• Status: proposed
• Date: 2026-11-06
• Issue: #28

Context

After session.sh up --apply the cluster is empty. Argo CD (ADR 0006) has to be installed, and told what to sync, without a manual step, every session. ADR 0006 said that Terraform installs Argo CD with Helm. A Helm provider in the session root would be configured from a cluster that the same apply creates, and a third root would split the state of one layer (ADR 0010).

Facts read from the Argo CD Helm chart repository:

• The newest chart is argo-cd 10.14.0, which installs Argo CD v3.6.1.
• The chart sets no CPU or memory requests for any component. Dex and the notifications controller are enabled by default. Redis is enabled by default.

Decision

• A bootstrap step in infrastructure/session.sh runs after the apply. It is a script step, not Terraform, and there is no third Terraform root.
• The step gets the administrator credential of the cluster into a kubeconfig file outside the repository, installs Argo CD with helm upgrade --install from a pinned chart version (10.14.0) and the values in gitops/bootstrap/argocd-values.yaml, waits until it is ready, sets the source range of the ingress (ADR 0016), and applies one root Application. PP_KUBE_CONTEXT makes it use an existing kubectl context instead, which is how it is run against kind.
• The root Application points at gitops/dev in this repository. The repository is public, so Argo CD needs no repository credential (ADR 0006).
• Values are minimal: Dex and the notifications controller are off, the ApplicationSet controller runs no replica, requests and limits are set from the capacity budget, and the server has no ingress.
• The UI is reached with kubectl port-forward only.
• The admin credential is generated by Argo CD when it starts. A step of session.sh (argocd-password) reads it from the cluster when asked and prints it to the terminal. It is not written to a file, a log or the repository, and the step refuses to print it when the output is not a terminal. It disappears with the cluster.
• The step can be repeated without harm. It was run against kind, including a repeat. Its ingress part runs against a stand-in for the add-on’s resource there, because kind has no add-on. That part is checked on AKS in a live session.

Consequences

• Terraform does not manage Argo CD. ADR 0006 is changed to say so.
• Helm and kubectl are needed on the maintainer’s machine.
• Upgrading Argo CD means changing the pinned version in the script.
• The credential is visible in the terminal and its scroll-back while it is shown.
• The memory and CPU of Argo CD are not measured. Session 1 measures them.

Alternatives considered

• Terraform with a Helm provider in the session root. The provider configuration depends on a cluster that does not exist when the plan runs.
• A third Terraform root for in-cluster resources. More state, and it breaks the two layers of ADR 0010.
• The AKS GitOps extension, which runs Flux. Not chosen in ADR 0006.