MOPC-Portal/src/lib/openai.ts

import OpenAI from 'openai'
import type { ChatCompletionCreateParamsNonStreaming } from 'openai/resources/chat/completions'
import Anthropic from '@anthropic-ai/sdk'
import { prisma } from './prisma'

// Hardcoded Claude model list (Anthropic API doesn't expose a models.list endpoint for all users)
export const ANTHROPIC_CLAUDE_MODELS = [
  'claude-opus-4-5-20250514',
  'claude-sonnet-4-5-20250514',
  'claude-haiku-3-5-20241022',
  'claude-opus-4-20250514',
  'claude-sonnet-4-20250514',
] as const

/**
 * AI client type returned by getOpenAI().
 * Both the OpenAI SDK and the Anthropic adapter satisfy this interface.
 * All AI services only use .chat.completions.create(), so this is safe.
 */
export type AIClient = OpenAI | AnthropicClientAdapter

type AnthropicClientAdapter = {
  __isAnthropicAdapter: true
  chat: {
    completions: {
      create(params: ChatCompletionCreateParamsNonStreaming): Promise<OpenAI.Chat.Completions.ChatCompletion>
    }
  }
}

// OpenAI client singleton with lazy initialization
const globalForOpenAI = globalThis as unknown as {
  openai: AIClient | undefined
  openaiInitialized: boolean
}

// ─── Provider Detection ─────────────────────────────────────────────────────

/**
 * Get the configured AI provider from SystemSettings.
 * Returns 'openai' (default), 'litellm' (ChatGPT subscription proxy), or 'anthropic' (Claude API).
 */
export async function getConfiguredProvider(): Promise<'openai' | 'litellm' | 'anthropic'> {
  try {
    const setting = await prisma.systemSettings.findUnique({
      where: { key: 'ai_provider' },
    })
    const value = setting?.value || 'openai'
    if (value === 'litellm') return 'litellm'
    if (value === 'anthropic') return 'anthropic'
    return 'openai'
  } catch {
    return 'openai'
  }
}

/**
 * Check if a model ID indicates LiteLLM ChatGPT subscription routing.
 * Models like 'chatgpt/gpt-5.2' use the chatgpt/ prefix.
 * Used by buildCompletionParams (sync) to strip unsupported token limit fields.
 */
export function isLiteLLMChatGPTModel(model: string): boolean {
  return model.toLowerCase().startsWith('chatgpt/')
}

// ─── Model Type Detection ────────────────────────────────────────────────────

/**
 * Reasoning models that require different API parameters:
 * - Use max_completion_tokens instead of max_tokens
 * - Don't support response_format: json_object (must instruct JSON in prompt)
 * - Don't support temperature parameter
 * - Don't support system messages (use developer or user role instead)
 */
const REASONING_MODEL_PREFIXES = ['o1', 'o3', 'o4']

/**
 * Models that use max_completion_tokens instead of max_tokens.
 * This includes reasoning models AND newer GPT models (GPT-5+).
 */
const NEW_TOKEN_PARAM_PREFIXES = ['o1', 'o3', 'o4', 'gpt-5', 'gpt-6', 'gpt-7']

/**
 * Models that don't support custom temperature values.
 * These only accept the default temperature (1).
 */
const NO_TEMPERATURE_PREFIXES = ['o1', 'o3', 'o4', 'gpt-5', 'gpt-6', 'gpt-7']

/**
 * Check if a model is a reasoning model (o1, o3, o4 series)
 * These models have additional restrictions (no temperature, no json_object, etc.)
 */
export function isReasoningModel(model: string): boolean {
  const modelLower = model.toLowerCase()
  return REASONING_MODEL_PREFIXES.some(prefix =>
    modelLower.startsWith(prefix) ||
    modelLower.includes(`/${prefix}`) ||
    modelLower.includes(`-${prefix}`)
  )
}

/**
 * Check if a model requires max_completion_tokens instead of max_tokens.
 * This includes reasoning models AND newer GPT models (GPT-5+).
 */
export function usesNewTokenParam(model: string): boolean {
  const modelLower = model.toLowerCase()
  return NEW_TOKEN_PARAM_PREFIXES.some(prefix =>
    modelLower.startsWith(prefix) ||
    modelLower.includes(`/${prefix}`) ||
    modelLower.includes(`-${prefix}`)
  )
}

/**
 * Check if a model supports custom temperature values.
 * Newer models (o-series, GPT-5+) only accept default temperature (1).
 */
export function supportsTemperature(model: string): boolean {
  const modelLower = model.toLowerCase()
  return !NO_TEMPERATURE_PREFIXES.some(prefix =>
    modelLower.startsWith(prefix) ||
    modelLower.includes(`/${prefix}`) ||
    modelLower.includes(`-${prefix}`)
  )
}

/**
 * Check if a model requires higher token limits due to reasoning overhead.
 * GPT-5 nano especially needs more tokens as reasoning consumes output budget.
 */
export function needsHigherTokenLimit(model: string): boolean {
  const modelLower = model.toLowerCase()
  return modelLower.includes('nano') || modelLower.includes('gpt-5')
}

/**
 * Get minimum recommended max_tokens for a model.
 * Reasoning models need higher limits because internal reasoning consumes tokens.
 */
export function getMinTokenLimit(model: string, requestedLimit?: number): number | undefined {
  // For GPT-5 nano, reasoning uses significant token budget
  // If user requests < 8000, bump it up or remove limit
  if (needsHigherTokenLimit(model)) {
    const minLimit = 16000 // Ensure enough headroom for reasoning
    if (!requestedLimit) return undefined // No limit = model default
    return Math.max(requestedLimit, minLimit)
  }
  return requestedLimit
}

// ─── Chat Completion Parameter Builder ───────────────────────────────────────

type MessageRole = 'system' | 'user' | 'assistant' | 'developer'

export interface ChatCompletionOptions {
  messages: Array<{ role: MessageRole; content: string }>
  maxTokens?: number
  temperature?: number
  jsonMode?: boolean
}

/**
 * Build chat completion parameters with correct settings for the model type.
 * Handles differences between standard models and reasoning models.
 */
export function buildCompletionParams(
  model: string,
  options: ChatCompletionOptions
): ChatCompletionCreateParamsNonStreaming {
  const isReasoning = isReasoningModel(model)

  // Convert messages for reasoning models (system -> developer)
  const messages = options.messages.map(msg => {
    if (isReasoning && msg.role === 'system') {
      return { role: 'developer' as const, content: msg.content }
    }
    return msg as { role: 'system' | 'user' | 'assistant' | 'developer'; content: string }
  })

  // For reasoning models requesting JSON, append JSON instruction to last user message
  if (isReasoning && options.jsonMode) {
    // Find last user message index (polyfill for findLastIndex)
    let lastUserIdx = -1
    for (let i = messages.length - 1; i >= 0; i--) {
      if (messages[i].role === 'user') {
        lastUserIdx = i
        break
      }
    }
    if (lastUserIdx !== -1) {
      messages[lastUserIdx] = {
        ...messages[lastUserIdx],
        content: messages[lastUserIdx].content + '\n\nIMPORTANT: Respond with valid JSON only, no other text.',
      }
    }
  }

  const params: ChatCompletionCreateParamsNonStreaming = {
    model,
    messages: messages as ChatCompletionCreateParamsNonStreaming['messages'],
  }

  // Token limit parameter differs between model types
  // Newer models (GPT-5+, o-series) use max_completion_tokens
  // Also ensure sufficient tokens for models with reasoning overhead (GPT-5 nano)
  const effectiveMaxTokens = getMinTokenLimit(model, options.maxTokens)
  if (effectiveMaxTokens) {
    if (usesNewTokenParam(model)) {
      params.max_completion_tokens = effectiveMaxTokens
    } else {
      params.max_tokens = effectiveMaxTokens
    }
  }

  // Newer models (o-series, GPT-5+) don't support custom temperature
  if (supportsTemperature(model) && options.temperature !== undefined) {
    params.temperature = options.temperature
  }

  // Reasoning models don't support response_format: json_object
  if (!isReasoning && options.jsonMode) {
    params.response_format = { type: 'json_object' }
  }

  // LiteLLM ChatGPT subscription models reject token limit fields
  if (isLiteLLMChatGPTModel(model)) {
    delete params.max_tokens
    delete params.max_completion_tokens
  }

  return params
}

/**
 * Get OpenAI API key from SystemSettings
 */
async function getOpenAIApiKey(): Promise<string | null> {
  try {
    const setting = await prisma.systemSettings.findUnique({
      where: { key: 'openai_api_key' },
    })
    return setting?.value || process.env.OPENAI_API_KEY || null
  } catch {
    // Fall back to env var if database isn't available
    return process.env.OPENAI_API_KEY || null
  }
}

/**
 * Get Anthropic API key from SystemSettings
 */
async function getAnthropicApiKey(): Promise<string | null> {
  try {
    const setting = await prisma.systemSettings.findUnique({
      where: { key: 'anthropic_api_key' },
    })
    return setting?.value || process.env.ANTHROPIC_API_KEY || null
  } catch {
    return process.env.ANTHROPIC_API_KEY || null
  }
}

/**
 * Get custom base URL for OpenAI-compatible providers.
 * Supports OpenRouter, Together AI, Groq, local models, etc.
 * Set via Settings → AI or OPENAI_BASE_URL env var.
 */
async function getBaseURL(): Promise<string | undefined> {
  try {
    const setting = await prisma.systemSettings.findUnique({
      where: { key: 'openai_base_url' },
    })
    return setting?.value || process.env.OPENAI_BASE_URL || undefined
  } catch {
    return process.env.OPENAI_BASE_URL || undefined
  }
}

/**
 * Create OpenAI client instance.
 * Supports custom baseURL for OpenAI-compatible providers
 * (OpenRouter, Groq, Together AI, local models, etc.)
 */
async function createOpenAIClient(): Promise<OpenAI | null> {
  const apiKey = await getOpenAIApiKey()
  const provider = await getConfiguredProvider()

  // LiteLLM proxy may not require a real API key
  const effectiveApiKey = apiKey || (provider === 'litellm' ? 'sk-litellm' : null)

  if (!effectiveApiKey) {
    console.warn('OpenAI API key not configured')
    return null
  }

  const baseURL = await getBaseURL()

  if (baseURL) {
    console.log(`[OpenAI] Using custom base URL: ${baseURL} (provider: ${provider})`)
  }

  return new OpenAI({
    apiKey: effectiveApiKey,
    ...(baseURL ? { baseURL } : {}),
  })
}

/**
 * Check if a model is a Claude Opus model (supports extended thinking).
 */
function isClaudeOpusModel(model: string): boolean {
  return model.toLowerCase().includes('opus')
}

/**
 * Create an Anthropic adapter that wraps the Anthropic SDK behind the
 * same `.chat.completions.create()` surface as OpenAI. This allows all
 * AI service files to work with zero changes.
 */
async function createAnthropicAdapter(): Promise<AnthropicClientAdapter | null> {
  const apiKey = await getAnthropicApiKey()
  if (!apiKey) {
    console.warn('Anthropic API key not configured')
    return null
  }

  const baseURL = await getBaseURL()
  const anthropic = new Anthropic({
    apiKey,
    ...(baseURL ? { baseURL } : {}),
  })

  if (baseURL) {
    console.log(`[Anthropic] Using custom base URL: ${baseURL}`)
  }

  return {
    __isAnthropicAdapter: true,
    chat: {
      completions: {
        async create(params: ChatCompletionCreateParamsNonStreaming): Promise<OpenAI.Chat.Completions.ChatCompletion> {
          // Extract system messages → Anthropic's system parameter
          const systemMessages: string[] = []
          const userAssistantMessages: Anthropic.MessageParam[] = []

          for (const msg of params.messages) {
            const content = typeof msg.content === 'string' ? msg.content : ''
            if (msg.role === 'system' || msg.role === 'developer') {
              systemMessages.push(content)
            } else {
              userAssistantMessages.push({
                role: msg.role === 'assistant' ? 'assistant' : 'user',
                content,
              })
            }
          }

          // Ensure messages start with a user message (Anthropic requirement)
          if (userAssistantMessages.length === 0 || userAssistantMessages[0].role !== 'user') {
            userAssistantMessages.unshift({ role: 'user', content: 'Hello' })
          }

          // Determine max_tokens (required by Anthropic, default 16384)
          const maxTokens = params.max_tokens ?? params.max_completion_tokens ?? 16384

          // Build Anthropic request
          const anthropicParams: Anthropic.MessageCreateParamsNonStreaming = {
            model: params.model,
            max_tokens: maxTokens,
            messages: userAssistantMessages,
            ...(systemMessages.length > 0 ? { system: systemMessages.join('\n\n') } : {}),
          }

          // Add temperature if present (Anthropic supports 0-1)
          if (params.temperature !== undefined && params.temperature !== null) {
            anthropicParams.temperature = params.temperature
          }

          // Extended thinking for Opus models
          if (isClaudeOpusModel(params.model)) {
            anthropicParams.thinking = { type: 'enabled', budget_tokens: Math.min(8192, maxTokens - 1) }
          }

          // Call Anthropic API
          let response = await anthropic.messages.create(anthropicParams)

          // Extract text from response (skip thinking blocks)
          let responseText = response.content
            .filter((block): block is Anthropic.TextBlock => block.type === 'text')
            .map((block) => block.text)
            .join('')

          // JSON retry: if response_format was set but response isn't valid JSON
          const wantsJson = params.response_format && 'type' in params.response_format && params.response_format.type === 'json_object'
          if (wantsJson && responseText) {
            try {
              JSON.parse(responseText)
            } catch {
              // Retry once with explicit JSON instruction
              const retryMessages = [...userAssistantMessages]
              const lastIdx = retryMessages.length - 1
              if (lastIdx >= 0 && retryMessages[lastIdx].role === 'user') {
                retryMessages[lastIdx] = {
                  ...retryMessages[lastIdx],
                  content: retryMessages[lastIdx].content + '\n\nIMPORTANT: You MUST respond with valid JSON only. No markdown, no extra text, just a JSON object or array.',
                }
              }

              const retryParams: Anthropic.MessageCreateParamsNonStreaming = {
                ...anthropicParams,
                messages: retryMessages,
              }

              response = await anthropic.messages.create(retryParams)
              responseText = response.content
                .filter((block): block is Anthropic.TextBlock => block.type === 'text')
                .map((block) => block.text)
                .join('')
            }
          }

          // Normalize response to OpenAI shape
          return {
            id: response.id,
            object: 'chat.completion' as const,
            created: Math.floor(Date.now() / 1000),
            model: response.model,
            choices: [
              {
                index: 0,
                message: {
                  role: 'assistant' as const,
                  content: responseText || null,
                  refusal: null,
                },
                finish_reason: response.stop_reason === 'end_turn' || response.stop_reason === 'stop_sequence' ? 'stop' : response.stop_reason === 'max_tokens' ? 'length' : 'stop',
                logprobs: null,
              },
            ],
            usage: {
              prompt_tokens: response.usage.input_tokens,
              completion_tokens: response.usage.output_tokens,
              total_tokens: response.usage.input_tokens + response.usage.output_tokens,
              prompt_tokens_details: undefined as any,
              completion_tokens_details: undefined as any,
            },
          }
        },
      },
    },
  }
}

/**
 * Get the AI client singleton.
 * Returns an OpenAI client or an Anthropic adapter (both expose .chat.completions.create()).
 * Returns null if the API key is not configured.
 */
export async function getOpenAI(): Promise<AIClient | null> {
  if (globalForOpenAI.openaiInitialized) {
    return globalForOpenAI.openai || null
  }

  const provider = await getConfiguredProvider()
  const client = provider === 'anthropic'
    ? await createAnthropicAdapter()
    : await createOpenAIClient()

  if (process.env.NODE_ENV !== 'production') {
    globalForOpenAI.openai = client || undefined
    globalForOpenAI.openaiInitialized = true
  }

  return client
}

/**
 * Reset the OpenAI client singleton (e.g., after settings change).
 * Next call to getOpenAI() will create a fresh client.
 */
export function resetOpenAIClient(): void {
  globalForOpenAI.openai = undefined
  globalForOpenAI.openaiInitialized = false
}

/**
 * Check if OpenAI is configured and available
 */
export async function isOpenAIConfigured(): Promise<boolean> {
  const provider = await getConfiguredProvider()
  if (provider === 'litellm') {
    const baseURL = await getBaseURL()
    return !!baseURL
  }
  if (provider === 'anthropic') {
    const apiKey = await getAnthropicApiKey()
    return !!apiKey
  }
  const apiKey = await getOpenAIApiKey()
  return !!apiKey
}

/**
 * List available models from OpenAI API
 */
export async function listAvailableModels(): Promise<{
  success: boolean
  models?: string[]
  error?: string
  manualEntry?: boolean
}> {
  try {
    const provider = await getConfiguredProvider()

    // LiteLLM proxy for ChatGPT subscription doesn't support models.list()
    if (provider === 'litellm') {
      return {
        success: true,
        models: [],
        manualEntry: true,
      }
    }

    // Anthropic: return hardcoded Claude model list
    if (provider === 'anthropic') {
      const apiKey = await getAnthropicApiKey()
      if (!apiKey) {
        return { success: false, error: 'Anthropic API key not configured' }
      }
      return {
        success: true,
        models: [...ANTHROPIC_CLAUDE_MODELS],
      }
    }

    const client = await getOpenAI()

    if (!client) {
      return {
        success: false,
        error: 'OpenAI API key not configured',
      }
    }

    const response = await (client as OpenAI).models.list()
    const chatModels = response.data
      .filter((m) => m.id.includes('gpt') || m.id.includes('o1') || m.id.includes('o3') || m.id.includes('o4'))
      .map((m) => m.id)
      .sort()

    return {
      success: true,
      models: chatModels,
    }
  } catch (error) {
    return {
      success: false,
      error: error instanceof Error ? error.message : 'Unknown error',
    }
  }
}

/**
 * Validate that a specific model is available
 */
export async function validateModel(modelId: string): Promise<{
  valid: boolean
  error?: string
}> {
  try {
    const client = await getOpenAI()

    if (!client) {
      return {
        valid: false,
        error: 'AI API key not configured',
      }
    }

    const provider = await getConfiguredProvider()

    // For Anthropic, use minimal max_tokens
    const params = buildCompletionParams(modelId, {
      messages: [{ role: 'user', content: 'test' }],
      maxTokens: provider === 'anthropic' ? 16 : 1,
    })

    await client.chat.completions.create(params)

    return { valid: true }
  } catch (error) {
    const message = error instanceof Error ? error.message : 'Unknown error'
    // Check for specific model errors
    if (message.includes('does not exist') || message.includes('model_not_found')) {
      return {
        valid: false,
        error: `Model "${modelId}" is not available with your API key`,
      }
    }
    return {
      valid: false,
      error: message,
    }
  }
}

/**
 * Test OpenAI connection with the configured model
 */
export async function testOpenAIConnection(): Promise<{
  success: boolean
  error?: string
  model?: string
  modelTested?: string
}> {
  try {
    const client = await getOpenAI()
    const provider = await getConfiguredProvider()

    if (!client) {
      const label = provider === 'anthropic' ? 'Anthropic' : 'OpenAI'
      return {
        success: false,
        error: `${label} API key not configured`,
      }
    }

    // Get the configured model
    const configuredModel = await getConfiguredModel()

    // Test with the configured model using correct parameters
    const params = buildCompletionParams(configuredModel, {
      messages: [{ role: 'user', content: 'Hello' }],
      maxTokens: provider === 'anthropic' ? 16 : 5,
    })

    const response = await client.chat.completions.create(params)

    return {
      success: true,
      model: response.model,
      modelTested: configuredModel,
    }
  } catch (error) {
    const message = error instanceof Error ? error.message : 'Unknown error'
    const configuredModel = await getConfiguredModel()

    // Check for model-specific errors
    if (message.includes('does not exist') || message.includes('model_not_found') || message.includes('not_found_error')) {
      return {
        success: false,
        error: `Model "${configuredModel}" is not available. Check Settings → AI to select a valid model.`,
        modelTested: configuredModel,
      }
    }

    return {
      success: false,
      error: message,
      modelTested: configuredModel,
    }
  }
}

// Default models for different use cases
export const AI_MODELS = {
  ASSIGNMENT: 'gpt-4o', // Best for complex reasoning
  QUICK: 'gpt-4o-mini', // Faster, cheaper for simple tasks
} as const

/**
 * Get the admin-configured AI model from SystemSettings.
 * Falls back to the provided default if not configured.
 */
export async function getConfiguredModel(fallback: string = AI_MODELS.ASSIGNMENT): Promise<string> {
  try {
    const setting = await prisma.systemSettings.findUnique({
      where: { key: 'ai_model' },
    })
    return setting?.value || process.env.OPENAI_MODEL || fallback
  } catch {
    return process.env.OPENAI_MODEL || fallback
  }
}