| name | kcc-direct-controller-implementer |
| description | Implement the controller, mappers, and fuzzer for a direct KCC resource, ensuring package isolation and CI compliance. Use this when implementing the main reconciliation logic for a "direct" resource. |
KCC Direct Controller Implementer
This skill guides the implementation of the controller, mappers, and fuzzer for direct KCC resources, with a focus on package isolation to prevent symbol collisions and rigorous CI validation.
Inputs
resource_kind: The KCC Kind.
package_path: The isolated package directory (e.g., pkg/controller/direct/vertexai/examplestore/).
proto_package: The GCP proto package (e.g., google.cloud.aiplatform.v1).
Workflow
-
Package Isolation:
You MUST implement all controller-related logic in the provided package_path. This prevents symbol collisions in mapper.generated.go.
-
Brownfield Migrations:
For brownfield resources (with existing terraform/DCL controllers), we do NOT immediately change the default controller to direct in Go type labels or static config map defaults.
- Keep the default controller as legacy (TF/DCL), so users can opt-in gradually.
- Both direct and legacy controllers are now automatically tested dynamically if a direct controller is available in
static_config.go (under SupportedControllers), so there is no longer any need to manually modify tests/e2e/unified_test.go.
-
Controller Structure & Patterns:
- Proto Format Desired State: Convert the KRM Spec to its Proto representation once in
AdapterForObject and store it as a proto struct pointer (e.g., *pb.MyResource or desired) in the adapter, rather than duplicating conversion logic in both Create and Update. The adapter should avoid holding references to raw KRM objects for desired state to keep the interfaces clean, consistent with actual (which is also of proto type), and avoid redundant conversions. This ensures that desired is stored in the same proto format as actual.
- Handling Non-API / KRM-only Spec Fields: If the KRM Spec has fields that are not represented in the GCP resource's proto message (such as client-side behavioral options or custom installation flags, e.g.,
SkipInitialVersionCreation), do NOT mix them into the proto. Instead, parse and store them as separate, explicitly-named fields on the adapter struct (e.g., desiredSkipInitialVersionCreation bool).
- NormalizeReferences: Always call
common.NormalizeReferences in AdapterForObject to resolve any resource references:
if err := common.NormalizeReferences(ctx, reader, obj, nil); err != nil {
return nil, fmt.Errorf("normalizing references: %w", err)
}
- Identity Parent Paths: Create a
ParentString() method on the resource's identity type (e.g., KMSCryptoKeyIdentity) instead of constructing formatting string patterns manually inside the controller. This keeps parent paths canonical and reusable.
- Client Creation Options & Go Client Preference:
- Prefer GAPIC REST Client over gRPC (
go-client): When building the GCP client in m.client(ctx), ALWAYS prefer using the official Google Cloud Go client library (GAPIC client, e.g., cloud.google.com/go/<service>/apivX, or discovery client google.golang.org/api/...). Many GAPIC libraries provide both REST (NewFooRESTClient) and gRPC (NewFooClient) constructors. You MUST prefer the REST variant (NewFooRESTClient) initialized with m.config.RESTClientOptions().
- Why: REST GAPIC clients integrate cleanly with HTTP golden traffic recording (
_http.log), mock layer HTTP alignment, and avoid gRPC transport/auth quirks, while still automatically wrapping Long-Running Operations (*service.FooOperation), OAuth scopes, and retry logic.
- Client Construction Example (GAPIC REST Primary Choice): Retrieve configuration options using
m.config.RESTClientOptions() and construct the REST client:
var opts []option.ClientOption
opts, err := m.config.RESTClientOptions()
if err != nil {
return nil, err
}
gcpClient, err := gcp.NewMyResourceRESTClient(ctx, opts...)
if err != nil {
return nil, fmt.Errorf("building MyResource REST client: %w", err)
}
- Fallback Order & Troubleshooting:
- GAPIC REST Client (
NewFooRESTClient): Primary preference.
- GAPIC gRPC Client (
NewFooClient): Try with m.config.GRPCClientOptions() if the REST constructor is missing or fails during live GCP testing.
- Raw Protobuf gRPC (
pb.NewFooClient(conn)): Only fall back to dialing raw gRPC conns (grpc.Dial) if no GAPIC Go client library struct/constructor exists for the API.
- Diff Comparison & Structured Diff (
tags.DiffForTopLevelFields): Always prefer using top-level field tags-based diff comparison via tags.DiffForTopLevelFields over recursive/magical comparison functions (such as common.CompareProtoMessageStructuredDiff or common.CompareProtoMessage which are deprecated/discouraged due to unpredictable behaviors in BasicDiff).
func compareResource(ctx context.Context, actual, desired *pb.MyResource) (*structuredreporting.Diff, *fieldmaskpb.FieldMask, error) {
maskedActual, err := mappers.OnlySpecFields(actual, MyResourceSpec_FromProto, MyResourceSpec_ToProto)
if err != nil {
return nil, nil, err
}
maskedActual.Name = desired.Name
clonedDesired := proto.CloneOf(desired)
populateDefaults := func(obj *pb.MyResource) {
if obj.SomeDefaultedField == nil {
obj.SomeDefaultedField = ...
}
}
populateDefaults(maskedActual)
populateDefaults(clonedDesired)
diffs, updateMask, err := tags.DiffForTopLevelFields(ctx, clonedDesired.ProtoReflect(), maskedActual.ProtoReflect())
if err != nil {
return nil, nil, err
}
return diffs, updateMask, nil
}
- Reconciling Empty or Incomplete LRO Responses: Many GCP APIs (such as Dataproc's
UpdateCluster LRO) return an empty response (google.protobuf.Empty), or do not fully populate read-only status fields (such as state, metrics, or instance names) during resource creation. If you map status directly from such incomplete/empty LRO responses, you will inadvertently clear status fields in Kubernetes.
- Rule: Always perform a GET operation (
Get<Resource>) immediately after a Create or Update LRO successfully completes to fetch the fully-populated resource before calling updateStatus.
- Propagating KRM Metadata Labels: Metadata labels (such as
managed-by-cnrm: true or custom user-supplied labels) must be explicitly mapped and propagated on both Create and Update operations so they are correctly synchronized to GCP.
- Structured Reporting & updateStatus: In the
Update method, use diffs.HasDiff() to report exact diffs back to the user via structuredreporting.ReportDiff. Always call a helper updateStatus function to update the Kube status at the end of both Create and Update reconciliation paths:
func (a *MyResourceAdapter) updateStatus(ctx context.Context, op directbase.Operation, latest *pb.MyResource) error {
mapCtx := &direct.MapContext{}
status := MyResourceStatus_FromProto(mapCtx, latest)
if mapCtx.Err() != nil {
return mapCtx.Err()
}
return op.UpdateStatus(ctx, status, nil)
}
- Delete Idempotency Check: In the
Delete method, check if the resource has already been deleted using direct.IsNotFound(err) to ensure idempotency, returning true, nil to gracefully exit:
if err != nil {
if direct.IsNotFound(err) {
return true, nil
}
return false, err
}
- Immutable Resources: If a direct resource is completely immutable in GCP (meaning no fields can be updated once created), the
Update method must STILL perform the comparison check on spec fields. If a diff is detected, return a descriptive error (e.g. fmt.Errorf("<Kind> is immutable and cannot be updated")) so that the error/diff is surfaced on the resource status rather than silently doing nothing. Also, register the model using registry.CannotBeDeleted() if deletion is not supported.
- Mutable-but-Unreadable Fields: If certain spec fields (such as keys, passwords, or specific metadata/launchStage fields) are mutable but cannot be read back from the GCP API (meaning they are write-only or missing from the GET response), never check
desired inside populateDefaults. Instead:
- Check the resource's
updateTime or change cookie to see if the GCP side is fully reconciled and unchanged (meaning generation == observedGeneration and GCP updateTime matches status updateTime on a successful/Ready status).
- If the update time matches and there are no external modifications, copy those mutable-but-unreadable fields from
desired to maskedActual in the comparison function to avoid false diffs.
- If the update time does not match (or the resource is not ready), do NOT copy them. This treats the unreadable fields as having changed (since we cannot prove they didn't), correctly forcing an update.
-
Mappers:
Verify mapper.generated.go and manual mappers. Ensure all references use the standard Ref pattern.
-
Fuzzer:
Implement <resource_lower>_fuzzer.go and register it with fuzztesting.RegisterKRMFuzzer.
- Fuzzer Implementation Guidelines:
- ALWAYS prefer and encourage using the fluent
f.SpecField(".foo") and f.StatusField(".bar") methods rather than inserting directly into SpecFields/StatusFields sets (e.g. avoid f.SpecFields.Insert...).
- For unhandled or unimplemented fields, prefer highly descriptive helper methods (such as
f.Unimplemented_NotYetTriaged(".baz"), f.Unimplemented_Identity(".id"), or other specific variants) rather than standard inserts into UnimplementedFields sets. This categorizes why we are not handling specific fields.
-
Final Generation & Reporting:
Run make ready-pr from the repository root. This is a critical step that:
- Runs
make fmt and make vet.
YOU MUST COMMIT ALL RESULTING CHANGES.
-
Validation & Last-Mile Tests:
Run the following tests to ensure CI compliance:
- Fuzzing:
dev/ci/presubmits/fuzz-roundtrippers
- E2E Scaffolding:
dev/ci/presubmits/tests-e2e-fixtures-direct
- Schema Integrity:
go test ./pkg/crd/template/...
- API Field Coverage:
go test ./tests/apichecks/....
- If
TestCRDFieldPresenceInTestsForAlpha fails, you may need to regenerate the exceptions file by running it with WRITE_GOLDEN_OUTPUT=1.
Journaling
Append any controller/reconciliation quirks or API check hurdles to .gemini/journals/<service>.md using the format described in the kcc-agentic-journaler skill.