providers

type AnthropicProvider struct {
	*BaseProvider
	// contains filtered or unexported fields
}

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

type BedrockProvider struct {
	*BaseProvider
	// contains filtered or unexported fields
}

type BuiltinFactory func(ProviderConfig) Provider

type CacheControl struct {
	Type string `json:"type"`          // "ephemeral"
	TTL  string `json:"ttl,omitempty"` // "" / "5m" / "1h" — Anthropic-style
}

type Compat struct {
	// ProviderName is used in errors, Response.Provider, etc.
	ProviderName string

	// DefaultBaseURL is the fallback when ProviderConfig.BaseURL is empty.
	DefaultBaseURL string

	// EndpointPath is appended to BaseURL for chat completions.
	// Default: "/chat/completions"
	EndpointPath string

	// ExtraHeaders are sent with every request (e.g. OpenRouter: HTTP-Referer).
	ExtraHeaders map[string]string

	// StreamHeaders are sent only for stream requests.
	StreamHeaders map[string]string

	// MaxTokensField overrides the JSON key for max tokens.  Default: "max_tokens".
	MaxTokensField string

	// MaxStopSequences limits stop sequences sent.  0 = unlimited.
	MaxStopSequences int

	// OmitStop suppresses stop sequences entirely.
	OmitStop bool

	// SupportsJSONSchema enables full json_schema in response_format.
	SupportsJSONSchema bool

	// JSONSchemaToPrompt injects the JSON schema into the last user message
	// instead of using response_format.  Used by GLM.
	JSONSchemaToPrompt bool

	// ThinkingMapper converts ThinkingConfig into provider-specific request
	// body fields. Returns nil to skip. If nil, thinking is omitted and a
	// portability warning is emitted.
	ThinkingMapper func(thinking *ThinkingConfig, model string) map[string]any

	// ResponseFormatMapper converts ResponseFormat to the provider-specific
	// value for the "response_format" key.  Return nil to omit.
	// If the function itself is nil, defaults to json_object-only support.
	ResponseFormatMapper func(rf *ResponseFormat) any

	// CustomMessageConverter replaces the default ConvertMessages.
	// Must return a JSON-serializable slice.
	CustomMessageConverter func(messages []Message) any

	// CustomToolConverter replaces the default ConvertTools.
	CustomToolConverter func(tools []Tool) any

	// ExtraTransform lets a provider consume or rewrite keys from req.Extra
	// before they are merged verbatim into the request body. Useful for
	// translating provider-agnostic options (e.g. cache_retention) into
	// provider-native fields (e.g. OpenRouter's top-level cache_control).
	//
	// It receives the full Extra map and the in-progress body, and returns
	// the residual Extra (with handled keys removed). The body may be
	// mutated in place. Returning nil is equivalent to consuming everything.
	// When ExtraTransform itself is nil, Extra is merged verbatim.
	ExtraTransform func(extra map[string]any, body map[string]any, req *Request) map[string]any

	// CleanSchema recursively cleans JSON schemas for strict mode.
	CleanSchema func(schema any) any

	// ReasoningField is the preferred JSON key for reasoning in message/delta.
	// When set, it is probed first; remaining default fields are tried as fallback.
	// Default probe order: "reasoning_content", "reasoning", "reasoning_text".
	ReasoningField string

	// ReasoningCondition controls when reasoning is extracted.
	//   ""/"always"                    — whenever the field is non-empty
	//   "model_contains:<substring>"   — only when model name contains substring
	ReasoningCondition string

	// ReasoningCumulative marks streaming reasoning fields that contain the
	// complete accumulated reasoning text on each chunk rather than a delta.
	// The stream reader emits only the newly added suffix.
	ReasoningCumulative bool

	// ContentAsInterface parses Message.Content as interface{} (string or array)
	// instead of plain string.  Used by OpenRouter.
	ContentAsInterface bool

	// ModelFromResponse takes the model name from response JSON.
	// When false, uses the request model.
	ModelFromResponse bool

	// HasCompletionTokenDetails looks for completion_tokens_details.reasoning_tokens.
	HasCompletionTokenDetails bool

	// HasCacheTokens looks for prompt_cache_hit_tokens / prompt_cache_miss_tokens.
	HasCacheTokens bool

	// DataPrefix is the SSE data line prefix.  Default: "data: ".
	DataPrefix string

	// SupportsStrictTools forwards tools[i].function.strict when true.
	// Leave false for providers that reject unknown OpenAI-only fields.
	SupportsStrictTools bool

	// RequireAllToolsStrict marks providers whose strict mode is request-wide.
	// Strict is forwarded only when every function tool opts in; otherwise it is
	// omitted for portability.
	RequireAllToolsStrict bool

	// SkipAPIKeyValidation skips the API key check.  Used by local
	// providers like Ollama that don't require authentication.
	SkipAPIKeyValidation bool
}

type ErrorType string

type FunctionCall struct {
	Name      string `json:"name"`
	Arguments string `json:"arguments"`
}

type FunctionDef struct {
	Name        string `json:"name"`
	Description string `json:"description"`
	Parameters  any    `json:"parameters"`
	// Strict enables provider-native strict tool calling (Structured Outputs
	// for function arguments). OpenAI-compatible Chat APIs use
	// `tools[i].function.strict`; OpenAI Responses uses its flat function tool
	// `strict` field; Anthropic and Bedrock use their native tool strict fields.
	//
	// Caller responsibilities for strict mode (per OpenAI spec):
	//   - every property must be listed in `required`; use a `["string","null"]`
	//     union to express optional fields
	//   - <= 5000 total object properties, <= 10 nesting levels
	//   - keywords like format/pattern/minLength/maximum are best-effort only
	//
	// nil leaves the provider default. Providers without documented strict tool
	// support omit the flag for cross-model portability.
	// See https://platform.openai.com/docs/guides/function-calling
	Strict *bool `json:"strict,omitempty"`
}

type GeminiProvider struct {
	*BaseProvider
}

type HTTPDoer interface {
	Do(req *http.Request) (*http.Response, error)
}

type JSONSchema struct {
	Name        string `json:"name"`
	Description string `json:"description,omitempty"`
	Schema      any    `json:"schema"`
	Strict      *bool  `json:"strict,omitempty"`
}

type LiteLLMError struct {
	Type     ErrorType `json:"type"`
	Code     string    `json:"code,omitempty"`
	Message  string    `json:"message"`
	Provider string    `json:"provider,omitempty"`
	Model    string    `json:"model,omitempty"`
	Cause    error     `json:"-"` // Original error, not serialized.

	// HTTP details (if applicable).
	StatusCode int               `json:"status_code,omitempty"`
	Headers    map[string]string `json:"headers,omitempty"`

	// Retry hints.
	Retryable  bool `json:"retryable"`
	RetryAfter int  `json:"retry_after,omitempty"` // seconds
}

type Message struct {
	Role             string           `json:"role"`
	Content          string           `json:"content"`
	ReasoningContent string           `json:"reasoning_content,omitempty"`
	ReasoningDetails []map[string]any `json:"reasoning_details,omitempty"`
	Contents         []MessageContent `json:"contents,omitempty"`
	ToolCalls        []ToolCall       `json:"tool_calls,omitempty"`
	ToolCallID       string           `json:"tool_call_id,omitempty"`
	IsError          bool             `json:"is_error,omitempty"` // tool result error flag (Anthropic)
	CacheControl     *CacheControl    `json:"cache_control,omitempty"`
}

type MessageContent struct {
	Type        string           `json:"type"`
	Text        string           `json:"text,omitempty"`
	ImageURL    *MessageImageURL `json:"image_url,omitempty"`
	ToolName    string           `json:"tool_name,omitempty"` // tool_reference blocks
	Annotations []map[string]any `json:"annotations,omitempty"`
	Logprobs    []map[string]any `json:"logprobs,omitempty"`
}

type MessageImageURL struct {
	URL    string `json:"url"`
	Detail string `json:"detail,omitempty"` // "auto", "low", or "high"
}

type ModelInfo struct {
	ID               string `json:"id"`
	Name             string `json:"name,omitempty"`
	Provider         string `json:"provider,omitempty"`
	Description      string `json:"description,omitempty"`
	ContextLength    int    `json:"context_length,omitempty"`
	InputTokenLimit  int    `json:"input_token_limit,omitempty"`
	OutputTokenLimit int    `json:"output_token_limit,omitempty"`
	Created          int64  `json:"created,omitempty"`
	OwnedBy          string `json:"owned_by,omitempty"`

	// Capability flags (best-effort, may be zero-valued for unknown models)
	SupportsTools    bool `json:"supports_tools,omitempty"`
	SupportsVision   bool `json:"supports_vision,omitempty"`
	SupportsThinking bool `json:"supports_thinking,omitempty"`
}

type ModelLister interface {
	ListModels(ctx context.Context) ([]ModelInfo, error)
}

type OpenAICompatMessage struct {
	Role             string           `json:"role"`
	Content          interface{}      `json:"content,omitempty"` // string or array (needed by OpenRouter)
	ToolCalls        []openaiToolCall `json:"tool_calls,omitempty"`
	ToolCallID       string           `json:"tool_call_id,omitempty"`
	ReasoningContent string           `json:"reasoning_content,omitempty"` // used by GLM/Qwen
	ReasoningDetails []map[string]any `json:"reasoning_details,omitempty"`
}

type OpenAICompatProvider struct {
	*BaseProvider
	// contains filtered or unexported fields
}

type OpenAIProvider struct {
	*BaseProvider
}

type OpenAIResponsesRequest struct {
	Model string `json:"model"`

	// Use standard messages to build input items for Responses API.
	Messages []Message `json:"messages"`

	// Top-level instructions. System/developer messages are also extracted and
	// combined into this field when building the official Responses payload.
	Instructions string `json:"instructions,omitempty"`

	// Conversation state. Conversation and PreviousResponseID are mutually
	// exclusive per the official Responses API.
	Conversation       any    `json:"conversation,omitempty"`
	PreviousResponseID string `json:"previous_response_id,omitempty"`

	// Output configuration
	MaxOutputTokens *int     `json:"max_output_tokens,omitempty"`
	MaxToolCalls    *int     `json:"max_tool_calls,omitempty"`
	Include         []string `json:"include,omitempty"`
	TopLogprobs     *int     `json:"top_logprobs,omitempty"`

	// Sampling parameters
	Temperature *float64 `json:"temperature,omitempty"`
	TopP        *float64 `json:"top_p,omitempty"`

	// Text configuration. ResponseFormat maps to text.format; TextVerbosity
	// maps to text.verbosity ("low", "medium", or "high").
	ResponseFormat *ResponseFormat `json:"response_format,omitempty"`
	TextVerbosity  string          `json:"text_verbosity,omitempty"`
	Truncation     string          `json:"truncation,omitempty"` // "auto" or "disabled"

	// Tool configuration
	Tools             []Tool                `json:"tools,omitempty"`
	OpenAITools       []OpenAIResponsesTool `json:"openai_tools,omitempty"`
	ToolChoice        any                   `json:"tool_choice,omitempty"`
	ParallelToolCalls *bool                 `json:"parallel_tool_calls,omitempty"`

	// Reasoning configuration for Responses reasoning models.
	ReasoningEffort  string `json:"reasoning_effort,omitempty"`  // none, low, medium, high, xhigh
	ReasoningSummary string `json:"reasoning_summary,omitempty"` // auto, concise, detailed

	// Unified thinking control (optional). If set, it maps to reasoning fields.
	Thinking *ThinkingConfig `json:"thinking,omitempty"`

	// Prompt caching and request metadata.
	PromptCacheKey       string            `json:"prompt_cache_key,omitempty"`
	PromptCacheRetention string            `json:"prompt_cache_retention,omitempty"`
	Metadata             map[string]string `json:"metadata,omitempty"`
	SafetyIdentifier     string            `json:"safety_identifier,omitempty"`

	// Service configuration.
	ServiceTier string `json:"service_tier,omitempty"` // auto, default, flex, priority
	Background  *bool  `json:"background,omitempty"`
	Store       *bool  `json:"store,omitempty"`

	// Prompt template reference, passed through as the official prompt object.
	Prompt map[string]any `json:"prompt,omitempty"`

	// APIKey overrides the provider-level API key for this single request.
	APIKey string `json:"-"`

	// OnPayload is called with the serialized JSON body before sending
	// the HTTP request. Useful for debugging and logging API calls.
	OnPayload func(provider string, payload []byte) `json:"-"`
}

type OpenAIResponsesTool map[string]any

type Provider interface {
	Name() string
	Validate() error
	Chat(ctx context.Context, req *Request) (*Response, error)
	Stream(ctx context.Context, req *Request) (StreamReader, error)
}

type ProviderConfig struct {
	APIKey     string           `json:"api_key"`
	BaseURL    string           `json:"base_url,omitempty"`
	Timeout    time.Duration    `json:"timeout,omitempty"`
	Extra      map[string]any   `json:"extra,omitempty"`
	Resilience ResilienceConfig `json:"resilience,omitempty"`
	HTTPClient HTTPDoer         `json:"-"`
}

type Request struct {
	Model          string          `json:"model"`
	Messages       []Message       `json:"messages"`
	MaxTokens      *int            `json:"max_tokens,omitempty"`
	Temperature    *float64        `json:"temperature,omitempty"`
	TopP           *float64        `json:"top_p,omitempty"`
	Tools          []Tool          `json:"tools,omitempty"`
	ToolChoice     any             `json:"tool_choice,omitempty"`
	ResponseFormat *ResponseFormat `json:"response_format,omitempty"`
	Stop           []string        `json:"stop,omitempty"`
	Thinking       *ThinkingConfig `json:"thinking,omitempty"`

	// APIKey overrides the provider-level API key for this single request.
	// When empty, the provider's default key is used.
	// Enables key rotation, OAuth short-lived tokens, and multi-tenant scenarios.
	APIKey string `json:"-"`

	// Provider-specific extensions
	Extra map[string]any `json:"extra,omitempty"`

	// OnPayload is called with the serialized JSON body before sending
	// the HTTP request. Useful for debugging and logging API calls.
	OnPayload func(provider string, payload []byte) `json:"-"`

	// OnWarning is called when a provider adapter keeps the request portable by
	// omitting or reducing an unsupported option instead of failing the call.
	OnWarning func(provider string, message string) `json:"-"`
}

type ResilienceConfig struct {
	MaxRetries     int           `json:"max_retries"`
	InitialDelay   time.Duration `json:"initial_delay"`
	MaxDelay       time.Duration `json:"max_delay"`
	Multiplier     float64       `json:"multiplier"`
	Jitter         bool          `json:"jitter"`
	RequestTimeout time.Duration `json:"request_timeout"`
	ConnectTimeout time.Duration `json:"connect_timeout"`
	// StreamIdleTimeout aborts a streaming response if no chunk arrives
	// within this window. Guards against silent connection stalls that
	// RequestTimeout cannot detect (server keeps the TCP connection open
	// but stops emitting tokens). 0 disables the watchdog.
	StreamIdleTimeout time.Duration `json:"stream_idle_timeout"`
}

type Response struct {
	Content          string           `json:"content"`
	Contents         []MessageContent `json:"contents,omitempty"`
	ToolCalls        []ToolCall       `json:"tool_calls,omitempty"`
	Usage            Usage            `json:"usage"`
	Model            string           `json:"model"`
	Provider         string           `json:"provider"`
	FinishReason     string           `json:"finish_reason,omitempty"`
	ReasoningContent string           `json:"reasoning_content,omitempty"`
	ReasoningDetails []map[string]any `json:"reasoning_details,omitempty"`

	// Extra captures provider-specific response fields not covered by the
	// standard schema (e.g. logprobs, annotations, system_fingerprint).
	Extra map[string]any `json:"extra,omitempty"`
}

type ResponseFormat struct {
	Type       string      `json:"type"`
	JSONSchema *JSONSchema `json:"json_schema,omitempty"`
}

type StreamChunk struct {
	Type             string         `json:"type"`
	Content          string         `json:"content,omitempty"`
	ContentIndex     *int           `json:"content_index,omitempty"`
	OutputIndex      *int           `json:"output_index,omitempty"`
	ItemID           string         `json:"item_id,omitempty"`
	ToolCallDelta    *ToolCallDelta `json:"tool_call_delta,omitempty"`
	FinishReason     string         `json:"finish_reason,omitempty"`
	Model            string         `json:"model,omitempty"`
	Provider         string         `json:"provider"`
	Done             bool           `json:"done"`
	ReasoningContent string         `json:"reasoning_content,omitempty"`
	ReasoningDone    bool           `json:"reasoning_done,omitempty"`
	Usage            *Usage         `json:"usage,omitempty"`
}

type StreamReader interface {
	Next() (*StreamChunk, error)
	Close() error
}

type ThinkingConfig struct {
	Type         string `json:"type"`            // "enabled" or "disabled"
	Level        string `json:"level,omitempty"` // "low", "medium", "high" — provider translates to API-specific param
	BudgetTokens *int   `json:"budget_tokens,omitempty"`
}

type Tool struct {
	Type         string      `json:"type"`
	Function     FunctionDef `json:"function"`
	DeferLoading bool        `json:"defer_loading,omitempty"`
}

type ToolCall struct {
	ID       string       `json:"id"`
	Type     string       `json:"type"`
	Function FunctionCall `json:"function"`
	// ThoughtSignature carries Gemini 3's opaque reasoning signature, passed
	// through verbatim across turns. Empty for providers that don't emit one.
	ThoughtSignature string `json:"thought_signature,omitempty"`
}

type ToolCallDelta struct {
	Index            int    `json:"index"`
	ID               string `json:"id,omitempty"`
	Type             string `json:"type,omitempty"`
	FunctionName     string `json:"function_name,omitempty"`
	ArgumentsDelta   string `json:"arguments_delta,omitempty"`
	ThoughtSignature string `json:"thought_signature,omitempty"`
	OutputIndex      *int   `json:"output_index,omitempty"`
	ItemID           string `json:"item_id,omitempty"`
}

type Usage struct {
	Provider string `json:"provider,omitempty"`
	Model    string `json:"model,omitempty"`

	PromptTokens     int `json:"prompt_tokens"`
	CompletionTokens int `json:"completion_tokens"`
	TotalTokens      int `json:"total_tokens"`
	ReasoningTokens  int `json:"reasoning_tokens,omitempty"`

	// Cache-related token statistics
	CacheCreationInputTokens int `json:"cache_creation_input_tokens,omitempty"` // Tokens written to cache
	CacheReadInputTokens     int `json:"cache_read_input_tokens,omitempty"`     // Tokens read from cache
}