Files
zeekayandHanzo Dev 10e5c10fc5 ai reaches the cluster through a seam, and stops running informers
cluster/ is live application code — /v1/finetune/* and the deploy/undeploy
actions reach it — and it genuinely needs cluster state, so this is the seam
case rather than the delete case.

cluster/k8s.go mirrors cloud-oss/k8s.go: a K8sClient interface, a BuildK8sClient
that never returns nil, and an unavailable default that fails closed with a
reason. Its error is deliberately NOT one of the sentinels, so "there is no
cluster" can never be read as "the object is absent". The factory takes a
kubeconfig because this service is multi-cluster — the config comes from the
tenant's provider row, not from the pod.

cache.go is deleted: 511 lines running six informers over services, deployments,
ingresses, nodes, events and namespaces. An application should not hold a
replicated mirror of cluster state. View now lists per request, and Describe
derives its URL from the view it already built instead of making a second round
of reads.

The typed k8s.io/api structs stay, decoded from the seam's maps through
apimachinery's converter, so every existing status and failure branch is
unchanged rather than rewritten. Types are not cluster access.

go list -deps finds no client-go. k8s.io/api, apimachinery, metrics (types) and
kustomize (renders, never dials) remain.

Co-authored-by: Hanzo Dev <[email protected]>
2026-08-05 19:11:47 -07:00
..