[LinoyBitan1]

Follow-up to #135: sandbox recreation could fail during archive verification or lose workspace files when trusted mounts changed; OpenClaw onboarding could retain an agent model that overrode the configured model; and the portal could report a workspace ready before its UI was reachable.

JIRA: https://redhat.atlassian.net/browse/APPENG-6793

Changes

• Preserve workspace files for environment and mount changes, normalize archive ownership to the sandbox user, retain permissions, and make completed restores safe to retry. Stop before changing a running sandbox when its labels cannot be read.
• Apply the configured model to the OpenClaw main agent while preserving its selected credential binding.
• Add optional inference.models inputs for attached native OpenAI and Anthropic providers. Keep credentials on providers and model selection on sandboxes; retain one startup default through inference.providerRef and inference.model. Optional providers remain disabled in the shipped profile.
• Reject empty inference models and report malformed profile catalog fields.
• Wait for the OpenShell UI before the portal reports readiness.
• Preserve unspecified Secret fields during CLI inference updates, support individual model/URL/provider updates, warn about Vault reconciliation, and remove duplicate native model defaults from the Secret template.

The CLI and Secret-template changes also cover the remaining relevant gaps from #133, using the credential/model separation introduced by #135.

Validation

• Prior focused suites passed: 181 installer/model/catalog tests and 155 CLI/template/portal tests.
• A fresh expanded regression rerun is in progress; its final result will be added here.
• Fresh shell syntax and git diff --check checks passed.

Live OpenShift testing used the pushed branch on k4vgv with optional providers enabled in the test profile:

• NVIDIA, OpenAI gpt-6-luna, and Anthropic claude-sonnet-4-6 each answered an inference request through the OpenClaw gateway with the intended provider and model.
• Repeated inference with NVIDIA detached: OpenAI Luna and Anthropic Sonnet 4.6 both returned 4 to the arithmetic prompt. The gateway model picker exposed only OpenAI and Anthropic.
• With Anthropic listed first and OpenAI selected in inference.providerRef, OpenAI Luna remained the startup default. Existing conversations retain their selected model.
• Authenticated chat access returned HTTP 200. The user also confirmed both remaining providers work in the chat UI.
• Sandbox recovery and portal lifecycle/readiness were exercised during the earlier follow-up deployment tests.

The two-provider experiment explicitly detached NVIDIA and removed its saved OpenClaw catalog entries from the existing workspace; removing a provider from the YAML alone does not prune existing runtime state.