The cluster picks up the cards
Reconstruction⚠ This is not a live capture: it is a condensed reconstruction, edited afterwards from the real transcript of the September 14–15, 2026 session. Ludo's prompts are quoted word for word, typos included, except for machine names replaced with the article's; tool output is rewritten and shortened, the timing is compressed and names are sanitized.
The request is simple: plug the new RTX 3050 into the first gaming VM and make
the GTX 960 available to Kubernetes. The NVIDIA driver goes on the host, and
after the reboot the tower turns into a ghost: the node shows Ready, its
lease keeps renewing, and no container starts anymore. The agent first blames
k3s’s PATH, then locks itself out by hammering sshd during boot. A photo of
the screen settles it: a chain of hung tasks in the kernel, modprobe stuck
behind bridge.
The driver is pulled off the host and handed to a compute VM, vm-03. While
writing it, the agent catches a synchronous virsh call in the libvirt hook,
one that could have blocked the very gaming VM it was meant to help. Then comes
borrowing the two 3050s: three cards will not boot, and the bisect blames the
cards when it was really comparing a VM with an install CD to one without. The
agent settles for a single card; Ludo pushes back, and <rom bar='off'/>
fixes everything.
The last highlight is the swap itself, tested live: a sudo masked by PATH,
a cordon that survives the reboot, a fixed-name systemd-run unit that refuses
to start again, a shutdown sent before the guest agent is up.
The full story, told by Bob, is in When nobody is playing, the cluster picks up the cards.