Summary

When bundling with --deployer argocd-helm and pushing to an OCI registry, the generated gpu-operator-pre ArgoCD Application uses spec.source.path instead of spec.source.chart. ArgoCD then tries to resolve the bundle's targetRevision as a tag on the OCI registry base rather than on the named chart, which always fails with not found.

The root aicr-stack (app-of-apps) Application has the same shape and also fails to load its target state.

The downstream effect on a real cluster is severe: gpu-operator-pre never syncs, the nvidia-kernel-module-params ConfigMap is never created, the GPU Operator ClusterPolicy stays notReady ("Failed to reconcile state-driver: ... ConfigMap "nvidia-kernel-module-params" not found"), the NVIDIA driver is never installed on GPU nodes, and downstream consumers (e.g. nvidia-dra-driver-gpu kubelet plugin) hang in Init forever waiting for /run/nvidia/driver.

Environment

• aicr built from main (post-v0.13.0)
• Deployer: argocd-helm
• Bundle artifact pushed to a private OCI registry as a single Helm chart, e.g. oci://<registry>/<org>/<chart>:<tag>
• ArgoCD: argo-cd Helm chart, default install
• Cluster: EKS with GB200 (Ubuntu 24.04)

Reproduction

aicr bundle \
  --recipe recipe.yaml \
  --deployer argocd-helm \
  --dynamic gpuoperator:driver.version \
  --output oci://<registry>/<org>/<chart>:<tag>

# On the target cluster (ArgoCD already installed, OCI repo secret registered):
helm install <release> oci://<registry>/<org>/<chart>:<tag> \
  --namespace argocd \
  --set repoURL=oci://<registry>/<org> \
  --set gpuoperator.driver.version=<version>

After install, kubectl get applications -n argocd shows:

NAME                       SYNC STATUS   HEALTH STATUS
aicr-stack                 Unknown       Healthy
gpu-operator-pre           Unknown       Healthy
gpu-operator               OutOfSync     Healthy
nvidia-dra-driver-gpu      OutOfSync     Progressing
...

Generated Application (broken)

# Application: gpu-operator-pre
spec:
  source:
    repoURL: oci://<registry>/<org>
    path: "005-gpu-operator-pre"          # ← wrong: 'path' is for Git
    targetRevision: <tag>
    helm:
      values: |
        driver:
          version: <version>

ArgoCD error:

Failed to load target state: failed to generate manifest for source 1 of 1:
rpc error: code = Unknown desc = failed to resolve revision "<tag>":
cannot get digest for revision <tag>:
<registry>/<org>:<tag>: not found

The <registry>/<org>:<tag> lookup fails because ArgoCD is treating repoURL as the full OCI artifact and path as a sub-directory inside it, which is Git-source semantics. For OCI Helm sources, ArgoCD needs chart: <name>.

Expected Generated Application

# Application: gpu-operator-pre
spec:
  source:
    repoURL: oci://<registry>/<org>
    chart: gpu-operator-pre                # ← correct for OCI Helm
    targetRevision: <tag>
    helm:
      values: |
        driver:
          version: <version>

By comparison, the sibling gpu-operator Application (which ArgoCD does resolve correctly) already uses chart: gpu-operator against an OCI Helm source.

Suggested Fix

In the argocd-helm deployer's Application template, switch path: NNN-<name> to chart: <name> for child charts that resolve to OCI Helm sources packaged inside the bundle artifact.

Impact

• Critical for any consumer trying to deploy AICR bundles via argocd-helm from an OCI registry
• Silent at bundle build time; only surfaces after helm install on the target cluster
• Cascades into the GPU stack never coming up (driver missing → DRA kubelet plugin blocked)