apps/backend/src/mcp/servers/DeepSeekServer.ts

import { MCPServer, Tool, Resource } from '@widget-tdc/mcp-types';
import axios, { AxiosInstance, AxiosResponse, AxiosError } from 'axios';
import { z } from 'zod';
import {
  DeepSeekResponse,
  ChatMessage,
  ModelInfo,
  ModelConfig
} from './deepseek-types.js';

const API_CONFIG = {
  BASE_URL: 'https://api.deepseek.com', // Updated to base URL, endpoints usually include /chat/completions
  DEFAULT_MODEL: 'deepseek-reasoner',
  ENDPOINTS: {
    CHAT: '/chat/completions'
  }
} as const;

const MODELS: ModelInfo[] = [
  {
    id: "deepseek-chat",
    name: "DeepSeek Chat",
    description: "General-purpose chat model optimized for dialogue"
  },
  {
    id: "deepseek-reasoner",
    name: "DeepSeek Reasoner",
    description: "Model optimized for reasoning and problem-solving"
  }
];

const MODEL_CONFIGS: ModelConfig[] = [
  {
    id: "temperature",
    name: "Temperature",
    type: "number",
    description: "Controls randomness in the output (0.0 to 2.0)",
    default: 0.7,
    minimum: 0,
    maximum: 2
  },
  {
    id: "max_tokens",
    name: "Maximum Tokens",
    type: "integer",
    description: "Maximum number of tokens to generate",
    default: 8000,
    minimum: 1
  },
  {
    id: "top_p",
    name: "Top P",
    type: "number",
    description: "Controls diversity via nucleus sampling (0.0 to 1.0)",
    default: 1.0,
    minimum: 0,
    maximum: 1
  },
  {
    id: "frequency_penalty",
    name: "Frequency Penalty",
    type: "number",
    description: "Reduces repetition by penalizing frequent tokens (-2.0 to 2.0)",
    default: 0.1,
    minimum: -2,
    maximum: 2
  },
  {
    id: "presence_penalty",
    name: "Presence Penalty",
    type: "number",
    description: "Reduces repetition by penalizing used tokens (-2.0 to 2.0)",
    default: 0,
    minimum: -2,
    maximum: 2
  }
];

export class DeepSeekServer implements MCPServer {
  name = "deepseek-mcp-server";
  version = "0.1.0";
  private axiosInstance: AxiosInstance;
  private conversationHistory: ChatMessage[] = [];
  private apiKey: string;

  constructor() {
    this.apiKey = process.env.DEEPSEEK_API_KEY || '';
    if (!this.apiKey) {
      console.warn("⚠️ DEEPSEEK_API_KEY environment variable is missing. DeepSeek Server will be disabled.");
    }

    this.axiosInstance = axios.create({
      baseURL: API_CONFIG.BASE_URL,
      headers: {
        'Authorization': `Bearer ${this.apiKey}`,
        'Content-Type': 'application/json'
      }
    });

    // Handle API errors
    this.axiosInstance.interceptors.response.use(
      (response: AxiosResponse) => response,
      (error: AxiosError) => {
        console.error("[DeepSeek API Error]", error.response?.data || error.message);
        throw error;
      }
    );
  }

  async listTools(): Promise<Tool[]> {
    if (!this.apiKey) return [];

    return [
      {
        name: "chat_completion",
        description: "Generate a chat completion using DeepSeek models",
        inputSchema: {
          type: "object",
          properties: {
            message: { type: "string", description: "Single message content" },
            messages: { 
              type: "array", 
              items: {
                type: "object",
                properties: {
                  role: { type: "string", enum: ["system", "user", "assistant"] },
                  content: { type: "string" }
                },
                required: ["role", "content"]
              }
            },
            model: { type: "string", default: "deepseek-reasoner" },
            temperature: { type: "number", minimum: 0, maximum: 2, default: 0.7 },
            max_tokens: { type: "integer", minimum: 1, default: 8000 },
            top_p: { type: "number", minimum: 0, maximum: 1, default: 1.0 },
            frequency_penalty: { type: "number", minimum: -2, maximum: 2, default: 0.1 },
            presence_penalty: { type: "number", minimum: -2, maximum: 2, default: 0 }
          }
        }
      },
      {
        name: "multi_turn_chat",
        description: "Engage in a multi-turn conversation with context history",
        inputSchema: {
          type: "object",
          properties: {
            messages: {
              oneOf: [
                { type: "string" },
                {
                  type: "array",
                  items: {
                    type: "object",
                    properties: {
                      role: { type: "string", enum: ["system", "user", "assistant"] },
                      content: {
                        type: "object",
                        properties: {
                          type: { type: "string", const: "text" },
                          text: { type: "string" }
                        }
                      }
                    }
                  }
                }
              ]
            },
            model: { type: "string", default: "deepseek-chat" },
            temperature: { type: "number", minimum: 0, maximum: 2, default: 0.7 },
            max_tokens: { type: "integer", minimum: 1, default: 8000 },
            top_p: { type: "number", minimum: 0, maximum: 1, default: 1.0 },
            frequency_penalty: { type: "number", minimum: -2, maximum: 2, default: 0.1 },
            presence_penalty: { type: "number", minimum: -2, maximum: 2, default: 0 }
          },
          required: ["messages"]
        }
      }
    ];
  }

  async callTool(name: string, args: any): Promise<any> {
    if (!this.apiKey) {
      throw new Error("DeepSeek API key not configured");
    }

    if (name === "chat_completion") {
      return this.handleChatCompletion(args);
    } else if (name === "multi_turn_chat") {
      return this.handleMultiTurnChat(args);
    }

    throw new Error(`Unknown tool: ${name}`);
  }

  private async handleChatCompletion(args: any) {
    const schema = z.object({
      message: z.string().optional(),
      messages: z.array(z.object({
        role: z.enum(['system', 'user', 'assistant']),
        content: z.string()
      })).optional(),
      model: z.string().default('deepseek-reasoner'),
      temperature: z.number().min(0).max(2).default(0.7),
      max_tokens: z.number().positive().int().default(8000),
      top_p: z.number().min(0).max(1).default(1.0),
      frequency_penalty: z.number().min(-2).max(2).default(0.1),
      presence_penalty: z.number().min(-2).max(2).default(0)
    });

    const parsed = schema.parse(args);
    
    let messages: ChatMessage[];
    if (parsed.message) {
      messages = [{ role: 'user', content: parsed.message }];
    } else if (parsed.messages) {
      messages = parsed.messages as ChatMessage[];
    } else {
      throw new Error("Either 'message' or 'messages' must be provided");
    }

    try {
      const response = await this.axiosInstance.post<DeepSeekResponse>(
        API_CONFIG.ENDPOINTS.CHAT,
        {
          messages,
          model: parsed.model,
          temperature: parsed.temperature,
          max_tokens: parsed.max_tokens,
          top_p: parsed.top_p,
          frequency_penalty: parsed.frequency_penalty,
          presence_penalty: parsed.presence_penalty
        }
      );

      return {
        content: [{
          type: "text",
          text: response.data.choices[0].message.content
        }]
      };
    } catch (error) {
      console.error("Error with deepseek-reasoner, falling back to deepseek-chat");
      
      try {
        const fallbackResponse = await this.axiosInstance.post<DeepSeekResponse>(
          API_CONFIG.ENDPOINTS.CHAT,
          {
            messages,
            model: 'deepseek-chat',
            temperature: parsed.temperature,
            max_tokens: parsed.max_tokens,
            top_p: parsed.top_p,
            frequency_penalty: parsed.frequency_penalty,
            presence_penalty: parsed.presence_penalty
          }
        );

        return {
          content: [{
            type: "text",
            text: "Note: Fallback to deepseek-chat due to reasoner error.\n\n" + 
                  fallbackResponse.data.choices[0].message.content
          }]
        };
      } catch (fallbackError) {
        if (axios.isAxiosError(fallbackError)) {
          throw new Error(`DeepSeek API error: ${fallbackError.response?.data?.error?.message ?? fallbackError.message}`);
        }
        throw fallbackError;
      }
    }
  }

  private async handleMultiTurnChat(args: any) {
    const schema = z.object({
      messages: z.union([
        z.string(),
        z.array(z.object({
          role: z.enum(['system', 'user', 'assistant']),
          content: z.object({
            type: z.literal('text'),
            text: z.string()
          })
        }))
      ]),
      model: z.string().default('deepseek-chat'),
      temperature: z.number().min(0).max(2).default(0.7),
      max_tokens: z.number().positive().int().default(8000),
      top_p: z.number().min(0).max(1).default(1.0),
      frequency_penalty: z.number().min(-2).max(2).default(0.1),
      presence_penalty: z.number().min(-2).max(2).default(0)
    });

    const parsed = schema.parse(args);

    let newMessage: { role: 'user' | 'system' | 'assistant', content: string };
    if (typeof parsed.messages === 'string') {
      newMessage = { role: 'user', content: parsed.messages };
    } else {
      const msg = parsed.messages[0];
      newMessage = { role: msg.role, content: msg.content.text };
    }

    // Add new message to history
    this.conversationHistory.push(newMessage);

    // Transform all messages for API
    const transformedMessages = this.conversationHistory.map(msg => ({
      role: msg.role,
      content: msg.content
    }));

    try {
      const response = await this.axiosInstance.post<DeepSeekResponse>(
        API_CONFIG.ENDPOINTS.CHAT,
        {
          messages: transformedMessages,
          model: parsed.model,
          temperature: parsed.temperature,
          max_tokens: parsed.max_tokens,
          top_p: parsed.top_p,
          frequency_penalty: parsed.frequency_penalty,
          presence_penalty: parsed.presence_penalty
        }
      );

      // Add assistant's response to history
      const assistantMessage: ChatMessage = {
        role: 'assistant',
        content: response.data.choices[0].message.content
      };
      this.conversationHistory.push(assistantMessage);

      return {
        content: [{
          type: "text",
          text: assistantMessage.content
        }]
      };
    } catch (error) {
      if (axios.isAxiosError(error)) {
        throw new Error(`DeepSeek API error: ${error.response?.data?.error?.message ?? error.message}`);
      }
      throw error;
    }
  }

  async listResources(): Promise<Resource[]> {
    const models = MODELS.map(model => ({
      uri: `models://${model.id}`,
      name: model.name,
      description: model.description,
      mimeType: "application/json"
    }));

    const configs = MODEL_CONFIGS.map(config => ({
      uri: `config://${config.id}`,
      name: config.name,
      description: config.description,
      mimeType: "application/json"
    }));

    return [...models, ...configs];
  }

  async readResource(uri: string): Promise<string | Buffer> {
    if (uri.startsWith('models://')) {
      const modelId = uri.replace('models://', '');
      const model = MODELS.find(m => m.id === modelId);
      if (model) {
        return JSON.stringify(model, null, 2);
      }
    }

    if (uri.startsWith('config://')) {
      // Handles config://{modelId}/{configId} but the SDK implementation was slightly different in how it matched templates.
      // The original implementation had "config://{modelId}" listing resources, and "config://{modelId}/{configId}" reading.
      // But my listResources returns "config://{configId}" essentially (based on my simplification).
      // Let's stick to the simpler version or adapt if needed.
      
      // If uri is exactly one of the config URIs from listResources
      const configId = uri.replace('config://', '');
      const config = MODEL_CONFIGS.find(c => c.id === configId);
      if (config) {
        return JSON.stringify(config, null, 2);
      }
    }
    
    throw new Error(`Resource not found: ${uri}`);
  }
}