import {
  type ChatCompletionMessageParam,
  type ChatCompletionMessageToolCall,
  type ChatCompletionTool,
} from "openai/resources/index.js";

/* -------------------------------------------------------------------------- */
/*                              Wire message types                            */
/* -------------------------------------------------------------------------- */

/** Assistant message extended with provider-specific extras (e.g. reasoning_content). */
export type OpenAICompatAssistantMessage = {
  role: "assistant";
  content: string;
  tool_calls?: ChatCompletionMessageToolCall[];
  [key: string]: unknown;
};

/** Union of wire message shapes accepted by the upstream completions endpoint. */
export type OpenAICompatMessageParam = ChatCompletionMessageParam | OpenAICompatAssistantMessage;

/* -------------------------------------------------------------------------- */
/*                          Outbound request body                             */
/*                                                                            */
/* Precise model of the OpenAI-COMPATIBLE SUPERSET we transmit: the base       */
/* completions fields plus provider-specific extensions (top_k, min_p,         */
/* reasoning_effort, preserve_thinking, priority, …). `stream: true` is a      */
/* literal so the SDK resolves its streaming overload. The trailing index      */
/* signature admits future vendor keys without reshaping this type.            */
/* -------------------------------------------------------------------------- */

export type ShartiCompletionsBody = {
  model: string;
  messages: OpenAICompatMessageParam[];
  stream: true;
  stream_options: { include_usage: true };

  // Base sampling / tooling fields
  temperature?: number;
  top_p?: number;
  max_tokens?: number;
  seed?: number;
  presence_penalty?: number;
  frequency_penalty?: number;
  tools?: ChatCompletionTool[];
  tool_choice?: "auto" | "required" | "none";
  parallel_tool_calls?: boolean;

  // Provider-specific extensions (OpenAI-compatible superset)
  top_k?: number;
  min_p?: number;
  repetition_penalty?: number;
  reasoning_effort?: string;
  thinking_token_budget?: number;
  preserve_thinking?: boolean;
  priority?: string;
  timeout?: number;

  [key: string]: unknown;
};

/* -------------------------------------------------------------------------- */
/*                             Upstream stream chunk                          */
/*                                                                            */
/* Deliberate structural SUPERSET of the OpenAI chunk plus LM Studio / Splash  */
/* extras (native reasoning, timing telemetry). Defined WITHOUT index          */
/* signatures so the SDK's concrete ChatCompletionChunk remains assignable     */
/* to it — the consumer therefore needs no cast at the boundary.               */
/* -------------------------------------------------------------------------- */

export type UpstreamUsage = {
  prompt_tokens?: number;
  completion_tokens?: number;
  total_tokens?: number;
};

export type UpstreamTiming = Record<string, unknown>;

export type UpstreamToolCallDelta = {
  index: number;
  id?: string;
  function?: { name?: string; arguments?: string };
};

export type UpstreamDelta = {
  content?: string | null;
  tool_calls?: UpstreamToolCallDelta[];
  /** Native reasoning channel (provider-specific). */
  reasoning_content?: unknown;
};

export type UpstreamChoice = {
  delta?: UpstreamDelta;
  finish_reason?: string | null;
};

export type UpstreamChunk = {
  choices?: UpstreamChoice[];
  usage?: UpstreamUsage | null;
  timings?: UpstreamTiming;
};

/* -------------------------------------------------------------------------- */
/*                              Streaming state                               */
/* -------------------------------------------------------------------------- */

/** Mutable state for a single in-flight tool call while assembling its arguments. */
export type ToolCallState = {
  id: string;
  name: string | null;
  index: number;
  arguments: string;
};

/* -------------------------------------------------------------------------- */
/*                          Aggregated statistics                             */
/* -------------------------------------------------------------------------- */

/** Metrics gathered from a single completed request stream. */
export type GenerationStats = {
  inputTokens: number | null;
  outputTokens: number | null;
  totalTokens: number | null;
  /** Decode throughput (tok/s): server-reported when available, else derived. */
  tokensPerSecond: number | null;
  timeToFirstTokenSeconds: number | null;
  totalTimeSeconds: number | null;
  promptPerSecond: number | null;
  cachedTokens: number | null;
  rawTimings: UpstreamTiming | null;
  finishReason: string | null;
};

import {
  type ChatCompletionMessageParam,
  type ChatCompletionMessageToolCall,
  type ChatCompletionTool,
} from "openai/resources/index.js";

/* -------------------------------------------------------------------------- */
/*                              Wire message types                            */
/* -------------------------------------------------------------------------- */

/** Assistant message extended with provider-specific extras (e.g. reasoning_content). */
export type OpenAICompatAssistantMessage = {
  role: "assistant";
  content: string;
  tool_calls?: ChatCompletionMessageToolCall[];
  [key: string]: unknown;
};

/** Union of wire message shapes accepted by the upstream completions endpoint. */
export type OpenAICompatMessageParam = ChatCompletionMessageParam | OpenAICompatAssistantMessage;

/* -------------------------------------------------------------------------- */
/*                          Outbound request body                             */
/*                                                                            */
/* Precise model of the OpenAI-COMPATIBLE SUPERSET we transmit: the base       */
/* completions fields plus provider-specific extensions (top_k, min_p,         */
/* reasoning_effort, preserve_thinking, priority, …). `stream: true` is a      */
/* literal so the SDK resolves its streaming overload. The trailing index      */
/* signature admits future vendor keys without reshaping this type.            */
/* -------------------------------------------------------------------------- */

export type ShartiCompletionsBody = {
  model: string;
  messages: OpenAICompatMessageParam[];
  stream: true;
  stream_options: { include_usage: true };

  // Base sampling / tooling fields
  temperature?: number;
  top_p?: number;
  max_tokens?: number;
  seed?: number;
  presence_penalty?: number;
  frequency_penalty?: number;
  tools?: ChatCompletionTool[];
  tool_choice?: "auto" | "required" | "none";
  parallel_tool_calls?: boolean;

  // Provider-specific extensions (OpenAI-compatible superset)
  top_k?: number;
  min_p?: number;
  repetition_penalty?: number;
  reasoning_effort?: string;
  thinking_token_budget?: number;
  preserve_thinking?: boolean;
  priority?: string;
  timeout?: number;

  [key: string]: unknown;
};

/* -------------------------------------------------------------------------- */
/*                             Upstream stream chunk                          */
/*                                                                            */
/* Deliberate structural SUPERSET of the OpenAI chunk plus LM Studio / Splash  */
/* extras (native reasoning, timing telemetry). Defined WITHOUT index          */
/* signatures so the SDK's concrete ChatCompletionChunk remains assignable     */
/* to it — the consumer therefore needs no cast at the boundary.               */
/* -------------------------------------------------------------------------- */

export type UpstreamUsage = {
  prompt_tokens?: number;
  completion_tokens?: number;
  total_tokens?: number;
};

export type UpstreamTiming = Record<string, unknown>;

export type UpstreamToolCallDelta = {
  index: number;
  id?: string;
  function?: { name?: string; arguments?: string };
};

export type UpstreamDelta = {
  content?: string | null;
  tool_calls?: UpstreamToolCallDelta[];
  /** Native reasoning channel (provider-specific). */
  reasoning_content?: unknown;
};

export type UpstreamChoice = {
  delta?: UpstreamDelta;
  finish_reason?: string | null;
};

export type UpstreamChunk = {
  choices?: UpstreamChoice[];
  usage?: UpstreamUsage | null;
  timings?: UpstreamTiming;
};

/* -------------------------------------------------------------------------- */
/*                              Streaming state                               */
/* -------------------------------------------------------------------------- */

/** Mutable state for a single in-flight tool call while assembling its arguments. */
export type ToolCallState = {
  id: string;
  name: string | null;
  index: number;
  arguments: string;
};

/* -------------------------------------------------------------------------- */
/*                          Aggregated statistics                             */
/* -------------------------------------------------------------------------- */

/** Metrics gathered from a single completed request stream. */
export type GenerationStats = {
  inputTokens: number | null;
  outputTokens: number | null;
  totalTokens: number | null;
  /** Decode throughput (tok/s): server-reported when available, else derived. */
  tokensPerSecond: number | null;
  timeToFirstTokenSeconds: number | null;
  totalTimeSeconds: number | null;
  promptPerSecond: number | null;
  cachedTokens: number | null;
  rawTimings: UpstreamTiming | null;
  finishReason: string | null;
};