llm

type AnthropicAgentAdapter struct {
	// contains filtered or unexported fields
}

type Client interface {
	// Complete sends a prompt to the LLM and returns a response.
	//
	// Inputs:
	//   ctx - Context for cancellation and timeout
	//   request - The completion request
	//
	// Outputs:
	//   *Response - The LLM response
	//   error - Non-nil if the request failed
	Complete(ctx context.Context, request *Request) (*Response, error)

	// Name returns the provider name (e.g., "anthropic", "openai").
	Name() string

	// Model returns the model being used.
	Model() string
}

type CompletionCall struct {
	Request   *Request
	Timestamp time.Time
}

type EmptyResponseError struct {
	// Duration is how long the request took before returning empty.
	Duration time.Duration

	// MessageCount is the number of input messages sent.
	MessageCount int

	// Model is the model that returned the empty response.
	Model string
}

type GeminiAgentAdapter struct {
	// contains filtered or unexported fields
}

type Message struct {
	// Role is "user", "assistant", or "system".
	Role string `json:"role"`

	// Content is the text content.
	Content string `json:"content"`

	// ToolCalls contains tool invocations (for assistant messages).
	ToolCalls []ToolCall `json:"tool_calls,omitempty"`

	// ToolResults contains tool results (for tool messages).
	ToolResults []ToolCallResult `json:"tool_results,omitempty"`
}

type MockClient struct {
	// contains filtered or unexported fields
}

type OllamaAdapter struct {
	// contains filtered or unexported fields
}

type OpenAIAgentAdapter struct {
	// contains filtered or unexported fields
}

type ReActParseResult struct {
	// Thought is the LLM's reasoning (may be empty).
	Thought string

	// Action is the tool name to invoke (empty if no action).
	Action string

	// ActionInput is the JSON arguments (empty if no action).
	ActionInput string

	// FinalAnswer is the final response (empty if not done).
	FinalAnswer string

	// RemainingText is any unparsed text.
	RemainingText string
}

type Request struct {
	// SystemPrompt is the system message.
	SystemPrompt string `json:"system_prompt,omitempty"`

	// Messages is the conversation history.
	Messages []Message `json:"messages"`

	// Tools defines available tools for the LLM.
	Tools []tools.ToolDefinition `json:"tools,omitempty"`

	// ToolChoice controls tool selection behavior.
	// If nil, defaults to "auto" (model decides).
	ToolChoice *ToolChoice `json:"tool_choice,omitempty"`

	// MaxTokens limits the response length.
	MaxTokens int `json:"max_tokens,omitempty"`

	// Temperature controls randomness (0.0-1.0). Set to 0.0 for most
	// deterministic output. Set to a negative value (e.g., -1) to omit
	// from the request and use the provider's default. The Go zero value
	// (0.0) is treated as an explicit "most deterministic" setting.
	Temperature float64 `json:"temperature,omitempty"`

	// StopSequences defines sequences that stop generation.
	StopSequences []string `json:"stop_sequences,omitempty"`

	// ModelOverride allows using a different model for this request.
	// Used for multi-model scenarios (e.g., tool routing with a fast model).
	// Empty string means use the client's default model.
	ModelOverride string `json:"model_override,omitempty"`

	// KeepAlive controls how long the model stays loaded in VRAM.
	// Values: "-1" = infinite, "5m" = 5 minutes (default), "0" = unload immediately.
	// Used to prevent model thrashing when alternating between models.
	KeepAlive string `json:"keep_alive,omitempty"`
}

type Response struct {
	// Content is the text response.
	Content string `json:"content"`

	// ToolCalls contains any tool calls the LLM wants to make.
	ToolCalls []ToolCall `json:"tool_calls,omitempty"`

	// StopReason indicates why generation stopped.
	// Values: "end", "max_tokens", "tool_use", "stop_sequence"
	StopReason string `json:"stop_reason"`

	// TokensUsed is the total tokens consumed (input + output).
	TokensUsed int `json:"tokens_used"`

	// InputTokens is the input token count.
	InputTokens int `json:"input_tokens"`

	// OutputTokens is the output token count.
	OutputTokens int `json:"output_tokens"`

	// Duration is how long the request took.
	Duration time.Duration `json:"duration"`

	// Model is the model that generated this response.
	Model string `json:"model,omitempty"`

	// TraceStep is the CRS trace step recorded for this LLM call.
	// Callers can use this to record the step in the CRS trace recorder.
	// May be nil if CRS recording was not applicable.
	TraceStep *crs.TraceStep `json:"trace_step,omitempty"`
}

type ToolCall struct {
	// ID is a unique identifier for this call.
	ID string `json:"id"`

	// Name is the tool name.
	Name string `json:"name"`

	// Arguments are the tool arguments as JSON.
	Arguments string `json:"arguments"`
}

type ToolCallResult struct {
	// ToolCallID links back to the tool call.
	ToolCallID string `json:"tool_call_id"`

	// Content is the result content.
	Content string `json:"content"`

	// IsError indicates if this is an error result.
	IsError bool `json:"is_error,omitempty"`
}

type ToolChoice struct {
	// Type controls tool selection behavior:
	// - "auto": Model decides whether to call tools (default)
	// - "any": Model MUST call at least one tool
	// - "tool": Model MUST call the specific named tool
	// - "none": Model cannot call tools (text response only)
	Type string `json:"type"`

	// Name is required when Type is "tool". Specifies which tool to force.
	Name string `json:"name,omitempty"`
}