Proposal: Map llama-skein reasoning capability in model discovery #34

Why

llama-skein now advertises a reasoning boolean per model in /v1/models
(companion change advertise-reasoning-capability). opencode's
discoverOpenAICompatibleModels defaulted discovered local models to
capabilities.reasoning=false, so reasoning models (streaming reasoning_content
first) rendered as a frozen "not responding" during the think phase.

What

• In discoverOpenAICompatibleModels, read item.reasoning from the /v1/models
  response and set capabilities.reasoning from it. A hand-configured provider
  model capability still wins over discovery (existing-first merge preserved).

Non-goals

• Changing reasoning rendering itself (already gated on capabilities.reasoning).
• Client regen: discovery reads the raw /models JSON, so no generated-client
  change is required for this mapping.

Disposition (2026-09-18)

Shipped. Shipped: provider.ts maps reasoning on discovery.