@fhxgs/dsh-model-hub

Configuration

The plugin registers one settings namespace, model-hub:

model-hub:
  picker:
    mode: include                 # 'all' | 'include'; absent means "the user has not answered"
    include:                      # honored under mode: include; provider and model must be paired
      - { provider: kimi-coding, model: k3 }
    preferredEffort: high         # optional; the middle level of the effort chain
  subagent: inherit               # 'inherit' | { provider, model, reasoningEffort? }

• An empty document resolves to { picker: { include: [] }, subagent: 'inherit' }. The materialized empty array reads as unanswered, never as an empty allowlist.
• The global default model is not here. It stays in the official agent-default-model namespace, which the Host half reads and never writes; only the browser half's destructive-confirmation dialog changes it, through the official settings.mutate with an expectedRevision.
• model-hub-providers is a second, deliberately separate namespace addressing the built-in provider domain — the catalog cache key digests every directory namespace, so keeping the visibility policy out of it stops a policy edit from invalidating the whole model catalog.
• A model-hub: section on disk that the schema rejects fails namespace registration, and the plugin does not mount at all. Recovery is to fix or delete that section and restart the profile.