- Shipped
- November 14, 2026 at 8:40 PM UTC
- Author
- Kamo
- Commit
- c740c1b
Forgejo tasks stayed "running" for minutes after "Job succeeded" because the hourly `docker builder prune -a` (3-7 min, 6-23 GB) holds containerd's GC, and the runner's `docker rm` of the job container waits in leases.Delete behind it (goroutine dump: ContainerRm -> ReleaseLayer -> leases.Delete blocked 3+ min). Tasks were released in the same second the prune finished. Walk the cache down in 3 GB --keep-storage steps with a pause between them and run every 15 min, so a queued container removal waits at most one step. Adds the host script to node-config as its source of truth (installed by hand at ****************