Multi-provider host auto-provisioning (HostProvider abstraction)

[mathaix]

Goal

One-click "Provision Host" from the UI on any cloud (GCP, AWS, Hetzner, …), using the same on-host bootstrap as bare-metal enrollment. Today auto-provisioning is GCP-only and takes a different (GCE-metadata) code path than enrollment.

Background / current state

There are two onboarding paths today:

Path	Providers	Mechanism
Enrollment (universal)	GCP, AWS, Hetzner, bare-metal	provision-host.sh + curl .../api/agent/install | bash -s -- <token> → /api/agent/register
Auto-provision from UI (GCP-only)	GCP	backend/internal/provisioner/provisioner.go → GCE Instances.Insert + metadata injection

The enrollment path is already provider-agnostic and is documented as the unified path (see docs/getting-started.md). The GCP auto-provisioner is a separate, GCP-specific code path (injects agent-token/tunnel-token via GCE metadata; the agent self-configures from the metadata server). No AWS/Hetzner/user-data/cloud-init provisioning code exists.

Proposed design

Make cloud auto-provision a thin wrapper around the same bootstrap the enrollment path runs:

1. Control plane mints an enrollment token.
2. Renders user-data = provision-host.sh (install deps) + curl .../api/agent/install | bash -s -- <token>.
3. Creates the VM on the chosen provider with that user-data (GCE startup-script / EC2 user-data / Hetzner cloud-init).
4. The host registers via /api/agent/register exactly like a bare host — tunnel, host record, agent token, and GCS creds all flow through the existing register endpoint. No provider-specific code runs on the host.

Provider-specific code reduces to a thin interface:

type HostProvider interface {
    CreateInstance(ctx context.Context, spec ProvisionSpec, userData string) (instanceID string, err error)
    DeleteInstance(ctx context.Context, instanceID string) error
}

• GCP impl first (wraps the existing compute Insert, minus the metadata/tunnel logic, plus startup-script).
• AWS impl = EC2 RunInstances with user-data.
• Hetzner impl = hcloud server create with user_data (cloud-init).

Scope / tasks

• Define HostProvider interface + ProvisionSpec (machine type/size, region/zone, image, nested-virt).
• Render the unified bootstrap user-data (reuse scripts/provision-host.sh + enrollment install).
• Refactor provisioner.go to: mint token → render user-data → provider.CreateInstance(...); drop the GCP-metadata path so GCP goes through /api/agent/register like every other host.
• GCP HostProvider impl (stock image + startup-script, or keep FC_SNAPSHOT_NAME fast path).
• AWS HostProvider impl (EC2, nested-virt-capable instance types e.g. *.metal).
• Hetzner HostProvider impl (hcloud).
• UI: provider/region/size selection in the "Provision Host" flow.
• Reconciler: provider-aware liveness (GCP has an instance API; others fall back to heartbeat-only — see heartbeat_checker.go).
• Cross-provider artifact distribution: confirm /register GCS-creds hand-off works for AWS/Hetzner pulling the agent + kernel, or add an HTTP artifact source.

Related

• Unified path documented in docs/getting-started.md (option A).
• Follow-up: GitHub Action to build + push ocm-agent + Firecracker kernel artifacts to the artifact bucket.

🤖 Generated with Claude Code