Deploy through plan and human approval
High-risk actions never reach execution without an independent approval boundary.
Platform engineering and private cloud
Build, import and operate application platforms.
Turns raw servers or existing Kubernetes environments into standardized RKE2, OKD or imported Kubernetes platforms with catalog, fleet, release, policy, disconnected delivery and Day‑2 operations.

The deploy request is recorded and a plan with its impact is built.
Kubernetes, registry, identity, policy, monitoring, upgrades and recovery become another operations burden if they are assembled without one product lifecycle.
Blueprints and profiles replace environment-specific assembly.
RKE2, OKD and imported clusters share a product boundary without fake symmetry.
Health, drift, node lifecycle and upgrade stay under fleet operations.
Desired state and artifacts retain identity and provenance.
Factory owns platform lifecycle while target distributions keep their native capabilities.
Build a managed platform or enroll an existing cluster.
Versioned platform capabilities and templates.
Operate many targets through one product model.
Forgejo records intent; Argo CD reconciles it.
zot keeps canonical OCI identity; oc-mirror v2 + zot provide sealed acquisition.
Application release and environment binding on observed platform capabilities.
Policy, backup run, restore and drill for managed targets.
Drain, replace, maintenance and distribution-aware upgrade.
The Factory does not migrate itself into the target distribution. RKE2 management stays self-contained while RKE2, OKD and imported Kubernetes are distinct target identities.
| Target | Entry path | Ownership | Use |
|---|---|---|---|
| Managed RKE2 | Factory build | Platform lifecycle under Product Control Plane | Standard private application platform |
| Existing Kubernetes | Import + discovery | Existing runtime + 4SO lifecycle boundary | Bring existing clusters under management |
| Managed OKD | Distribution-aware build | OKD owns CVO/MCO/OVN/OLM/SCC behavior | OpenShift-compatible target |
| Existing OKD | Import + capability discovery | Native OKD + 4SO product workflows | Manage without duplicate platform stacks |
A target remains a managed product resource through capability discovery, releases, node change, upgrade and recovery.
Each profile combines distribution identity, capability discovery, policy, supply chain and lifecycle. OKD and RKE2 are deliberately not forced into artificial implementation symmetry.
A lightweight, self-contained Kubernetes platform.
Uses OKD-native capabilities rather than duplicating them.
Adopt an existing cluster without rebuilding it.
Application delivery over a managed target.
Acquire, mirror and install without runtime Internet dependency.
Operate targets after creation.
For engineers who want to know how the product works underneath: components, boundaries, failure behavior and the operations it runs.
The architecture avoids making Git, Argo or the target cluster an accidental second product source of truth.
| Layer | Component / authority | Responsibility | Boundary |
|---|---|---|---|
| Product authority | PostgreSQL + Platform API | Org/Project, Catalog, Blueprint, Fleet, Release, durable operations and audit | Product and lifecycle authority; not a runtime target |
| Desired state | Forgejo | Versioned desired configuration and release inputs | Desired state, not runtime truth |
| Reconciliation | Argo CD | Applies desired state and detects drift | A reconciler; never an independent source of truth |
| Artifact plane | zot OCI registry | Image/artifact delivery, mirroring and disconnected inventory | Artifact identity and provenance are preserved |
| Management plane | Self-contained RKE2 | Control services, identity, PostgreSQL, Git/registry and operator APIs | Kept separate from the target distribution |
| Targets | RKE2 · OKD · Existing Kubernetes | Capability discovery, workload runtime and Day-2 lifecycle | Distribution-native capabilities are kept; no duplicate stack is imposed |
Target and delivery contract. Factory does not build a vague abstraction over Kubernetes: product authority, desired state, reconciliation, the artifact plane and the distribution-native runtime keep separate boundaries.
| Domain | Component / mechanism | Behavior |
|---|---|---|
| Managed OKD | openshift-install / oc + Redfish boot media | Bare-metal provisioning and cluster lifecycle follow OKD’s own distribution contract. |
| Disconnected OKD | oc-mirror v2 → zot | Acquisition, inventory and mirror are sealed and reviewable before install. |
| Application runtime trait | Optional Dapr sidecar capability | Dapr can serve service invocation, pub/sub, bindings and resiliency, but never takes over product state, workflow or secret authority. |
Risk, approval, retry policy and runtime verification belong to the operation model rather than an informal runbook.
High-risk actions never reach execution without an independent approval boundary.
Forgejo intent and Argo reconciliation do not create a second product authority.
Artifact identity is sealed before target installation.
Raw infrastructure to registered target.
Failure and recoveryPLAN_FAILED remains pre-mutation; unknown side effects require reconciliation.
Enrollment without rebuilding the cluster.
Failure and recoveryIdentity or capability conflict returns to the operator.
Release binds desired state to observed runtime.
Failure and recoveryUnknown outcome is read back before any redispatch.
Supply chain is sealed before install.
Failure and recoveryMissing or substituted artifacts fail closed.
Day‑2 remains distribution-aware.
Failure and recoveryRollback is only used where the underlying workflow supports it.
Drift and data protection deserve visible lifecycle paths too.
Resolve desired/observed differences without overwriting distribution-native ownership.
Failure and recoveryIf the difference belongs to a target-native subsystem, Factory preserves that ownership instead of forcing symmetry.
Data protection continues through isolated restore and workload verification.
Failure and recoveryAn unverified artifact or unverified restore is not accepted as recovery evidence.
| Action | Admission / Preconditions | Execution owner | Success criterion | Failure / recovery |
|---|---|---|---|---|
| Create Platform | Resolvable profile + provider/target inputs | Factory operation executor | Observed target matches desired profile | PLAN_FAILED / retry / recovery path |
| Import Cluster | Identity + access + capability discovery | Import workflow | Registered target with observed capabilities | Conflict requires operator resolution |
| Deploy Application | Immutable release + binding + policy | Forgejo/Argo + bounded executor | Observed deployment/service evidence | Unknown result is not blindly retried |
| Replace Node | Exact node identity + safety admission | Leased/fenced operation | Replacement ready; old node safely removed | Lease loss → retry wait or operator |
| Upgrade | Compatibility + health + evidence gate | Distribution-aware workflow | Observed version/capability health | Rollback only where real |
| Disconnected Delivery | Source lock + digest + inventory | acquire/mirror workflow | Target artifact identity matches sealed inventory | Mismatch fails closed |
Failure states: PLAN_FAILED / FAILED / ROLLBACK_FAILED / NEEDS_OPERATOR / CANCELLED. Retry applies only to allowed failure classes; rollback is claimed only where the workflow supports it.
The operator works with targets, fleet, releases and lifecycle actions rather than per-cluster tribal procedures.
4SO owns the product lifecycle; the target keeps the native semantics that make it trustworthy.