// Provider-wire compat for "modular". Modular Cloud serves open models on an
// OpenAI-compatible API. The provider is catalog-only: models, prices, and
// limits come from the models.dev `modular` entry. Reasoning controls differ
// per model and were measured against the API.

provider "modular" {
	default-model "google/gemma-4-31b-it"
	env "MODULAR_API_KEY"

	thinking-mode "effort"
	// The API takes OpenRouter's `reasoning` object for every model, so a model
	// models.dev adds later works without its own block: on sends
	// `reasoning.effort`, off sends `reasoning.enabled = false`.
	thinking-format "openrouter"

	// MiniMax M3.1 Flash takes a nested `reasoning.effort` and always thinks:
	// `reasoning.enabled = false` still returns reasoning tokens.
	models "MiniMaxAI/MiniMax-M3.1-Flash" {
		thinking-requires-effort #true
		thinking-efforts "low" "medium" "high" "xhigh" "max"
	}
	// MiniMax M3 and Gemma 4 31B switch reasoning with `reasoning.enabled` and
	// ignore any effort, so the ladder is one on-rung beside off: off sends
	// `enabled: false`, and on replaces the effort with `enabled: true`.
	models "minimax/minimax-m3" "google/gemma-4-31b-it" {
		thinking-efforts "high"
		when-thinking {
			extra-body {
				reasoning {
					enabled #true
				}
			}
		}
	}
}