVG cover art.

All three shipped, one atomic commit each, and every commit merged to main unchanged after review. The gate was npm run lint plus npm run build, re-run independently by the harness after every task. The run cost $1.15 on claude-fable-5 and took about ten minutes of wall clock.

Nocturne is a loop for well-specified backlog work, not a substitute for judgment. It does what the tickets say and proves it; deciding what the tickets should say, and reviewing what comes out, stays with a human.

The verification stance is the point. Autonomous coding usually fails not because the model cannot write code but because it reports success too easily. Nocturne's answer is structural: independent gate re-runs, diff guards, and real commits as the only evidence that counts.