Currently, to override the default capabilities (like constrained: 'all') for a model using the openAICompatible plugin, one must implement a custom resolver that manually defines the model using defineCompatOpenAIModel and compatOaiModelRef.

This is verbose and inconvenient, especially when just trying to enable a capability that the model supports but isn't enabled by default in the generic model info (e.g. native structured output for Fireworks AI).

It would be much better if:

1. openAICompatible plugin accepted a capabilities or modelInfo override in its configuration.
2. Or if ai.generate respected the info property passed in modelRef at runtime (currently resolveModel ignores it).

Example of current workaround:

openAICompatible({
  name: 'fireworks',
  apiKey: process.env.FIREWORKS_API_KEY,
  baseURL: '...',
  resolver: async (client, actionType, actionName) => {
    if (actionType === 'model') {
      return defineCompatOpenAIModel({
        name: actionName,
        client,
        modelRef: compatOaiModelRef({
          name: actionName,
          info: {
            supports: {
              multiturn: true,
              media: true,
              tools: true,
              toolChoice: true,
              systemRole: true,
              constrained: 'all', // <--- The goal
            },
          },
        }),
      });
    }
    return undefined;
  },
})