☾Nightshift

Running coding agents by hand doesn't scale: one terminal, one task, and you have to stay at the computer. I wanted agents to work through a queue on several servers, safely, while I'm away, and only hand me results that pass checks.

• Postgres-backed queue with atomic pickup (FOR UPDATE SKIP LOCKED), so two workers never take the same task.
• Each run gets its own git worktree, runs the agent, then a verify command; a failed verify gets one repair pass, otherwise the task ends as partial, not done.
• High-risk tasks go through a plan gate: a read-only planning run, approval in the app, and only then the real run.
• Android client for creating tasks, approving plans and cancelling runs, with push notifications and OTA updates.

1. 01 ClientNative Android app (Kotlin) + `ns` CLI
2. 02 Control planeFastify REST API, auth, notification outbox, FCM
3. 03 QueuePostgreSQL task state machine with leases and heartbeats
4. 04 WorkersDependency-free Node worker per server: worktree → claude -p → verify → PR

This project is about making agents trustworthy enough to run unattended: they never merge on their own, failures are labelled honestly, follow-up ideas are only proposed and never started automatically, and every run is isolated in its own branch.

• ✓Server, worker, migrations and Android client are present in the repository.
• ✓16 test files; Dockerfile and Compose for the control plane.

• !Personal infrastructure, experimental. No CI pipeline yet.

Screenshots are being prepared with anonymised data. Real client data will not be shown.