frostyard lab holds the GitOps QA pipeline for the
snosi bootc images: Argo Workflows
templates, cluster manifests and the QA dashboard for the lab's single-node
k3s host. The snosi maintainers use it to learn whether a published image is
good; a change reaches the cluster only by a merge to main, which Argo CD
reconciles.
Status: in production.
Documentation: docs/README.md
| Path | What it is |
|---|---|
argo/workflow-templates/ |
The WorkflowTemplates that Argo CD syncs (Application frostyard-lab): the digest poller, the QA pipeline, and the container, VM, install and lifecycle lanes |
argo/*.yaml |
Submit files for one-off manual runs |
manifests/ |
CronWorkflows, RBAC, semaphores, digest state and result publication (Application frostyard-lab-infra) |
argocd/ |
The Argo CD Applications, applied once by hand at bootstrap |
site/, e2e/ |
The Astro QA dashboard on GitHub Pages, and its Playwright suite |
scripts/ |
The run collector, the dev submit helpers, the lifecycle preflight and the template gate |
tests/, test/ |
Offline pytest contracts for the templates, manifests and scripts; Node tests of the template gate |
policies/ |
The deny-by-default agent governance policy and its checker |
Justfile |
Operator wrappers over kubectl and argo |
docs/ |
Decisions, designs, the lifecycle lane spec, runbooks and open plans |
Results are public on the QA dashboard,
which renders runs.json from the data/runs branch
(ADR-0021). To look at the cluster,
start read-only with the current kubeconfig context:
just status # Argo CD sync and health, and the enabled lanes
just runs # recent workflow runsTo change a pipeline, open a pull request: after the merge, Argo CD applies it,
and selfHeal reverts anything applied by hand. Never kubectl apply a
WorkflowTemplate or a manifest. The clients, every recipe and the manual runs
are in operating the lab; a cluster from zero is in
bootstrap; the field contract of the manual bootc
lifecycle lane is in its spec.
A CronWorkflow polls each product's registry digest. When the digest moves, the QA pipeline runs the testsuite BDD suites against that exact digest, in the image booted as a nested systemd container, and stores the digest only after QA passes, so a failure retries on the next poll. Incus VM lanes cover what a container cannot see: Secure Boot, the installer and the update path. A collector publishes the results to the dashboard in 1 direction; the dashboard never queries the cluster. Details: architecture, VM lanes and reporting.
- Evidence is advisory: no lane is a release gate, and a VM lane is not hardware qualification (quality).
- The dashboard shows bounded workflow summaries, not full diagnostics, and can be stale (reporting).
- The lab has no artifact store: the workflow log is the evidence surface (ADR-0009).
- More known gaps are in architecture. The superseded roadmap is history; it still holds G4 pre-promotion gating and the open questions.
mise install # the pinned Node
python -m pip install -r requirements-test.txt
make ci # what CI runsThe full setup, the test suites and the Playwright run are in CONTRIBUTING.md.
- AGENTS.md: the rules for agents and contributors.
- docs/README.md: the decisions, designs, specs and plans.
- SECURITY.md: report a vulnerability privately.
- LICENSE: MIT.