modelsetup

type Candidate struct {
	ID      string
	Listed  ListedModel
	Catalog *CatalogModel // nil when models.dev does not describe the model
	// Backend is the protocol the model is served over, from models.dev; for
	// a model models.dev does not describe it is the provider's usual one,
	// and a test request finds the real one (see Probe).
	Backend string
	// Unsupported says why the model cannot be used, such as a protocol
	// kvit-coder does not speak. Empty when it can.
	Unsupported string
}

type Catalog struct {
	FetchedAt time.Time                  `json:"fetched_at"`
	Providers map[string]CatalogProvider `json:"providers"`
}

type CatalogModel struct {
	ID               string            `json:"id"`
	Name             string            `json:"name"`
	Reasoning        bool              `json:"reasoning"`
	ReasoningOptions []ReasoningOption `json:"reasoning_options,omitempty"`
	ToolCall         bool              `json:"tool_call"`
	Status           string            `json:"status,omitempty"`
	Limit            struct {
		Context int `json:"context"`
		Output  int `json:"output"`
	} `json:"limit"`
	Cost *struct {
		Input  float64 `json:"input"`
		Output float64 `json:"output"`
	} `json:"cost,omitempty"`
	// Provider overrides the provider's client library for this model, which
	// is how one gateway serving three protocols is described.
	Provider *struct {
		NPM string `json:"npm"`
	} `json:"provider,omitempty"`
}

type CatalogProvider struct {
	ID     string                  `json:"id"`
	Name   string                  `json:"name"`
	NPM    string                  `json:"npm"`
	API    string                  `json:"api"`
	Env    []string                `json:"env"`
	Models map[string]CatalogModel `json:"models"`
}

type Kind int

type ListedModel struct {
	ID string
	// Context is the context size the endpoint reports, or 0: vLLM puts
	// max_model_len in its model list, and for llama.cpp ServerContext asks
	// /props. Hosted gateways report nothing here; models.dev has it.
	Context int
}

type ProbeResult struct {
	// Backend is the protocol that answered; it differs from the row's when
	// the row's own was refused and another one answered.
	Backend string
	// Err is why no protocol answered; nil when one did.
	Err error
	// KeyRefused is set when the endpoint answered 401 or 403.
	KeyRefused bool
}

type Provider struct {
	// Key is the short name put at the end of a row's id ("kimi-k3-go"), so
	// the same model from two providers gets two ids.
	Key string
	// Name is what the list shows, and Short is what goes in brackets after
	// a model's name ("Kimi K3 (Go)").
	Name, Short string
	Kind        Kind
	// BaseURL is where requests go. MessagesURL is the address for models
	// served over Claude's messages protocol, for which kvit-coder appends
	// /v1/messages itself, so it stops before /v1.
	BaseURL, MessagesURL string
	// KeyEnv is the environment variable a row reads its key from; a key
	// typed into :setup is saved under the same name in credentials.json.
	KeyEnv string
	// KeyPage is the web page where a key is issued.
	KeyPage string
	// CatalogID is the provider's id on models.dev.
	CatalogID string
	// EffortField is where a chat-completions request carries the effort for
	// this provider; see llm.effort_field in docs/configuration.md.
	EffortField string
	// Headers are put on every row for this provider.
	Headers []string
}

type ReasoningOption struct {
	Type   string   `json:"type"`
	Values []string `json:"values,omitempty"`
}