# Deployment reference files Use one authored YAML per environment: `staging-deploy.yaml`, `prod-deploy.yaml`, or any filename. Account/infrastructure settings live under `advanced`; no separate setup-profile or local values/installer JSON is required. The CLI generates its own private plan, values, credentials and recovery files. Never discard those to retry an uncertain operation. | Mode | Reference | Current status | | --- | --- | --- | | Azure | [azure/deployment.yaml](azure/deployment.yaml) | Experimental coordinator; fresh combined cloud qualification pending | | AWS | [aws/deployment.yaml](aws/deployment.yaml) | Experimental coordinator; fresh combined cloud qualification pending | | local-k8s | [local-k8s/deployment.yaml](local-k8s/deployment.yaml) | Existing local kind cluster; reuse journaled installer | | local-docker | Existing `deploy/service/stack.mjs` / Compose workflow | One-YAML CLI orchestration is roadmap; command refuses effects | | remote-k8s | Existing cluster + explicit kubeconfig/network route required | Roadmap until implemented and tested on a real remote cluster | Reference identities, images, versions and paths are placeholders. Replace them. Examples do not authorize provisioning, provide real image digests or promise a spending cap. They omit `trial`: no automatic expiry by default. For disposable data only, optionally add `trial: {expiresAfterHours: 6, deleteData: true}`. ## Azure and AWS ```sh owcluster deploy azure up --config azure/deployment.yaml --plan-only \ --max-estimated-cost-usd 100 owcluster deploy aws up --config aws/deployment.yaml --plan-only ``` A cost ceiling is optional, is plan-only, and checks an estimate rather than actual billing. It fails when modeled prices are unknown. Review the price timestamp, assumptions and excluded variable costs. Omit `--plan-only` after reviewing the provisioning/grant bundle. Noninteractive use requires its exact `--accept-plan` and `--accept-bootstrap` digests. Add `--keep-access` for the private management route and use the generated `connection.json` to open the UI. Replace `advanced.infrastructure` pins and identity/resource references in the same YAML. Azure needs its actual subscription, installer object ID (not app/client ID), existing workload/control groups, Entra administrator group, SSH public key, operator IPv4, AKS/Ubuntu versions and versioned Key Vault password reference. The documentation IPv4 and image digests are fake. The Key Vault uses RBAC and ARM deployment access. Workload/control prerequisites and the cleanup image must exist. The private key stays local; select `advanced.azure.sshPrivateKeyFile` if needed. AWS needs its actual account/profile, installer ARN, region/availability zones, reviewed versions/add-ons, versioned S3 cleanup package/checksum and regional ARM64 AMI for the optional SSM bridge. Replace the fake AMI and cleanup object reference. Use the matching reviewed identity scopes and current regional preflight. Set `advanced.release` to a generated release manifest from `up/package.py`: ARM64 for AWS, AMD64 for Azure, including a UI image if enabled. Release manifests are generated artifacts, not another hand-authored environment config. Versions and package distribution remain subject to qualification; a CLI login alone is not permission to provision. Export model keys using the named environment variables or use file references. No secret values belong in deployment YAML. [Installer details](../up/README.md) explain current prerequisites, pricing scope, trial data deletion, private access and mode limitations. ## Existing local Kubernetes Both local and remote Kubernetes require an already running cluster, a working kubeconfig/context, permissions and network access. Current `local-k8s` is qualified only for explicit `kind-*` contexts and dedicated `ow-service-m4-*` namespaces. It does not create Kubernetes from SSH machines or install a database. Edit the one local YAML: kubeconfig/context/namespace, service/runner/UI image digests, chart path, model/provider references, existing OpenShell coordinates and an existing PostgreSQL URL reference + public CA file. The database connection verifies TLS; the credentials and model keys are read privately only after plan approval. Kubernetes must be able to reach PostgreSQL and pull/use those images. `scripted` is a local fixture option, not a substitute for live-provider qualification. ```sh owcluster deploy local-k8s plan --config local-k8s/deployment.yaml --state .ows/local owcluster deploy local-k8s up --config local-k8s/deployment.yaml --state .ows/local \ --accept-plan PLAN_DIGEST owcluster deploy local-k8s status --state .ows/local owcluster management-ui status --state .ows/local/installation ``` Use the actual digest printed by plan. The shared installer checks cluster and namespace identities, OpenShell compatibility/storage/network prerequisites and original Helm receipts. `advanced.openshell.mode: existing` reuses the named installation; `install` accepts a reviewed pinned archive and inline `values` that the CLI materializes internally. It never silently upgrades customer-managed OpenShell. The existing installer enforces its pinned archive/image requirements. No new sandbox/model run is implied by successful installation readiness. The CLI generates the private API certificate and maps `managementUI.port` to the loopback preview. Keep and trust that CA; do not disable TLS verification. Set UI/recording preferences in YAML or use the independent feature commands. Native lower-level `owcluster deploy local` / `deploy cluster` remain compatibility interfaces for existing journals; new examples use `local-k8s`. The new `local-docker` and `remote-k8s` command names are discoverable but report `phase: roadmap`, return failure, and create no resources. They must not be used as proof of successful deployment or qualification. Docker's existing separate scripts remain available pending a complete one-file CLI adapter.