// src/services/agent.ts
// Copyright (C) 2025 DTP Technologies, LLC
// All Rights Reserved

import env from "../config/env.js";

import path from "node:path";
import fs from "node:fs/promises";
import { EventEmitter } from "node:events";

import { IUser } from "../models/user.js";
import { IChatSession, ChatSession } from "../models/chat-session.js";
import { ChatHistory, ChatHistoryStatus } from "../models/chat-history.js";

import { DtpService } from "../lib/service.js";

import {
  getToolByName,
  getToolsBySlugs,
  getToolsExcludingCategory,
  getToolsByMode,
} from "../tools/index.js";
import { DtpTool } from "../lib/tool.js";
import { ChatSessionType } from "../models/chat-session.js";

import { processImageForLlm } from "../lib/image-utils.js";
import AiService from "./ai.js";
import type {
  ChatMessage,
  ToolDefinition,
  ToolCall,
  ChatChunk,
} from "../lib/ai-client.js";

interface ChatContext {
  messages: ChatMessage[];
  tools: ToolDefinition[];
}

class AgentService extends DtpService {
  private events = new EventEmitter();
  private pendingUserMessages: Array<{ content: string; displayName: string }> =
    [];
  private abortController: AbortController | null = null;

  on(event: string, listener: (...args: any[]) => void): this {
    this.events.on(event, listener);
    return this;
  }

  off(event: string, listener: (...args: any[]) => void): this {
    this.events.off(event, listener);
    return this;
  }

  private emit(event: string, ...args: any[]): boolean {
    return this.events.emit(event, ...args);
  }

  /**
   * Queue a user message to be injected into the conversation during processing.
   * The message will be sent at the next opportunity (after tool results).
   */
  queueUserMessage(content: string, displayName: string): void {
    this.pendingUserMessages.push({ content, displayName });
    this.log.info("User message queued", {
      displayName,
      contentLength: content.length,
      totalQueued: this.pendingUserMessages.length,
    });
  }

  /**
   * Clear any pending user messages (e.g., when aborting).
   */
  clearPendingUserMessages(): void {
    this.pendingUserMessages = [];
  }

  /**
   * Get all pending user messages and clear the queue.
   * Returns messages to be injected into the conversation.
   */
  getPendingUserMessages(): Array<{ content: string; displayName: string }> {
    const messages = [...this.pendingUserMessages];
    this.pendingUserMessages = [];
    return messages;
  }

  /**
   * Check if there are any pending user messages.
   */
  hasPendingMessages(): boolean {
    return this.pendingUserMessages.length > 0;
  }

  /**
   * Abort the current agent operation.
   * Sets the abort signal to stop the streaming chat loop.
   */
  abort(): void {
    if (this.abortController) {
      this.abortController.abort();
      this.log.info("Agent operation aborted");
    }
    this.clearPendingUserMessages();
  }

  /**
   * Set the abort controller for the current operation.
   * Called at the start of streamChat to enable abort functionality.
   */
  private setAbortController(controller: AbortController): void {
    this.abortController = controller;
  }

  /**
   * Clear the abort controller after operation completes.
   */
  private clearAbortController(): void {
    this.abortController = null;
  }

  get name(): string {
    return "AgentService";
  }

  get slug(): string {
    return "agent";
  }

  async start(): Promise<void> {
    this.log.info("service started");
  }

  async stop(): Promise<void> {
    this.log.info("service stopped");
  }

  async chat(sessionId: string, prompt: string): Promise<void> {
    const session = await ChatSession.findById(sessionId).populate("user");
    if (!session) {
      throw new Error("Session not found");
    }

    const user = session.user as IUser;
    const userId = user._id.toString();
    const sessionObjId = session._id as any;

    const agentModel = await AiService.getAgentModel(userId);
    if (!agentModel) {
      throw new Error("No agent model configured");
    }

    // Create ChatHistory record at turn start with "processing" status
    const chatHistory = new ChatHistory({
      user: user._id,
      session: session._id,
      prompt,
      mode: session.mode,
      toolCalls: [],
      fileOperations: [],
      response: { thinking: "", message: "" },
      status: ChatHistoryStatus.Processing,
      isSubagent: false,
      inputTokens: 0,
      outputTokens: 0,
    });
    const savedHistory = await chatHistory.save();
    const historyId = savedHistory._id.toString();

    const context = await this.buildAgentContext(session, prompt);

    let fullResponse = "";
    let fullThinking = "";
    let collectedToolCalls: any[] = [];
    let subagentHistoryIds: string[] = [];
    let inputTokens = 0;
    let outputTokens = 0;

    try {
      const result = await this.streamChat(
        session,
        context.messages,
        context.tools,
        agentModel,
        userId,
        historyId,
      );
      fullResponse = result.response;
      fullThinking = result.thinking;
      collectedToolCalls = result.collectedToolCalls;
      subagentHistoryIds = result.subagentHistoryIds;
      inputTokens = result.inputTokens;
      outputTokens = result.outputTokens;

      // Finalize ChatHistory with success status
      await ChatHistory.findByIdAndUpdate(historyId, {
        status: ChatHistoryStatus.Success,
        "response.thinking": fullThinking,
        "response.message": fullResponse,
        toolCalls: (collectedToolCalls ?? []).map((tc) => ({
          tool: {
            name: tc.name,
            callId: tc.callId,
            parameters: tc.parameters,
          },
          response: tc.response,
          fileOperation: tc.fileOperation ?? undefined,
          subagentStats: tc.subagentStats ?? undefined,
        })),
        fileOperations: (collectedToolCalls ?? [])
          .filter((tc) => tc.fileOperation != null)
          .map((tc) => tc.fileOperation),
        inputTokens,
        outputTokens,
        subagentHistory: subagentHistoryIds || [],
      });

      await ChatSession.findByIdAndUpdate(sessionObjId, {
        $inc: {
          "stats.turnCount": 1,
          "stats.inputTokens": inputTokens,
          "stats.outputTokens": outputTokens,
        },
        lastMessageAt: new Date(),
      });

      if (session.stats.turnCount === 0) {
        await this.autoNameSession(session, prompt, fullResponse);
      }
    } catch (error) {
      this.log.error("Chat error", { error, sessionId, prompt });

      // Finalize ChatHistory with failed status
      await ChatHistory.findByIdAndUpdate(historyId, {
        status: ChatHistoryStatus.Failed,
        error: {
          message: error instanceof Error ? error.message : String(error),
          stack: error instanceof Error ? error.stack : undefined,
          timestamp: new Date(),
        },
      });

      await ChatSession.findByIdAndUpdate(sessionObjId, {
        $inc: {
          "stats.turnCount": 1,
          "stats.inputTokens": 0,
          "stats.outputTokens": 0,
        },
        lastMessageAt: new Date(),
      });

      throw error;
    }
  }

  async spawnSubagent(
    session: IChatSession,
    agentType: "explore" | "general",
    prompt: string,
  ): Promise<{
    response: string;
    history: Awaited<ReturnType<typeof ChatHistory.find>>;
    stats: { inputTokens: number; outputTokens: number; toolCallCount: number };
    historyIds: string[];
  }> {
    const user = session.user as IUser;
    const userId = user._id.toString();
    const sessionObjId = session._id as any;

    const agentModel = await AiService.getAgentModel(userId);
    if (!agentModel) {
      throw new Error("No agent model configured");
    }

    const context = await this.buildSubagentContext(session, agentType, prompt);

    try {
      const {
        response: fullResponse,
        stats,
        historyIds,
      } = await this.streamSubagentChat(
        session,
        context.messages,
        context.tools,
        agentModel,
        userId,
      );

      const history = await ChatHistory.find({
        session: sessionObjId,
      })
        .sort({ createdAt: -1 })
        .lean();

      await ChatSession.findByIdAndUpdate(sessionObjId, {
        lastMessageAt: new Date(),
      });

      return { response: fullResponse, history, stats, historyIds };
    } catch (error) {
      this.log.error("Subagent error", {
        error,
        agentType,
        prompt,
      });

      await this.saveChatHistory(
        session,
        prompt,
        undefined,
        undefined,
        error instanceof Error
          ? { message: error.message, stack: error.stack }
          : { message: String(error) },
      );

      throw error;
    }
  }

  private async buildSubagentContext(
    session: IChatSession,
    agentType: "explore" | "general",
    currentPrompt: string,
  ): Promise<ChatContext> {
    const messages: ChatMessage[] = [];

    const systemContent = await this.buildSubagentSystemPrompt(
      session,
      agentType,
    );
    messages.push({ role: "system", content: systemContent });

    await fs.writeFile(
      path.join(
        env.installRoot,
        "logs",
        `subagent.system.${agentType}.${session._id}.md`,
      ),
      systemContent,
      "utf-8",
    );

    messages.push({ role: "user", content: currentPrompt });

    const modeFilteredTools = getToolsByMode(session.mode);
    const subagentTools = getToolsExcludingCategory("browser")
      .filter((t) => modeFilteredTools.includes(t))
      .filter((t) => t.slug !== "subagent")
      .map((t: any) => t.definition);

    this.log.debug("tools available to subagent", {
      toolCount: subagentTools.length,
      toolNames: subagentTools.map((t: any) => t.function.name),
    });

    return { messages, tools: subagentTools };
  }

  private async buildSubagentSystemPrompt(
    session: IChatSession,
    agentType: "explore" | "general",
  ): Promise<string> {
    const user = session.user as IUser;
    const promptFilePath = path.join(
      env.installRoot,
      "data",
      "prompts",
      "subagent",
      agentType,
      "system.md",
    );

    let prompt = await fs.readFile(promptFilePath, "utf-8");
    prompt +=
      "\n\n## SESSION INFORMATION\n\n- Session ID: " +
      session._id +
      "\n- Session Type: " +
      session.type +
      "\n- Created At: " +
      session.createdAt.toISOString();
    prompt +=
      "\n\n## USER INFORMATION\n\n- User ID: " +
      user._id +
      "\n- Username: " +
      user.username +
      "\n- Display Name: " +
      user.displayName;
    prompt +=
      "\n\n## SYSTEM INFORMATION\n\n- Current Time: " +
      new Date().toISOString() +
      "\n- Gadget Version: " +
      env.pkg.version +
      "\n- Timezone: " +
      env.timezone;

    try {
      const agentConfig = await fs.readFile("AGENTS.md", "utf-8");
      this.log.debug("integrating AGENTS.md into subagent system prompt", {
        subagent: agentType,
      });
      prompt += "\n\n## AGENT CONFIGURATION\n\n" + agentConfig;
    } catch (error) {
      this.log.debug("AGENTS.md file not found or failed to load.", {
        subagent: agentType,
        error: error instanceof Error ? error.message : String(error),
      });
    }

    return this.renderSystemPromptTemplate(prompt, session.mode);
  }

  private async streamSubagentChat(
    session: IChatSession,
    messages: ChatMessage[],
    tools: ToolDefinition[],
    llm: string | undefined | null,
    userId: string,
  ): Promise<{
    response: string;
    thinking: string;
    stats: { inputTokens: number; outputTokens: number; toolCallCount: number };
    historyIds: string[];
  }> {
    let fullResponse = "";
    let fullThinking = "";
    let inputTokens = 0;
    let outputTokens = 0;
    let toolCallCount = 0;
    const historyIds: string[] = [];

    if (!llm) {
      throw new Error("No LLM model configured for this subagent");
    }

    const client = await AiService.getAgentClient(userId);
    client.on("finish_reason", (reason) => {
      this.emit("finish_reason", { reason });
    });

    let chatOptions;
    try {
      const providerInfo = await AiService.getAgentProviderInfo(userId);
      if (providerInfo) {
        chatOptions = await AiService.getModelChatOptions(
          userId,
          providerInfo.providerId,
          providerInfo.model,
        );
      }
    } catch (err) {
      this.log.warn("Failed to get model settings, using defaults", {
        error: err,
      });
    }

    let currentMessages: ChatMessage[] = [...messages];
    let continueLoop = true;
    let lastFinishReason: string | undefined;
    let iterations = 0;

    while (continueLoop) {
      iterations++;
      continueLoop = false;

      let iterationThinking = "";
      let iterationResponse = "";
      const toolCallMap = new Map<number, ToolCall>();
      const collectedToolCalls: Array<{
        name: string;
        callId: string;
        parameters: Array<{ name: string; value: string }>;
        response: string;
        fileOperation?: unknown;
        subagentStats?: unknown;
      }> = [];

      this.log.debug("Starting subagent chat request", {
        iteration: iterations,
        messageCount: currentMessages.length,
        hasTools: tools.length > 0,
        model: llm,
      });

      let stream: AsyncGenerator<ChatChunk>;
      try {
        stream = await client.streamChat(
          currentMessages,
          tools,
          llm,
          chatOptions,
        );
      } catch (chatError) {
        this.log.error("Failed to create subagent stream", {
          error:
            chatError instanceof Error ? chatError.message : String(chatError),
          model: llm,
        });
        throw chatError;
      }

      for await (const chunk of stream) {
        if (chunk.thinking) {
          iterationThinking += chunk.thinking;
          fullThinking += chunk.thinking;
        }

        if (chunk.content) {
          iterationResponse += chunk.content;
          fullResponse += chunk.content;
        }

        if (chunk.done && chunk.finish_reason) {
          lastFinishReason = chunk.finish_reason;
        }

        if (chunk.toolCall) {
          const tcDelta = chunk.toolCall;
          const index = tcDelta.index ?? 0;
          let existing = toolCallMap.get(index);

          if (!existing) {
            existing = {
              id:
                tcDelta.id ?? "call-" + Math.random().toString(36).slice(2, 9),
              type: "function",
              function: {
                name: tcDelta.function?.name ?? "",
                arguments: "",
              },
            };
            toolCallMap.set(index, existing);
          } else {
            if (tcDelta.id) existing.id = tcDelta.id;
            if (tcDelta.function?.name)
              existing.function.name = tcDelta.function.name;
          }

          if (tcDelta.function?.arguments !== undefined) {
            const deltaArgs =
              typeof tcDelta.function.arguments === "string"
                ? tcDelta.function.arguments
                : JSON.stringify(tcDelta.function.arguments);

            const currentArgs = existing.function.arguments;
            const trimmedDelta = deltaArgs.trim();
            const isFullObject =
              trimmedDelta.startsWith("{") && trimmedDelta.endsWith("}");

            if (
              isFullObject &&
              (currentArgs.length === 0 || currentArgs === deltaArgs)
            ) {
              existing.function.arguments = deltaArgs;
            } else {
              existing.function.arguments += deltaArgs;
            }
          }
        }
      }

      const finalToolCalls = Array.from(toolCallMap.values());

      currentMessages.push({
        role: "assistant",
        content: iterationResponse,
        tool_calls: finalToolCalls.length > 0 ? finalToolCalls : undefined,
      });

      if (finalToolCalls.length > 0) {
        continueLoop = true;
        for (const toolCall of finalToolCalls) {
          const toolName = toolCall.function.name;
          const toolArgsRaw = toolCall.function.arguments;

          let result = await this.executeTool(toolName, toolArgsRaw, session);

          let toolArgs: any = {};
          try {
            toolArgs = JSON.parse(toolArgsRaw);
          } catch {
            // ignore, executeTool already handled it
          }

          // Extract metadata from result
          let parsedResult: any = null;
          try {
            parsedResult = JSON.parse(result);
          } catch {
            // ignore
          }

          toolCallCount++;
          inputTokens += Math.ceil(toolArgsRaw.length / 4);
          outputTokens += Math.ceil(result.length / 4);

          this.log.debug("Subagent tool result for LLM", {
            toolName,
            resultLength: result.length,
            preview:
              result.length > 100 ? result.substring(0, 100) + "..." : result,
          });

          currentMessages.push({
            role: "tool",
            content: result,
            tool_call_id: toolCall.id,
          });

          const parameters: Array<{ name: string; value: string }> = [];
          for (const [k, v] of Object.entries(toolArgs)) {
            const value = typeof v === "string" ? v : JSON.stringify(v);
            // Skip parameters with empty string values
            if (value !== "") {
              parameters.push({ name: k, value });
            }
          }

          collectedToolCalls.push({
            name: toolName,
            callId: toolCall.id,
            parameters,
            response: result,
            fileOperation: parsedResult?.data?.fileOperation,
            subagentStats: parsedResult?.data?.subagentStats,
          });
        }
      } else {
        this.log.info(
          "Subagent loop terminating: no tool calls found in this iteration.",
        );
      }

      // Check for queued user messages and inject them before next LLM call
      const pendingMessages = this.getPendingUserMessages();
      if (pendingMessages.length > 0) {
        this.log.info("Injecting queued user messages in subagent", {
          count: pendingMessages.length,
        });
        for (const msg of pendingMessages) {
          currentMessages.push({
            role: "user",
            content: msg.content,
          });
          // Emit event so UI can display the message
          this.log.debug("Emitting queued_message_sent event", {
            displayName: msg.displayName,
            content: msg.content.substring(0, 50),
          });
          this.emit("queued_message_sent", {
            displayName: msg.displayName,
            content: msg.content,
          });
        }
        // Continue loop to process the injected messages
        continueLoop = true;
      }

      // Save this turn's history
      const turnHistoryId = await this.saveChatHistory(
        session,
        iterations === 1 ? messages[messages.length - 1]?.content || "" : "...",
        iterationThinking,
        iterationResponse,
        undefined,
        collectedToolCalls,
        undefined,
        true,
        inputTokens,
        outputTokens,
      );
      historyIds.push(turnHistoryId);

      if (lastFinishReason === "length") {
        this.log.info("Subagent loop continuing: finish_reason was 'length'.");
        currentMessages.push({
          role: "user",
          content: "Please continue.",
        });
        continueLoop = true;
        lastFinishReason = undefined;
      }
    }

    // Update global session stats
    await ChatSession.findByIdAndUpdate(session._id, {
      $inc: { "stats.toolCallCount": toolCallCount },
    });

    // Ensure we have a response to return to the parent
    if (!fullResponse.trim()) {
      if (toolCallCount > 0) {
        fullResponse = `Subagent task completed. Executed ${toolCallCount} tool calls to perform the requested actions.`;
      } else {
        fullResponse = "Subagent task completed without additional actions.";
      }
      this.log.info("Subagent provided empty response; synthesized summary.", {
        toolCallCount,
      });
    }

    this.log.info("Subagent chat stream finished.", {
      iterations,
      fullResponseLength: fullResponse.length,
      toolCallCount,
    });

    return {
      response: fullResponse,
      thinking: fullThinking,
      stats: { inputTokens, outputTokens, toolCallCount },
      historyIds,
    };
  }

  private async autoNameSession(
    session: IChatSession,
    prompt: string,
    response: string,
  ): Promise<void> {
    try {
      const user = session.user as IUser;
      const userId = user._id.toString();
      const utilityModel = await AiService.getUtilityModel(userId);

      const namePrompt =
        "Generate a short descriptive name (3-5 words) for this conversation based on the first user message and AI response. Just return the name, nothing else.\n\nUser message: " +
        prompt.trim() +
        "\n\nAI response: " +
        response.substring(0, 500);

      const suggestedName = await this.generate(userId, namePrompt, {
        model: utilityModel,
      });
      const finalName = suggestedName
        .trim()
        .replace(/^"|"$/g, "")
        .substring(0, 100);

      if (finalName) {
        session.name = finalName;
        await session.save();
        this.log.info("Auto-named session", {
          sessionId: session._id,
          name: finalName,
        });
        this.emit("session_updated", {
          sessionId: session._id.toString(),
          name: finalName,
        });
      }
    } catch (error) {
      this.log.warn("Failed to auto-name session", {
        sessionId: session._id,
        error,
      });
    }
  }

  async generate(
    userId: string,
    prompt: string,
    options?: { model?: string; systemPrompt?: string; tools?: string[] },
  ): Promise<string> {
    const messages: ChatMessage[] = [];

    if (options?.systemPrompt) {
      messages.push({ role: "system", content: options.systemPrompt });
    }

    messages.push({ role: "user", content: prompt });

    const tools = options?.tools
      ? getToolsBySlugs(options.tools).map((t) => t.definition as any)
      : ([] as any[]);

    const model = options?.model;
    if (!model) {
      throw new Error("No LLM model specified for generation");
    }

    const client = await AiService.getUtilityClient(userId);
    client.on("finish_reason", (reason) => {
      this.emit("finish_reason", { reason });
    });
    const response = await client.chat(messages, tools, model);

    return response.content;
  }

  private async buildAgentContext(
    session: IChatSession,
    currentPrompt: string,
  ): Promise<ChatContext> {
    const messages: ChatMessage[] = [];
    const mode = session.mode;

    const systemContent = await this.buildAgentSystemPrompt(session);
    messages.push({ role: "system", content: systemContent });

    await fs.writeFile(
      path.join(
        env.installRoot,
        "logs",
        `agent.system.${mode}.${session._id}.md`,
      ),
      systemContent,
      "utf-8",
    );

    const history = await ChatHistory.find({
      session: session._id,
      isSubagent: { $ne: true },
    })
      .sort({ createdAt: -1 })
      .lean();

    for (const turn of history.reverse()) {
      messages.push({ role: "user", content: turn.prompt });

      // Include error information for failed turns
      if (turn.status === "failed" && turn.error) {
        const errorContext = `\n\n[ERROR: ${turn.error.message}]`;
        const toolCallContext =
          turn.toolCalls && turn.toolCalls.length > 0
            ? `\n[Tool calls made before error: ${turn.toolCalls.map((tc) => tc.tool.name).join(", ")}]`
            : "";
        const thinkingContext = turn.response?.thinking
          ? `\n[Thinking before error: ${turn.response.thinking.substring(0, 200)}${turn.response.thinking.length > 200 ? "..." : ""}]`
          : "";

        messages.push({
          role: "assistant",
          content: `[Turn failed${errorContext}${toolCallContext}${thinkingContext}]`,
        });
      } else if (turn.response?.message) {
        messages.push({ role: "assistant", content: turn.response.message });
      }
    }

    messages.push({ role: "user", content: currentPrompt });

    const modeFilteredTools = getToolsByMode(mode);
    let availableTools =
      session.type === ChatSessionType.Extension
        ? modeFilteredTools
        : getToolsExcludingCategory("browser").filter((t) =>
            modeFilteredTools.includes(t),
          );

    // Filter tools based on user configuration
    const userId = (session.user as IUser)._id.toString();
    const userConfiguredTools: DtpTool[] = [];

    for (const tool of availableTools) {
      if (tool.requiresUserConfig()) {
        const isConfigured = await tool.checkUserConfig(userId);
        if (isConfigured) {
          userConfiguredTools.push(tool);
        } else {
          this.log.debug(
            `Tool ${tool.slug} excluded: user configuration required but not found`,
            {
              userId,
              toolSlug: tool.slug,
            },
          );
        }
      } else {
        userConfiguredTools.push(tool);
      }
    }

    availableTools = userConfiguredTools;

    const tools = availableTools.map((t: any) => t.definition);

    this.log.debug("tools available to agent this turn", {
      toolCount: tools.length,
      toolNames: tools.map((t) => t.function.name),
    });

    return { messages, tools };
  }

  private async buildAgentSystemPrompt(session: IChatSession): Promise<string> {
    const user = session.user as IUser;
    const mode = session.mode;
    const promptFilePath = path.join(
      env.installRoot,
      "data",
      "prompts",
      "agent",
      mode,
      "system.md",
    );

    let prompt = await fs.readFile(promptFilePath, "utf-8");
    prompt +=
      "\n\n## SESSION INFORMATION\n\n- Session ID: " +
      session._id +
      "\n- Session Type: " +
      session.type +
      "\n- Created At: " +
      session.createdAt.toISOString();
    prompt +=
      "\n\n## USER INFORMATION\n\n- User ID: " +
      user._id +
      "\n- Username: " +
      user.username +
      "\n- Display Name: " +
      user.displayName;
    prompt +=
      "\n\n## SYSTEM INFORMATION\n\n- Current Time: " +
      new Date().toISOString() +
      "\n- Gadget Version: " +
      env.pkg.version +
      "\n- Timezone: " +
      env.timezone;

    try {
      const agentConfig = await fs.readFile("AGENTS.md", "utf-8");
      this.log.debug("integrating AGENTS.md into system prompt");
      prompt += "\n\n## AGENT CONFIGURATION\n\n" + agentConfig;
    } catch (error) {
      this.log.debug("AGENTS.md file not found or failed to load.", {
        error: error instanceof Error ? error.message : String(error),
      });
    }

    prompt += "\n\n## PINBOARD\n\n";
    if (session.pins.length > 0) {
      for (const pin of session.pins) {
        prompt += "- [" + pin._id?.toString() + "] " + pin.content + "\n";
      }
    } else {
      prompt += "The pinboard is empty. Use the `pin_add` tool to add notes.\n";
    }

    return this.renderSystemPromptTemplate(prompt, mode);
  }

  async renderSystemPromptTemplate(
    prompt: string,
    _mode: string,
  ): Promise<string> {
    let content;

    /*
     * subagents.md
     */
    content = await fs.readFile(
      path.join(env.installRoot, "data", "prompts", "common", "subagents.md"),
      "utf-8",
    );
    prompt = prompt.replace("{{subagent_section}}", content.trim());

    /*
     * scope-block.md
     */
    content = await fs.readFile(
      path.join(env.installRoot, "data", "prompts", "common", "scope-block.md"),
      "utf-8",
    );
    prompt = prompt.replace("{{scope_block}}", content.trim());

    return prompt;
  }

  private async streamChat(
    session: IChatSession,
    messages: ChatMessage[],
    tools: ToolDefinition[],
    llm: string | undefined | null,
    userId: string,
    historyId: string,
  ): Promise<{
    response: string;
    thinking: string;
    collectedToolCalls: Array<{
      name: string;
      callId: string;
      parameters: Array<{ name: string; value: string }>;
      response: string;
      fileOperation?: unknown;
      subagentStats?: unknown;
    }>;
    subagentHistoryIds: string[];
    inputTokens: number;
    outputTokens: number;
  }> {
    let fullResponse = "";
    let fullThinking = "";
    const collectedToolCalls: Array<{
      name: string;
      callId: string;
      parameters: Array<{ name: string; value: string }>;
      response: string;
      fileOperation?: unknown;
      subagentStats?: unknown;
    }> = [];
    const subagentHistoryIds: string[] = [];
    let toolCallCount = 0;
    let inputTokens = 0;
    let outputTokens = 0;

    // Estimate input tokens from message content (rough estimate: 4 chars per token)
    const inputText = messages.map((m) => m.content).join(" ");
    inputTokens = Math.ceil(inputText.length / 4);

    if (!llm) {
      throw new Error("No LLM model configured for this agent");
    }

    const client = await AiService.getAgentClient(userId);
    client.on("finish_reason", (reason) => {
      this.emit("finish_reason", { reason });
    });

    let chatOptions;
    try {
      const providerInfo = await AiService.getAgentProviderInfo(userId);
      if (providerInfo) {
        chatOptions = await AiService.getModelChatOptions(
          userId,
          providerInfo.providerId,
          providerInfo.model,
        );
      }
    } catch (err) {
      this.log.warn("Failed to get model settings, using defaults", {
        error: err,
      });
    }

    let currentMessages: ChatMessage[] = [...messages];
    let continueLoop = true;
    let lastFinishReason: string | undefined;
    let iterations = 0;

    // Create abort controller for this operation
    const abortController = new AbortController();
    this.setAbortController(abortController);

    while (continueLoop) {
      // Check for abort signal at the start of each iteration
      if (abortController.signal.aborted) {
        this.log.info("Agent operation was aborted");
        this.clearAbortController();
        throw new Error("Operation aborted");
      }
      iterations++;
      continueLoop = false;

      // Check for queued user messages at the START of each iteration
      const pendingMessagesAtStart = this.getPendingUserMessages();
      if (pendingMessagesAtStart.length > 0) {
        this.log.info("Injecting queued user messages at start of iteration", {
          count: pendingMessagesAtStart.length,
          iteration: iterations,
        });
        for (const msg of pendingMessagesAtStart) {
          currentMessages.push({
            role: "user",
            content: msg.content,
          });
        }
        // Continue loop to process the injected messages
        continueLoop = true;
      }

      let iterationThinking = "";
      let iterationResponse = "";
      const toolCallMap = new Map<number, ToolCall>();

      this.log.debug("Starting AI chat request", {
        iteration: iterations,
        messageCount: currentMessages.length,
        hasTools: tools.length > 0,
        model: llm,
      });

      let stream: AsyncGenerator<ChatChunk>;
      try {
        stream = await client.streamChat(
          currentMessages,
          tools,
          llm,
          chatOptions,
        );
      } catch (chatError) {
        this.log.error("Failed to create chat stream", {
          error:
            chatError instanceof Error ? chatError.message : String(chatError),
          model: llm,
        });
        throw chatError;
      }

      for await (const chunk of stream) {
        // Check for abort signal during streaming
        if (abortController.signal.aborted) {
          this.log.info("Agent operation was aborted during streaming");
          this.clearAbortController();
          throw new Error("Operation aborted");
        }

        if (chunk.thinking) {
          iterationThinking += chunk.thinking;
          fullThinking += chunk.thinking;
          this.emit("thinking", { text: chunk.thinking });
        }

        if (chunk.content) {
          iterationResponse += chunk.content;
          fullResponse += chunk.content;
          outputTokens += Math.ceil(chunk.content.length / 4);
          this.emit("message", { text: chunk.content });
        }

        if (chunk.done && chunk.finish_reason) {
          lastFinishReason = chunk.finish_reason;
        }

        if (chunk.toolCall) {
          const tcDelta = chunk.toolCall;
          const index = tcDelta.index ?? 0;
          let existing = toolCallMap.get(index);

          if (!existing) {
            existing = {
              id:
                tcDelta.id ?? "call-" + Math.random().toString(36).slice(2, 9),
              type: "function",
              function: {
                name: tcDelta.function?.name ?? "",
                arguments: "",
              },
            };
            toolCallMap.set(index, existing);

            if (existing.function.name) {
              this.emit("tool_call", {
                name: existing.function.name,
                arguments: "",
              });
            }
          } else {
            if (tcDelta.id) existing.id = tcDelta.id;
            if (tcDelta.function?.name) {
              const oldName = existing.function.name;
              existing.function.name = tcDelta.function.name;
              if (!oldName && existing.function.name) {
                this.emit("tool_call", {
                  name: existing.function.name,
                  arguments: "",
                });
              }
            }
          }

          if (tcDelta.function?.arguments !== undefined) {
            const deltaArgs =
              typeof tcDelta.function.arguments === "string"
                ? tcDelta.function.arguments
                : JSON.stringify(tcDelta.function.arguments);

            const currentArgs = existing.function.arguments;
            const trimmedDelta = deltaArgs.trim();
            const isFullObject =
              trimmedDelta.startsWith("{") && trimmedDelta.endsWith("}");

            if (
              isFullObject &&
              (currentArgs.length === 0 || currentArgs === deltaArgs)
            ) {
              existing.function.arguments = deltaArgs;
            } else {
              existing.function.arguments += deltaArgs;
            }
          }
        }
      }

      const finalToolCalls = Array.from(toolCallMap.values());

      currentMessages.push({
        role: "assistant",
        content: iterationResponse,
        tool_calls: finalToolCalls.length > 0 ? finalToolCalls : undefined,
      });

      // Incrementally update ChatHistory with thinking and response from this iteration
      if (iterationThinking || iterationResponse) {
        await ChatHistory.findByIdAndUpdate(historyId, {
          "response.thinking": iterationThinking,
          "response.message": iterationResponse,
        });
      }

      if (finalToolCalls.length > 0) {
        continueLoop = true;
        for (const toolCall of finalToolCalls) {
          const toolName = toolCall.function.name;
          const toolArgsRaw = toolCall.function.arguments;

          let result = await this.executeTool(toolName, toolArgsRaw, session);

          toolCallCount++;

          let toolArgs: any = {};
          try {
            toolArgs = JSON.parse(toolArgsRaw);
          } catch {
            // ignore
          }

          let parsedResult: any = null;
          try {
            parsedResult = JSON.parse(result);
          } catch {
            // ignore
          }

          this.log.debug("Tool result for LLM", {
            toolName,
            resultLength: result.length,
            preview:
              result.length > 100 ? result.substring(0, 100) + "..." : result,
          });

          if (parsedResult?.success && parsedResult.data?.imageBase64) {
            try {
              const processed = await processImageForLlm(
                parsedResult.data.imageBase64,
              );
              currentMessages.push({
                role: "tool",
                content: `Screenshot captured (${processed.metadata.width}x${processed.metadata.height})`,
                images: [processed.base64],
                tool_call_id: toolCall.id,
              });
            } catch (imgErr) {
              currentMessages.push({
                role: "tool",
                content: result,
                tool_call_id: toolCall.id,
              });
            }
          } else {
            currentMessages.push({
              role: "tool",
              content: result,
              tool_call_id: toolCall.id,
            });
          }

          const parameters: Array<{ name: string; value: string }> = [];
          for (const [k, v] of Object.entries(toolArgs)) {
            const value = typeof v === "string" ? v : JSON.stringify(v);
            // Skip parameters with empty string values
            if (value !== "") {
              parameters.push({ name: k, value });
            }
          }

          if (parsedResult?.data?.historyIds) {
            subagentHistoryIds.push(...parsedResult.data.historyIds);
          }

          collectedToolCalls.push({
            name: toolName,
            callId: toolCall.id,
            parameters,
            response: result,
            fileOperation: parsedResult?.data?.fileOperation,
            subagentStats: parsedResult?.data?.subagentStats,
          });

          this.emit("tool_result", {
            tool: toolName,
            result: result,
            fileOperation: parsedResult?.data?.fileOperation,
            subagentStats: parsedResult?.data?.subagentStats,
            subagentPrompt: parsedResult?.data?.subagentPrompt,
            subagentType: parsedResult?.data?.subagentType,
          });

          // Incrementally update ChatHistory with tool calls from this iteration
          await ChatHistory.findByIdAndUpdate(historyId, {
            $push: {
              toolCalls: {
                tool: {
                  name: toolName,
                  callId: toolCall.id,
                  parameters,
                },
                response: result,
                fileOperation: parsedResult?.data?.fileOperation ?? undefined,
                subagentStats: parsedResult?.data?.subagentStats ?? undefined,
              },
            },
            $inc: {
              inputTokens: Math.ceil(toolArgsRaw.length / 4),
              outputTokens: Math.ceil(result.length / 4),
            },
          });
        }
      } else {
        this.log.info(
          "Agent loop terminating: no tool calls found in this iteration.",
        );
      }

      // Check for queued user messages and inject them before next LLM call
      const newPendingMessages = this.getPendingUserMessages();
      if (newPendingMessages.length > 0) {
        this.log.info("Injecting queued user messages", {
          count: newPendingMessages.length,
        });
        for (const msg of newPendingMessages) {
          currentMessages.push({
            role: "user",
            content: msg.content,
          });
          // Emit event so UI can display the message
          this.emit("queued_message_sent", {
            displayName: msg.displayName,
            content: msg.content,
          });
        }
        // Continue loop to process the injected messages
        continueLoop = true;
      }

      if (lastFinishReason === "length") {
        this.log.info("Agent loop continuing: finish_reason was 'length'.");
        currentMessages.push({
          role: "user",
          content: "Please continue.",
        });
        continueLoop = true;
        lastFinishReason = undefined;
      }
    }

    await ChatSession.findByIdAndUpdate(session._id, {
      $inc: { "stats.toolCallCount": toolCallCount },
    });

    this.emit("done", {
      fullResponse,
      fullThinking,
      inputTokens,
      outputTokens,
    });

    this.log.info("Agent chat stream finished.", {
      iterations,
      fullResponseLength: fullResponse.length,
      toolCallCount,
      inputTokens,
      outputTokens,
    });

    // Clear abort controller on successful completion
    this.clearAbortController();

    // Clear any remaining queued messages on successful completion
    this.clearPendingUserMessages();

    return {
      response: fullResponse,
      thinking: fullThinking,
      collectedToolCalls,
      subagentHistoryIds,
      inputTokens,
      outputTokens,
    };
  }

  private async executeTool(
    toolName: string,
    args: string,
    session: IChatSession,
  ): Promise<string> {
    const tool = getToolByName(toolName);
    if (!tool) {
      const error = "Unknown tool: " + toolName;
      this.log.error("Tool not found", { toolName });
      return JSON.stringify({ success: false, error, message: error });
    }

    try {
      this.log.debug("Tool execution started", { toolName, args });

      const parsedArgs = JSON.parse(args);
      const context = { session };

      const result = await tool.execute(context, parsedArgs);

      this.log.debug("Tool execution completed", {
        toolName,
        resultLength: result.length,
      });
      return result;
    } catch (error) {
      const errorMessage =
        error instanceof Error ? error.message : String(error);

      this.log.error("Tool execution failed: " + toolName, {
        tool: toolName,
        arguments: args,
        error: errorMessage,
        stack: error instanceof Error ? error.stack : undefined,
      });

      try {
        this.emit("tool_error", {
          tool: toolName,
          message: "Tool execution failed: " + errorMessage,
        });
      } catch (emitError) {
        this.log.error("Failed to emit tool error event", {
          toolName,
          emitError,
        });
      }

      return JSON.stringify({
        success: false,
        error: "TOOL_EXECUTION_FAILED",
        message: errorMessage,
      });
    }
  }

  private async saveChatHistory(
    session: IChatSession,
    prompt: string,
    thinking: string | undefined,
    message: string | undefined,
    error?: { message: string; stack?: string },
    toolCalls?: Array<{
      name: string;
      callId: string;
      parameters: Array<{ name: string; value: string }>;
      response: string;
      fileOperation?: unknown;
      subagentStats?: unknown;
    }>,
    subagentHistoryIds?: string[],
    isSubagent = false,
    inputTokens = 0,
    outputTokens = 0,
  ): Promise<string> {
    const status = error ? "failed" : "success";
    const user = session.user as IUser;

    const toolCallDocs = (toolCalls ?? []).map((tc) => ({
      tool: {
        name: tc.name,
        callId: tc.callId,
        parameters: tc.parameters,
      },
      response: tc.response,
      fileOperation: tc.fileOperation ?? undefined,
      subagentStats: tc.subagentStats ?? undefined,
    }));

    const fileOpDocs = toolCallDocs
      .filter((tc) => tc.fileOperation != null)
      .map((tc) => tc.fileOperation);

    const history = new ChatHistory({
      user: user._id,
      session: session._id,
      prompt,
      mode: session.mode,
      toolCalls: toolCallDocs,
      fileOperations: fileOpDocs,
      response: { thinking, message },
      status,
      isSubagent,
      subagentHistory: subagentHistoryIds || [],
      error: error
        ? { message: error.message, stack: error.stack, timestamp: new Date() }
        : undefined,
      inputTokens,
      outputTokens,
    });
    const saved = await history.save();
    return saved._id.toString();
  }
}

export default new AgentService();