runnerd doctor checks the host and the config, prints each check as ok, info, WARN, FAIL or skip, and gives a fix under each warning and failure. runnerctl doctor shows the running daemon’s last run, or runs runnerd doctor when no daemon runs; the dashboard’s Settings page shows the same, and its warnings and failures are listed as needing attention. A check that cannot run because an earlier one failed, such as everything Docker answers while Docker is down, is skipped.
start Docker; unset DOCKER_HOST or point it at a unix:// socket; sudo systemctl enable --now docker, sudo usermod -aG docker "$USER", then log in again
cgroups
Docker does not use cgroup v2 with the systemd driver
boot with cgroup v2, and remove any native.cgroupdriver=cgroupfs from /etc/docker/daemon.json
isolation
isolation = "unsafe-dind": runners run privileged and a job can take over the host
Docker has no sysbox-runc runtime, or Sysbox’s sysbox-mgr or sysbox-fs is stopped
the install for your distribution; sudo systemctl enable --now sysbox
kernel modules
/lib/modules/<release> is missing, as after a kernel update without a reboot
reboot to run the installed kernel
image
a profile’s image is missing and cannot be pulled, was never built, is published only for another platform, or carries another runnerd version than the daemon
docker pull, docker build --target <stage>, or the image tag of this version
the token cannot cancel or re-run jobs; info when [github] job_actions = false or a fine-grained token cannot be checked without trying
gh auth refresh with the repo scope
repository
a repository is public: anyone who opens a pull request can run code on this host
serve private repositories only, or have GitHub require approval before it runs workflows from fork pull requests
Two checks only inform: GitHub Actions’ cache says whether jobs’ actions/cache is served here or by GitHub (Caches), and /dev/kvm whether the device is there; nothing runnerd runs today needs it.
The dashboard asks to sign in again: runnerctl web signs this browser in with a new link. A link works once.
A job fails with short read: expected … bytes but got 0 while it pulls an image: the mirror served a layer with no data. Run runnerd doctor; its mirror check names the layers and the fix. runnerd runs the mirror without the cache that causes this, and before it starts the mirror it removes layers whose data is gone or empty (ADR 0144).