import OpenAI from "openai/index.js";
import type { ChatCompletionCreateParamsStreaming, ChatCompletionTool } from "openai/resources/index.js";
import { CONFIG_DEFAULTS, type ShartiConfig } from "./config";
import type { OpenAICompatMessageParam, ShartiCompletionsBody } from "./types";
import { asBool, asNumber, asString } from "./utils/coerce";
import { formatBlock } from "./utils/block";
import { ShartiLogger } from "./utils/logger";

/* -------------------------------------------------------------------------- */
/*                              Client factory                                */
/* -------------------------------------------------------------------------- */

/** Construct an OpenAI-compatible client for the configured endpoint. */
export function createOpenAI(baseUrl: string, apiKey: string) {
  return new OpenAI({
    apiKey: apiKey || "no-key",
    baseURL: baseUrl || undefined,
  });
}

/**
 * Issue a streaming chat-completions request.
 *
 * Centralizes the single, localized escape hatch at the OpenAI-compatible boundary: the
 * (superset) body is asserted against the SDK's streaming-params contract. Because the body
 * carries `stream: true`, the SDK resolves its streaming overload, so the awaited result is a
 * proper `Stream<ChatCompletionChunk>` — no downstream re-casting required.
 */
export async function executeCompletions(
  client: OpenAI,
  body: ShartiCompletionsBody,
  signal: AbortSignal,
) {
  return client.chat.completions.create(body as ChatCompletionCreateParamsStreaming, { signal });
}

/* -------------------------------------------------------------------------- */
/*                             Payload assembly                               */
/*                                                                            */
/* Structural parameters are always sent. Tuning parameters are gated by the  */
/* `customSampling` toggle so an untouched profile sends a minimal request.   */
/* Coercion fallbacks draw from CONFIG_DEFAULTS so a default is declared once.*/
/* -------------------------------------------------------------------------- */

export function buildRequestPayload(
  config: ShartiConfig,
  messages: OpenAICompatMessageParam[],
  tools: ChatCompletionTool[] | undefined,
): ShartiCompletionsBody {
  const payload: ShartiCompletionsBody = {
    model: asString(config.get("model")),
    messages,
    stream: true,
    // Ensure the final chunk carries token usage stats.
    stream_options: { include_usage: true },
  };

  // Tools: structural part of the request — attached whenever available.
  if (tools && tools.length > 0) {
    payload.tools = tools;
    payload.tool_choice = asString(config.get("toolChoice"), CONFIG_DEFAULTS.toolChoice) as ShartiCompletionsBody["tool_choice"];
    payload.parallel_tool_calls = asBool(config.get("parallelToolCalls"), CONFIG_DEFAULTS.parallelToolCalls);
  }

  // Behavioural flags — always sent regardless of customSampling.
  payload.preserve_thinking = true;
  payload.priority = asString(config.get("priority"), CONFIG_DEFAULTS.priority);
  const timeout = asNumber(config.get("timeout"), CONFIG_DEFAULTS.timeout);
  if (timeout > 0) {
    payload.timeout = timeout;
  }

  // Tuning parameters — omitted wholesale when customSampling is OFF.
  if (!asBool(config.get("customSampling"), CONFIG_DEFAULTS.customSampling)) {
    return payload;
  }

  payload.temperature = asNumber(config.get("temperature"), CONFIG_DEFAULTS.temperature);
  payload.top_k = asNumber(config.get("topK"), CONFIG_DEFAULTS.topK);
  payload.top_p = asNumber(config.get("topP"), CONFIG_DEFAULTS.topP);
  payload.min_p = asNumber(config.get("minP"), CONFIG_DEFAULTS.minP);
  payload.presence_penalty = asNumber(config.get("presencePenalty"), CONFIG_DEFAULTS.presencePenalty);
  payload.frequency_penalty = asNumber(config.get("frequencyPenalty"), CONFIG_DEFAULTS.frequencyPenalty);
  payload.repetition_penalty = asNumber(config.get("repetitionPenalty"), CONFIG_DEFAULTS.repetitionPenalty);
  payload.seed = asNumber(config.get("seed"), CONFIG_DEFAULTS.seed);
  payload.max_tokens = asNumber(config.get("maxTokens"), CONFIG_DEFAULTS.maxTokens);

  const reasoningEffort = asString(config.get("reasoningEffort"), CONFIG_DEFAULTS.reasoningEffort);
  if (reasoningEffort !== "default" && reasoningEffort.length > 0) {
    payload.reasoning_effort = reasoningEffort;
  }

  const thinkingTokenBudget = asNumber(config.get("thinkingTokenBudget"), CONFIG_DEFAULTS.thinkingTokenBudget);
  if (thinkingTokenBudget >= 0) {
    payload.thinking_token_budget = thinkingTokenBudget;
  }

  return payload;
}

/* -------------------------------------------------------------------------- */
/*                          Diagnostic payload dump                           */
/*                                                                            */
/* Derives the debug view directly from the constructed payload, so any knob   */
/* added in buildRequestPayload surfaces in diagnostics automatically.        */
/* -------------------------------------------------------------------------- */

/** Infrastructure keys that are not user-facing tuning knobs (hidden from the dump). */
const HIDDEN_DEBUG_KEYS = new Set([
  "model",
  "messages",
  "stream",
  "stream_options",
  "preserve_thinking",
]);

/**
 * Display order for the Settings dump, mirroring the field order in config.ts
 * (Sampling → Reasoning → Tools → Request). `tools` is not a config field; it is
 * slotted at the end of the Tools group.
 */
const SETTINGS_DISPLAY_ORDER: string[] = [
  "temperature",
  "top_k",
  "top_p",
  "min_p",
  "presence_penalty",
  "frequency_penalty",
  "repetition_penalty",
  "seed",
  "max_tokens",
  "reasoning_effort",
  "thinking_token_budget",
  "tools",
  "tool_choice",
  "parallel_tool_calls",
  "priority",
  "timeout",
];

const SETTINGS_ORDER_RANK: Record<string, number> = Object.fromEntries(
  SETTINGS_DISPLAY_ORDER.map((key, index) => [key, index]),
);

/** Optional per-key presentation tweaks for the debug dump. */
const DISPLAY_TRANSFORMERS: Partial<Record<string, (v: unknown) => unknown>> = {
  tools: (v) => (Array.isArray(v) && v.length > 0 ? true : false),
  timeout: (v) => `${v}s`,
};

export function logDebugPayload(logger: ShartiLogger, payload: Record<string, unknown>): void {
  const rows: Array<[string, unknown]> = Object.entries(payload)
    .filter(([k, v]) => !HIDDEN_DEBUG_KEYS.has(k) && v !== undefined && v !== null)
    .sort(
      ([a], [b]) =>
        (SETTINGS_ORDER_RANK[a] ?? Number.MAX_SAFE_INTEGER) -
        (SETTINGS_ORDER_RANK[b] ?? Number.MAX_SAFE_INTEGER),
    )
    .map(([k, v]) => [k, DISPLAY_TRANSFORMERS[k] ? DISPLAY_TRANSFORMERS[k]!(v) : v]);

  const block = formatBlock("⚙️ Settings:", rows);
  if (block) logger.debug(block);
}

import OpenAI from "openai/index.js";
import type { ChatCompletionCreateParamsStreaming, ChatCompletionTool } from "openai/resources/index.js";
import { CONFIG_DEFAULTS, type ShartiConfig } from "./config";
import type { OpenAICompatMessageParam, ShartiCompletionsBody } from "./types";
import { asBool, asNumber, asString } from "./utils/coerce";
import { formatBlock } from "./utils/block";
import { ShartiLogger } from "./utils/logger";

/* -------------------------------------------------------------------------- */
/*                              Client factory                                */
/* -------------------------------------------------------------------------- */

/** Construct an OpenAI-compatible client for the configured endpoint. */
export function createOpenAI(baseUrl: string, apiKey: string) {
  return new OpenAI({
    apiKey: apiKey || "no-key",
    baseURL: baseUrl || undefined,
  });
}

/**
 * Issue a streaming chat-completions request.
 *
 * Centralizes the single, localized escape hatch at the OpenAI-compatible boundary: the
 * (superset) body is asserted against the SDK's streaming-params contract. Because the body
 * carries `stream: true`, the SDK resolves its streaming overload, so the awaited result is a
 * proper `Stream<ChatCompletionChunk>` — no downstream re-casting required.
 */
export async function executeCompletions(
  client: OpenAI,
  body: ShartiCompletionsBody,
  signal: AbortSignal,
) {
  return client.chat.completions.create(body as ChatCompletionCreateParamsStreaming, { signal });
}

/* -------------------------------------------------------------------------- */
/*                             Payload assembly                               */
/*                                                                            */
/* Structural parameters are always sent. Tuning parameters are gated by the  */
/* `customSampling` toggle so an untouched profile sends a minimal request.   */
/* Coercion fallbacks draw from CONFIG_DEFAULTS so a default is declared once.*/
/* -------------------------------------------------------------------------- */

export function buildRequestPayload(
  config: ShartiConfig,
  messages: OpenAICompatMessageParam[],
  tools: ChatCompletionTool[] | undefined,
): ShartiCompletionsBody {
  const payload: ShartiCompletionsBody = {
    model: asString(config.get("model")),
    messages,
    stream: true,
    // Ensure the final chunk carries token usage stats.
    stream_options: { include_usage: true },
  };

  // Tools: structural part of the request — attached whenever available.
  if (tools && tools.length > 0) {
    payload.tools = tools;
    payload.tool_choice = asString(config.get("toolChoice"), CONFIG_DEFAULTS.toolChoice) as ShartiCompletionsBody["tool_choice"];
    payload.parallel_tool_calls = asBool(config.get("parallelToolCalls"), CONFIG_DEFAULTS.parallelToolCalls);
  }

  // Behavioural flags — always sent regardless of customSampling.
  payload.preserve_thinking = true;
  payload.priority = asString(config.get("priority"), CONFIG_DEFAULTS.priority);
  const timeout = asNumber(config.get("timeout"), CONFIG_DEFAULTS.timeout);
  if (timeout > 0) {
    payload.timeout = timeout;
  }

  // Tuning parameters — omitted wholesale when customSampling is OFF.
  if (!asBool(config.get("customSampling"), CONFIG_DEFAULTS.customSampling)) {
    return payload;
  }

  payload.temperature = asNumber(config.get("temperature"), CONFIG_DEFAULTS.temperature);
  payload.top_k = asNumber(config.get("topK"), CONFIG_DEFAULTS.topK);
  payload.top_p = asNumber(config.get("topP"), CONFIG_DEFAULTS.topP);
  payload.min_p = asNumber(config.get("minP"), CONFIG_DEFAULTS.minP);
  payload.presence_penalty = asNumber(config.get("presencePenalty"), CONFIG_DEFAULTS.presencePenalty);
  payload.frequency_penalty = asNumber(config.get("frequencyPenalty"), CONFIG_DEFAULTS.frequencyPenalty);
  payload.repetition_penalty = asNumber(config.get("repetitionPenalty"), CONFIG_DEFAULTS.repetitionPenalty);
  payload.seed = asNumber(config.get("seed"), CONFIG_DEFAULTS.seed);
  payload.max_tokens = asNumber(config.get("maxTokens"), CONFIG_DEFAULTS.maxTokens);

  const reasoningEffort = asString(config.get("reasoningEffort"), CONFIG_DEFAULTS.reasoningEffort);
  if (reasoningEffort !== "default" && reasoningEffort.length > 0) {
    payload.reasoning_effort = reasoningEffort;
  }

  const thinkingTokenBudget = asNumber(config.get("thinkingTokenBudget"), CONFIG_DEFAULTS.thinkingTokenBudget);
  if (thinkingTokenBudget >= 0) {
    payload.thinking_token_budget = thinkingTokenBudget;
  }

  return payload;
}

/* -------------------------------------------------------------------------- */
/*                          Diagnostic payload dump                           */
/*                                                                            */
/* Derives the debug view directly from the constructed payload, so any knob   */
/* added in buildRequestPayload surfaces in diagnostics automatically.        */
/* -------------------------------------------------------------------------- */

/** Infrastructure keys that are not user-facing tuning knobs (hidden from the dump). */
const HIDDEN_DEBUG_KEYS = new Set([
  "model",
  "messages",
  "stream",
  "stream_options",
  "preserve_thinking",
]);

/**
 * Display order for the Settings dump, mirroring the field order in config.ts
 * (Sampling → Reasoning → Tools → Request). `tools` is not a config field; it is
 * slotted at the end of the Tools group.
 */
const SETTINGS_DISPLAY_ORDER: string[] = [
  "temperature",
  "top_k",
  "top_p",
  "min_p",
  "presence_penalty",
  "frequency_penalty",
  "repetition_penalty",
  "seed",
  "max_tokens",
  "reasoning_effort",
  "thinking_token_budget",
  "tools",
  "tool_choice",
  "parallel_tool_calls",
  "priority",
  "timeout",
];

const SETTINGS_ORDER_RANK: Record<string, number> = Object.fromEntries(
  SETTINGS_DISPLAY_ORDER.map((key, index) => [key, index]),
);

/** Optional per-key presentation tweaks for the debug dump. */
const DISPLAY_TRANSFORMERS: Partial<Record<string, (v: unknown) => unknown>> = {
  tools: (v) => (Array.isArray(v) && v.length > 0 ? true : false),
  timeout: (v) => `${v}s`,
};

export function logDebugPayload(logger: ShartiLogger, payload: Record<string, unknown>): void {
  const rows: Array<[string, unknown]> = Object.entries(payload)
    .filter(([k, v]) => !HIDDEN_DEBUG_KEYS.has(k) && v !== undefined && v !== null)
    .sort(
      ([a], [b]) =>
        (SETTINGS_ORDER_RANK[a] ?? Number.MAX_SAFE_INTEGER) -
        (SETTINGS_ORDER_RANK[b] ?? Number.MAX_SAFE_INTEGER),
    )
    .map(([k, v]) => [k, DISPLAY_TRANSFORMERS[k] ? DISPLAY_TRANSFORMERS[k]!(v) : v]);

  const block = formatBlock("⚙️ Settings:", rows);
  if (block) logger.debug(block);
}