Validation and sources
What the automated checks cover, how to check a build by hand, and the upstream sources the role relies on.
Checked automatically
CI runs these on every push, without network access to hosts:
| Check | Covers |
|---|---|
cargo test --locked | The terminal UI's fields, profile edits, checks before Run, SSH discovery and rendered screens. The APIs of rookeryd and rookery serve against the committed documents. Links to a local rookeryd |
python3 -m unittest discover -s tests | Disk selection and its refusals, profile validation, release packaging, seed generation and GRUB patching. With Ansible installed, also the rendered templates for every profile |
bash -n | The install script, bootstrap, container, role and plugin scripts |
ansible-playbook --syntax-check | The role |
cargo fmt, cargo clippy | Formatting and lints |
vale | The prose of the README, these docs and the plugin guides |
tests/e2e | Rookery against a test host in a container. A profile is applied over SSH, and a second time without changes. Then the tests use the host with SSH keys, actions, terminals, files, secrets and a plugin from a folder. rookery serve is tested with its roles, its set-up and its page, and with a restore of the installation from its database and key file. They use a development build of Rookery |
See Building and releasing for running them locally.
Boot tests
The container suite can't boot anything. A cloud seed and the USB installer are booted in a virtual machine with KVM instead. ci/run.sh boot cloud and ci/run.sh boot iso run them. They take longer and are not part of every push.
| Test | What happens |
|---|---|
boot cloud | The Ubuntu 26.04 cloud image boots with the seed of rookery host seed as its NoCloud user data. It downloads the release archive over HTTPS and applies the role on its first boot |
boot iso | The installer of rookery host iso boots with only a disk of another serial, and must refuse it without writing to it. It then boots with that disk and the right one, erases and installs the right one and powers off. The installed system boots without the installer and sets itself up |
After the first boot, Rookery reaches the host over SSH and checks:
- rookeryd and a plugin's service run as systemd units, and rookeryd starts, stops and restarts units and reads the journal
- the firewall lets in SSH and nothing else, and SSH takes keys only and refuses root
- the status action ends without failed units, and a second apply in a row doesn't change anything
rookery rollout --all --rebootrefuses while a session runs, then reboots the host, waits for it and checks it- the end-to-end tests of commands, rookeryd's API and secrets
The guest runs on QEMU's user network with a package cache and a release server on the container's side. The cache is in build/boot/cache and makes a second run a lot quicker. The profile keeps the desktop and leaves out toolchains, Docker and Cockpit. It also leaves out the office and graphics programs that a host can do without, when ROOKERY_CI_BOOT_EXTRAS=1 is not set. On GitHub the Boot workflow runs either one when started by hand. Started by hand or by its weekly schedule, it installs those programs as well. Logs, timings and the disk and memory used are in build/boot/KIND/.
Checking a build by hand
docker compose -f container/compose.yml build runs the whole role with every feature on, up -d serves the desktop through noVNC and T3 on loopback ports, and exec -T desktop /usr/local/bin/rookery-smoke runs the smoke test. Under Docker's default seccomp profile Chromium cannot create its sandbox, and the smoke test reports the sandboxed browser launch as skipped.
ci/run.sh boot cloud does this unattended with a light profile. By hand, provision/bin/qemu.py cloud boots the Ubuntu cloud image. Apply the role over SSH with the qemu profile, then run rookery run HOST smoke. A second apply should not change anything, and after a reboot the desktop, noVNC, T3 and Docker should be up without anyone logging in. See Testing in QEMU.
ci/run.sh boot iso does this unattended with disk mode serial. By hand, provision/bin/qemu.py install boots the remastered ISO with disk mode auto. It selects the virtual disk, installs without input and powers off. Test the selected disk policy this way before booting a physical machine.
After an apply, sshd -T reports password, keyboard-interactive and root logins off, and only key logins for the operator. Key login still works.
Secure Boot with the remastered image, Wi-Fi and real disks are not covered by any of these checks.
Package availability and external installers change over time. If a dependency fails, provisioning records failed, retries from its timer, and never writes the success marker.
Primary references
- Ubuntu autoinstall and its validation
- cloud-init NoCloud datasource
- T3 v0.1.8 release, Linux x64 archive SHA-256
b9c49dc00c23fbbc5961bdaff2d88e556d96c8dbd638c4501e2937c70650a9e5 - T3 installer at the release commit and its background services
- Playwright for Python v1.65.0
- ansible-core 2.21.6 and Ansible's Python support matrix
- Mozilla's native Firefox packages
- TigerVNC Xvnc
- Tailscale Serve
- noVNC server requirements
- The references on Tools and Graphics
Docker comes from Ubuntu's packages rather than Docker's own repository. The cloud-init document is generated directly rather than from templates, so the seed builder has no dependencies.
Last updated on