model

type DataRef struct {
	MimeType string         `json:"mime_type,omitempty"`
	Value    map[string]any `json:"value"`
}

type FileRef struct {
	URI      string `json:"uri"`
	MimeType string `json:"mime_type,omitempty"`
	Name     string `json:"name,omitempty"`
}

type MediaRef struct {
	URL       string `json:"url,omitempty"`
	Base64    string `json:"base64,omitempty"`
	MediaType string `json:"media_type,omitempty"`
}

type Message struct {
	Role  Role   `json:"role"`
	Parts []Part `json:"parts"`
}

type Part struct {
	Type       PartType    `json:"type"`
	Text       string      `json:"text,omitempty"`
	Image      *MediaRef   `json:"image,omitempty"`
	Audio      *MediaRef   `json:"audio,omitempty"`
	File       *FileRef    `json:"file,omitempty"`
	Data       *DataRef    `json:"data,omitempty"`
	ToolCall   *ToolCall   `json:"tool_call,omitempty"`
	ToolResult *ToolResult `json:"tool_result,omitempty"`
}

type PartType string

type Role string

type StreamChunk struct {
	Role         Role       `json:"role,omitempty"`
	Content      string     `json:"content,omitempty"`
	ToolCalls    []ToolCall `json:"tool_calls,omitempty"`
	FinishReason string     `json:"finish_reason,omitempty"`
}

type TokenUsage struct {
	InputTokens  int64 `json:"input_tokens"`
	OutputTokens int64 `json:"output_tokens"`
	TotalTokens  int64 `json:"total_tokens"`

	// CachedInputTokens is the subset of InputTokens that hit the
	// provider's prompt cache and are therefore billed at a reduced
	// rate (typically 10x cheaper). It is always <= InputTokens.
	//
	// Different providers expose this under different names but the
	// semantics are aligned: OpenAI / DeepSeek / Azure / Qwen report
	// `usage.prompt_tokens_details.cached_tokens`, Anthropic /
	// Minimax report `usage.cache_read_input_tokens`, ByteDance
	// Doubao reports `usage.prompt_tokens_details.cached_tokens`.
	// The Generate / stream paths normalise these into this single
	// field so callers can compute a cache hit-rate uniformly via
	// CachedInputTokens / InputTokens without provider-specific
	// branching.
	//
	// Zero means either the provider does not expose cache stats
	// or no cache hit occurred on this call. omitempty keeps the
	// wire format stable for providers that never set it.
	CachedInputTokens int64 `json:"cached_input_tokens,omitempty"`

	// Model is the resolved LLM model name this usage came from
	// (e.g. "gpt-4o", "claude-3-7-sonnet-20250219"). Empty when the
	// reporter does not know which model produced the call (test
	// engines, aggregate reports). Hosts that bucket usage by model
	// for budgeting / quota enforcement consume this field.
	Model string `json:"model,omitempty"`

	// LatencyMs is the wall-clock latency of the producing call in
	// milliseconds. Zero when not measured. Used by sandbox hosts to
	// surface per-call timing on the same dimension as token counts
	// (avoids a parallel timing channel).
	LatencyMs int64 `json:"latency_ms,omitempty"`

	// CostMicros is the cost of the producing call in micro-units of
	// the host's configured currency (micro-USD = USD * 1_000_000).
	// Integer math is used so cumulative budgets do not drift.
	// Zero when no pricing catalog is configured. Hosts enforcing $
	// budgets accumulate this field.
	CostMicros int64 `json:"cost_micros,omitempty"`
}

type ToolCall struct {
	ID        string `json:"id"`
	Name      string `json:"name"`
	Arguments string `json:"arguments"`
}

type ToolDefinition struct {
	Name        string         `json:"name"`
	Description string         `json:"description"`
	InputSchema map[string]any `json:"input_schema"`
}

type ToolResult struct {
	ToolCallID string `json:"tool_call_id"`
	Content    string `json:"content"`
	IsError    bool   `json:"is_error,omitempty"`
}

type Usage struct {
	InputTokens       int64 `json:"input_tokens"`
	CachedInputTokens int64 `json:"cached_input_tokens,omitempty"`
	OutputTokens      int64 `json:"output_tokens"`
}