[PROJECT] Extensible Install w/ Kustomize & Helm #3158

Parent Issue tracking Extensible Install Roadmap

Goals

• Provider-Neutral Core: agent-substrate/substrate contains no provider-specific code, manifests, or tooling. All provider-specific content lives out-of-repo.
  Existing GKE-specific install content to move to ai-on-gke/substrate-gke. (other providers may use the same pattern in separate repositories).
• One Source of Truth: The Kustomize tree becomes single-source-of-truth for what gets installed for the core substrate. Other artifacts are generated from it.
• Installable Releases: Each release publishes anonymously pullable images, kustomize bases that kubectl apply -k and GitOps tools can use directly, and Helm charts generated from the same render: one for CRDs and one for each stage of agent-substrate system.
• Fork-free provider extensibility: Providers and operators change installation via kustomize components and documented settings or via Helm pre- and post-hooks in umbrella charts, not via patching core configs or imperative core-side installation.
• Separate operator-controlled values from base: Operator controlled values are not contained in the base kustomize files – defaults are supplied by overlays making re-applies safe for operator controlled fields (e.g. replicas, cpu, etc). Operator controlled values survive redeploys, upgrades, etc.
• Safe Upgrade ordering: Install paths follow the prescribed component ordering in the upgrade skew policy

Workstreams

Phases

1. Foundation. Interfaces agreed; component flag released; GKE Component created; GCP provisioning tools and staging out of core; warning releases for upcoming breaking changes.
2. GKE parity. substrate-gke installs through its Component; ate-setup can render the install; sizing moves out of the base manifests.
3. Core cleanup and release artifacts. GKE defaults and GCP logic removed from core; bootstrap, ordering, and staged upgrades proven in CI; release bundle published; Helm charts generated and tested.
4. Publish and finish. Charts published; GKE kustomize and Helm paths; last provider plugins moved out of core.

Integration points:

• the GKE parity release (GKE parity release: install through the GKE Component ai-on-gke/substrate-gke#151) gates core's removal of GKE defaults;
• the first published release bundle gates the GKE kustomize path and chart publishing.