Local-first onboarding

[rlnorthcutt]

Problem statement

There's currently no path to start with local-only inference instead of a cloud provider, and the setup wizard fully blocks users who have no provider configured yet, even though the tool is more useful to explore than a hard block suggests.

What success looks like

The install wizard can detect available resources (GPU, RAM) and suggest a right-sized local model, and a user can skip the wizard entirely and still get into the tool, with a clear path back to setup if nothing is configured.

Impacted repositories

• omnideck

Acceptance criteria

• The wizard inspects available GPU/RAM and offers 3 local-model choices with one recommended.
• A user can skip the wizard entirely and land in the tool.
• If no provider (cloud or local) is configured, a persistent banner says so and links back to the wizard.

[rlnorthcutt]

Related

• See First-class local inference #9 (First-class local inference) — the container GPU access, memory sizing, and docs work this onboarding flow depends on. Related but distinct: First-class local inference #9 is the underlying capability; this epic is the wizard UX built on top of it.