Agentic development supervisor

Ship the work that happens after you go offline.

Nightshift turns queued specs into reviewed GitHub work. It watches a shift queue, dispatches builds through Graveyard, sends results through Handoff review, and stops when budgets, permissions, or quality gates say stop.

[supervisor] queue=pending idle=30s
[dispatch] claimed feat.checkout-recovery.shift.json
[codex] Graveyard build loop started
[guard] budget ok, rate ok, breaker closed
[handoff] verdict: REQUEST_CHANGES
[dispatch] requeued with recovery hint
[supervisor] next shift in 30s
01

Queue

GitHub comments and local scripts create shifts that move from pending to active.

02

Build

The Codex adapter runs the Graveyard build loop with subscription-first auth.

03

Review

Handoff reviews the result and Nightshift updates the PR check run.

04

Guard

Budgets, rate limits, circuit breakers, and notifications keep the agent bounded.

Current build

Ralph Loop with teeth.

Nightshift grew out of AgentForge and the Graveyard runner. The new shape is simpler: a launchd-friendly supervisor, a queue contract, a Codex build adapter, Handoff review, and an audit trail for every shift.

Original proof run

AgentForge was the Ralph Loop demo.

The first version was a self-referential dashboard: the app visualized the autonomous build loop that created it. Nightshift keeps that idea, then turns it into a reusable supervisor for queued GitHub work.

Features
29/30
Avg score
9.69
Iterations
41
Tokens
109k
Cost
$0.95
Elapsed
76m