What changed, release by release.

A faster first node behind a dead mirror.

• On AlmaLinux and Rocky the first fetch of package information gets ninety seconds before the installer leaves a silent mirror for the distribution's own servers. It used to get four minutes.

Its pinned build passed 30 of 30 checks on Rocky Linux 10, AlmaLinux 9 and Debian 13. Behind the silent mirror, adding a node on Rocky Linux 10 took 280 seconds where 1.1.11 took 412.

A first machine that does not wait on a dead mirror.

• When the installer gives up on a silent package mirror, the server remembers it for six hours. The agent it has just installed used to ask the same mirror again and wait four more minutes, and the first machine on Rocky Linux 10 waited with it.

Its pinned build passed 30 of 30 checks on Rocky Linux 10, Rocky Linux 9 and Ubuntu 24.04. On Rocky Linux 10, behind the same silent mirror, the first machine was built in 33 seconds.

A quieter and less patient installer.

• Adding a compute node prints its own stages. A warning appears as a sentence under the stage it belongs to, and the rest goes to /var/log/klyrn-vm/node-install.log.
• On AlmaLinux and Rocky, asking the package mirrors what they have now gives up on a silent mirror after ninety seconds and asks the distribution's own servers. It used to wait five minutes.

Installed from the public command on twelve Linux systems: eleven passed 30 of 30 checks. On Rocky Linux 10, behind a silent package mirror, the first machine took longer to start than the check waits; it then ran and passed the rest.

The refusal, on the page.

• When the node an administrator chose cannot take a machine, the create page shows the reason itself, such as how much memory was short, and what to do. Before, it said the same sentence twice ahead of the reason.

Installed from the public command on twelve Linux systems, 30 of 30 checks on each.

A refused machine says why.

• When a machine cannot be placed, the message names the reason: how much memory, disk or processor was short, or that the node is held, in maintenance or not connected. Before, it only said the node could not place it.
• A customer is told which of their own numbers did not fit, without the names or the sold capacity of the servers.

Installed from the public command on twelve Linux systems, 30 of 30 checks on each.

The machine list on a laptop.

• The list controls are one wrapping row: two lines on a wide screen and three at 1280 pixels, where they were four. Nothing is hidden at any width.

The panel says more of what it knows, in less space.

• The machine list's controls fit two rows: the node and project filters and the search share the status row, and grouping moved into the View menu.
• The bar names the page once its heading has scrolled away.
• A compute node whose row carries an amber edge says why.
• The Overview names its work list by what is in it: running work, or the last finished.
• Every connected node's storage pools are measured once an hour. They were measured only when somebody opened the page, so figures could be days old, and the storage alerts are computed from them.

Two faults that only showed on a server in its first minutes.

• A machine waits for the kernel's filter modules before it starts. On AlmaLinux 10 and other Enterprise Linux 10 systems the node installs them in the background when it first connects, and a machine created in those seconds failed to start with an ebtables error.
• When the package mirror list costs dnf a timeout, the installer asks the distribution's own servers (Rocky Linux and AlmaLinux), with the signing key the system already trusts. A mirror that accepts connections and never answers no longer stops an install however many times it is tried.

For servers and machines that have only just started.

• First boot installs the guest agent only when the image does not already have it, and for five minutes at most. It used to run the package manager on every first boot; on a day one Rocky Linux mirror misbehaved, that install hung for good on a package that was already there and held the package lock.
• The installer waits for another package install on the server and names it, where it used to wait in silence and run out of time.
• dnf gives up on a slow or silent mirror and moves to the next, and an install that ran out of time is tried again.
• The bar names the page once its heading has scrolled away, and the customer's summary card no longer cuts its own words.

For the first hour on a new server, and for paying customers.

• A machine can be created the moment a new compute node connects: the node reports what it has twice more in its first half minute. On Rocky Linux 9 the storage pool was reported a minute late and a machine created in that minute was refused. (This was in 1.1.1, which shipped without notes.)
• A command that ran out of time says so, and that running it again carries on. A node install on a slow package mirror used to end with "dnf exited -1:" and nothing after it.
• A licence warns of its end seven days ahead, not thirty. One bought by the month warned its owner every day, starting the day they paid.
• Ubuntu 26.04, Rocky Linux 8 and 10 and CentOS Stream 10 are named as proven compute node systems, each with a full run behind it.
• KLYRN VM licences can now be bought at my.klyrn.com: order the nodes beyond the free one, pay, paste this controller's install id, and paste the licence file it gives you under Settings, Licence.

Install on the system you already run, and reach a first machine with nothing set up by hand.

• The installer runs on any systemd Linux. Proven end to end on Ubuntu 22.04, 24.04 and 26.04, Debian 12 and 13, AlmaLinux 8, 9 and 10, Rocky Linux 8, 9 and 10 and CentOS Stream 10.
• One server is enough: klyrn-vm admin node add-local makes the controller's own server a compute node, and the installer takes --single-server to do both at once.
• The first compute node brings a private network, and image checksums are fetched at once, so the first Create works.
• The installer opens firewalld as well as ufw.
• klyrn-vm doctor calls unfinished setup a next step, not a failure.
• Cancelling a task is safe: it reports as cancelled and nothing else on the host is disturbed.
• A machine must stay stopped or missing for ninety seconds before anybody is told.
• The side menu fits one screen and shows where you are.