llm

type ChatRequest struct {
	// Model name (e.g., "gpt-4", "claude-3-sonnet", "llama2")
	Model string `json:"model"`

	// Conversation messages
	Messages []Message `json:"messages"`

	// Whether to stream the response
	Stream *bool `json:"stream,omitempty"`

	// System prompt (some providers handle this separately from messages)
	System string `json:"system,omitempty"`

	// Generation parameters (unified across providers)
	MaxTokens   *int     `json:"max_tokens,omitempty"`
	Temperature *float64 `json:"temperature,omitempty"`
	TopP        *float64 `json:"top_p,omitempty"`
	TopK        *int     `json:"top_k,omitempty"`
	Stop        []string `json:"stop,omitempty"`
	Seed        *int     `json:"seed,omitempty"`

	// Tools are the tool definitions offered to the model, preserved as
	// raw provider JSON. The definitions are large and provider-shaped;
	// callers that only need the count (e.g. shadow-call classification:
	// the security monitor sends zero tools, the main conversation sends
	// the full set) should use len(Tools).
	Tools []json.RawMessage `json:"tools,omitempty"`

	// Provider-specific fields that don't map to common parameters
	Extra map[string]any `json:"extra,omitempty"`

	// RawRequest preserves the original request payload for cases where
	// parsing is incomplete or for debugging.
	RawRequest json.RawMessage `json:"raw_request,omitempty"`
}

type ChatResponse struct {
	// Model that generated the response
	Model string `json:"model"`

	// Response timestamp
	CreatedAt time.Time `json:"created_at,omitzero"`

	// The assistant's response message
	Message Message `json:"message"`

	// Whether generation is complete (for streaming)
	Done bool `json:"done"`

	// Stop reason (e.g., "stop", "length", "tool_use", "end_turn")
	StopReason string `json:"stop_reason,omitempty"`

	// Token usage and timing metrics
	Usage *Usage `json:"usage,omitempty"`

	// Provider-specific fields that don't map to common parameters
	Extra map[string]any `json:"extra,omitempty"`

	// RawResponse preserves the original response payload for cases where
	// parsing is incomplete or for debugging.
	RawResponse json.RawMessage `json:"raw_response,omitempty"`
}

type ContentBlock struct {
	Type string `json:"type"` // "text", "image", "tool_use", "tool_result", "thinking"

	// Text content (type="text")
	Text string `json:"text,omitempty"`

	// Image content (type="image")
	ImageURL    string `json:"image_url,omitempty"`    // URL to image
	ImageBase64 string `json:"image_base64,omitempty"` // Base64-encoded image data
	MediaType   string `json:"media_type,omitempty"`   // MIME type (e.g., "image/png")

	// Tool use (type="tool_use") - assistant requesting tool execution
	ToolUseID string         `json:"tool_use_id,omitempty"`
	ToolName  string         `json:"tool_name,omitempty"`
	ToolInput map[string]any `json:"tool_input,omitempty"`

	// Tool result (type="tool_result") - result from tool execution
	ToolResultID string `json:"tool_result_id,omitempty"` // References the tool_use_id
	ToolOutput   string `json:"tool_output,omitempty"`
	IsError      bool   `json:"is_error,omitempty"`

	// Thinking (type="thinking") - Anthropic extended-thinking blocks.
	// Anthropic emits thinking as content_block_delta frames with type
	// "thinking_delta" followed by a "signature_delta" that authenticates the
	// block. Consumers treat Thinking as opaque text; the signature is persisted
	// so downstream tooling can verify integrity.
	Thinking          string `json:"thinking,omitempty"`
	ThinkingSignature string `json:"thinking_signature,omitempty"`

	// Content (type="web_search_tool_result" and other server-tool results) -
	// the raw result payload Anthropic returns inline on the block, captured
	// verbatim as JSON so the variable result-object shapes survive without
	// imposing a schema. ToolResultID links it to the paired server_tool_use.
	Content json.RawMessage `json:"content,omitempty"`
}

type ConversationTurn struct {
	Provider string        `json:"provider"`
	Request  *ChatRequest  `json:"request"`
	Response *ChatResponse `json:"response"`
}

type ErrorResponse struct {
	Error string `json:"error"`
}

type Message struct {
	Role    string         `json:"role"`    // "system", "user", "assistant", "tool"
	Content []ContentBlock `json:"content"` // Array of content blocks
}

type RequestParams struct {
	System      string   `json:"system,omitempty"`
	MaxTokens   *int     `json:"max_tokens,omitempty"`
	Temperature *float64 `json:"temperature,omitempty"`
	Stream      *bool    `json:"stream,omitempty"`
	ToolCount   *int     `json:"tool_count,omitempty"`
}

type StreamChunk struct {
	// Model that generated the chunk
	Model string `json:"model"`

	// Chunk timestamp
	CreatedAt time.Time `json:"created_at,omitzero"`

	// The content of this chunk (typically a partial message)
	Message Message `json:"message"`

	// Whether this is the final chunk
	Done bool `json:"done"`

	// Index for providers that support multiple parallel completions
	Index int `json:"index,omitempty"`

	// Stop reason (only present on final chunk)
	StopReason string `json:"stop_reason,omitempty"`

	// Usage metrics (typically only present on final chunk)
	Usage *Usage `json:"usage,omitempty"`
}

type Usage struct {
	// Token counts
	PromptTokens     int `json:"prompt_tokens,omitempty"`
	CompletionTokens int `json:"completion_tokens,omitempty"`
	TotalTokens      int `json:"total_tokens,omitempty"`

	// Cache token counts (Anthropic prompt caching)
	CacheCreationInputTokens int `json:"cache_creation_input_tokens,omitempty"`
	CacheReadInputTokens     int `json:"cache_read_input_tokens,omitempty"`

	// TotalDurationNs is the proxy-measured wall-clock time, in nanoseconds,
	// from when the proxy received the client request to when the upstream
	// response was fully assembled. Set uniformly by the proxy across providers
	// — Anthropic and OpenAI don't surface a duration field on the wire, and
	// Ollama's server-internal `total_duration` is intentionally overwritten so
	// aggregate stats compare apples to apples.
	TotalDurationNs int64 `json:"total_duration_ns,omitempty"`

	// PromptDurationNs is provider-reported prompt-evaluation time, in
	// nanoseconds. Populated by providers that surface it (currently Ollama
	// only); left at zero otherwise.
	PromptDurationNs int64 `json:"prompt_duration_ns,omitempty"`
}