//! Agentic 循环 — 流式接收 → 工具执行 → 结果回传 LLM → 循环

use std::sync::Arc;
use std::sync::atomic::Ordering;

use tauri::{AppHandle, Emitter};
use tokio::sync::Mutex;

use df_ai::ai_tools::AiToolRegistry;
use df_ai::context::TokenEstimator;
use df_ai::provider::{ChatMessage, CompletionRequest, LlmProvider};

use df_storage::db::Database;
use df_storage::models::AiProviderRecord;

use crate::state::{AppState, LlmConcurrency};

use super::conversation::{save_conversation, TokenAccumulator};
use super::knowledge_inject::maybe_spawn_extraction;
use super::prompt::{build_system_prompt, get_active_provider};
use super::stream_recv::stream_llm;
use super::title::{ensure_conversation_title, spawn_ensure_title};
use super::audit::process_tool_calls;

use super::{AiChatEvent, AiSession};

/// Agentic 循环最大迭代次数
pub(crate) const MAX_AGENT_ITERATIONS: usize = 10;

/// Agentic 循环：流式接收 → 工具执行 → 结果回传 LLM → 循环
///
/// 退出条件：
/// - LLM 只返回文本（无 tool_calls）→ 正常结束
/// - 有工具需要审批 → 暂停循环（generating 保持 true），等 ai_approve 恢复
/// - 达到最大迭代次数 → 正常结束
pub(crate) async fn run_agentic_loop(
    session_arc: Arc<Mutex<AiSession>>,
    tools_arc: Arc<AiToolRegistry>,
    db: Arc<Database>,
    app_handle: AppHandle,
    provider_config: AiProviderRecord,
    system_prompt: String,
    conv_id: String,
    knowledge_config: crate::state::KnowledgeConfig,
    llm_concurrency: LlmConcurrency,
) {
    let provider: Box<dyn LlmProvider> = df_ai::build_provider(
        &provider_config.provider_type,
        &provider_config.base_url,
        &provider_config.api_key,
        &provider_config.default_model,
    );
    let tool_defs = tools_arc.tool_definitions();
    // 停止信号副本：stream_llm 与每轮迭代共享读取，避免重复加锁
    let stop_flag = session_arc.lock().await.stop_flag.clone();

    // token 累加器:loop 生命周期内各轮叠加,退出时传 save_conversation(累加模式落库)
    let mut tokens = TokenAccumulator::default();

    for iteration in 0..MAX_AGENT_ITERATIONS {
        // 用户请求停止 → 收尾退出（已生成文本已在上一轮入库）
        if stop_flag.load(Ordering::SeqCst) {
            let usage = df_ai::provider::TokenUsage {
                prompt_tokens: tokens.prompt(),
                completion_tokens: tokens.completion(),
                total_tokens: tokens.total(),
            };
            // 入口 stop:本轮可能尚未 stream(首轮即停),不记 model——避免把未实际生成的 model 写入 models 数组
            save_conversation(&session_arc, &db, &conv_id, Some(&usage), None).await;
            // 标题生成后台化：不阻塞 Completed emit（失败有 extract_title 兜底）
            spawn_ensure_title(&provider_config, &db, &conv_id, &app_handle, &session_arc, &llm_concurrency);
            let mut session = session_arc.lock().await;
            session.generating = false;
            // generating 复位后再 emit Completed：保证前端收事件时后端已可接下一条(发送队列续发不被"正在生成中"拒绝)
            let _ = app_handle.emit("ai-chat-event", AiChatEvent::AiCompleted { total_tokens: usage.total_tokens, prompt_tokens: tokens.prompt(), completion_tokens: tokens.completion(), conversation_id: Some(conv_id.clone()) });
            return;
        }

        // 新一轮通知前端（第二轮起），前端需新建 assistant 消息
        if iteration > 0 {
            let _ = app_handle.emit("ai-chat-event", AiChatEvent::AiAgentRound {
                round: (iteration + 1) as u32,
                conversation_id: Some(conv_id.clone()),
            });
        }

        // 构建请求消息（超预算时自动裁剪旧消息，保护工具调用三元组 + 最近 6 条）
        let messages = {
            let session = session_arc.lock().await;
            let sys_tokens = TokenEstimator::default().estimate_text(&system_prompt);
            let (history_msgs, _trimmed) = session.messages.build_for_request(sys_tokens);
            let mut msgs = vec![ChatMessage::system(&system_prompt)];
            msgs.extend(history_msgs);
            msgs
        };

        // 预估输入 token(兜底:部分 provider 如 GLM 流式 usage 不报 prompt_tokens,后段用它补)
        let estimated_prompt: u32 = {
            let est = TokenEstimator::default();
            messages.iter().map(|m| est.estimate_message(m)).sum()
        };

        let request = CompletionRequest {
            model: provider_config.default_model.clone(),
            messages,
            temperature: Some(0.7),
            max_tokens: Some(8192),
            stream: true,
            tools: if tool_defs.is_empty() { None } else { Some(tool_defs.clone()) },
            tool_choice: None,
        };

        // LLM 并发限流（全局 + 单对话双层），仅覆盖 stream_llm 调用本身;
        // 工具执行(process_tool_calls)是本地操作无 RPM 成本,permit 在 stream 后立即释放避免占槽
        let _global_permit = llm_concurrency.acquire_global().await;
        let _per_conv_permit = llm_concurrency.acquire_per_conv().await;
        // 流式接收（内部处理 idle timeout / 断连检测 / 停止信号）
        let (full_text, tool_calls_acc, round_usage) = match stream_llm(&*provider, request, &app_handle, &stop_flag, &conv_id).await {
            Some(result) => result,
            None => {
                // 错误已在 stream_llm 中 emit，直接结束
                let mut session = session_arc.lock().await;
                session.generating = false;
                return;
            }
        };
        // stream 结束立即释放 permit,后续工具执行不受限流(本地操作无 RPM 成本)
        drop(_global_permit);
        drop(_per_conv_permit);

        // 累加本轮 token:provider 流式 usage 的 prompt_tokens 为 0 时(GLM 等),用预估输入兜底
        let round_prompt = if round_usage.prompt_tokens == 0 { estimated_prompt } else { round_usage.prompt_tokens };
        tokens.add(round_prompt, round_usage.completion_tokens);

        // 追加 assistant 消息到历史
        let has_tool_calls = !tool_calls_acc.is_empty();
        {
            let mut session = session_arc.lock().await;
            if has_tool_calls {
                let mut order: Vec<u32> = tool_calls_acc.keys().copied().collect();
                order.sort_unstable();
                let ai_tool_calls: Vec<df_ai::provider::ToolCall> = order.iter()
                    .map(|i| {
                        let draft = &tool_calls_acc[i];
                        df_ai::provider::ToolCall::new(&draft.id, &draft.name, &draft.args)
                    })
                    .collect();
                let mut msg = ChatMessage::assistant_with_tools(&full_text, ai_tool_calls);
                msg.model = Some(provider_config.default_model.clone());
                session.messages.push(msg);
            } else if !full_text.is_empty() {
                let mut msg = ChatMessage::assistant(&full_text);
                msg.model = Some(provider_config.default_model.clone());
                session.messages.push(msg);
            }
        }

        // 停止信号：已生成文本入库后退出，不再执行后续工具调用
        if stop_flag.load(Ordering::SeqCst) {
            let usage = df_ai::provider::TokenUsage {
                prompt_tokens: tokens.prompt(),
                completion_tokens: tokens.completion(),
                total_tokens: tokens.total(),
            };
            save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
            // 标题生成后台化：不阻塞 Completed emit（失败有 extract_title 兜底）
            spawn_ensure_title(&provider_config, &db, &conv_id, &app_handle, &session_arc, &llm_concurrency);
            let mut session = session_arc.lock().await;
            session.generating = false;
            // generating 复位后再 emit Completed：保证前端收事件时后端已可接下一条(发送队列续发不被"正在生成中"拒绝)
            let _ = app_handle.emit("ai-chat-event", AiChatEvent::AiCompleted { total_tokens: usage.total_tokens, prompt_tokens: tokens.prompt(), completion_tokens: tokens.completion(), conversation_id: Some(conv_id.clone()) });
            return;
        }

        // 无工具调用 → 最终文本响应，循环结束
        if !has_tool_calls { break; }

        // 处理工具调用（Low 自动执行 / Medium+High 待审批）
        let pending_count = {
            let mut session = session_arc.lock().await;
            process_tool_calls(&mut session, tool_calls_acc, &tools_arc, &db, &app_handle, &conv_id).await
        };

        // 有待审批 → 暂停循环，等待用户审批后通过 ai_approve → try_continue_agent_loop 恢复
        if pending_count > 0 {
            let usage = df_ai::provider::TokenUsage {
                prompt_tokens: tokens.prompt(),
                completion_tokens: tokens.completion(),
                total_tokens: tokens.total(),
            };
            save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
            return; // generating 保持 true
        }

        // 全部自动执行完成 → 继续下一轮
    }

    // 正常完成
    let usage = df_ai::provider::TokenUsage {
        prompt_tokens: tokens.prompt(),
        completion_tokens: tokens.completion(),
        total_tokens: tokens.total(),
    };
    // 落库 + 标题 + 知识提炼打包后台化：不阻塞 generating 复位与 Completed 事件
    // save 先行(extract/title 都读已落库消息);extract 内部 fire-and-forget,与 title 可能并发
    // (均受 per_conv 信号量约束,读写不同字段互不干扰)
    // 并发取舍：与新对话新 loop 的 save 存在低概率并发 upsert，最多丢少量 token 累加（非功能错误，可接受）
    let usage_total = usage.total_tokens;
    {
        let session_arc = session_arc.clone();
        let db = db.clone();
        let conv_id = conv_id.clone();
        let provider_config = provider_config.clone();
        let knowledge_config = knowledge_config.clone();
        let app_handle = app_handle.clone();
        let llm_concurrency = llm_concurrency.clone();
        tauri::async_runtime::spawn(async move {
            save_conversation(&session_arc, &db, &conv_id, Some(&usage), Some(&provider_config.default_model)).await;
            // 知识提炼:需读已落库的对话消息,故在 save 之后
            if let Err(e) = maybe_spawn_extraction(&session_arc, &db, &conv_id, &provider_config, &knowledge_config, llm_concurrency.clone()).await {
                tracing::warn!("知识提炼触发失败(非阻断): {}", e);
            }
            ensure_conversation_title(&provider_config, &db, &conv_id, &app_handle, &session_arc, llm_concurrency).await;
        });
    }

    let mut session = session_arc.lock().await;
    session.generating = false;
    // generating 复位后再 emit Completed：落库/标题/提炼已在后台，前端立即感知完成
    let _ = app_handle.emit("ai-chat-event", AiChatEvent::AiCompleted { total_tokens: usage_total, prompt_tokens: tokens.prompt(), completion_tokens: tokens.completion(), conversation_id: Some(conv_id.clone()) });
}

/// 检查是否所有待审批已处理，如果是则恢复 agentic 循环
pub(crate) async fn try_continue_agent_loop(app: &AppHandle, state: &AppState) {
    let should_continue = {
        let session = state.ai_session.lock().await;
        session.generating && session.pending_approvals.is_empty()
    };

    if !should_continue { return; }

    let provider_config = match get_active_provider(state).await {
        Ok(p) => p,
        Err(_) => return,
    };
    let (lang, conv_id) = {
        let session = state.ai_session.lock().await;
        let lang = session.agent_language.clone().unwrap_or_else(|| "zh-CN".to_string());
        let conv_id = session.active_conversation_id.clone().unwrap_or_default();
        (lang, conv_id)
    };
    let system_prompt = build_system_prompt(state, &lang).await;

    let session_arc = state.ai_session.clone();
    let tools_arc = state.ai_tools.clone();
    let db = state.db.clone();
    let app_handle = app.clone();
    let knowledge_config = state.knowledge_config.lock().await.clone();
    let llm_concurrency = state.llm_concurrency.clone();

    // 恢复循环前通知前端新建 assistant 消息：审批(通过/拒绝)后新一轮文本
    // 不应追加到发起工具调用的旧消息，用 AiAgentRound 隔开
    let _ = app.emit("ai-chat-event", AiChatEvent::AiAgentRound {
        round: 0,
        conversation_id: Some(conv_id.clone()),
    });

    tauri::async_runtime::spawn(async move {
        run_agentic_loop(session_arc, tools_arc, db, app_handle, provider_config, system_prompt, conv_id, knowledge_config, llm_concurrency).await;
    });
}