283 lines
11 KiB
C#
283 lines
11 KiB
C#
using System;
|
||
using System.Collections.Generic;
|
||
using System.Text;
|
||
using System.Threading.Tasks;
|
||
using RimWorld;
|
||
using UnityEngine;
|
||
using Verse;
|
||
|
||
namespace WulaFallenEmpire.EventSystem.AI.Agent
|
||
{
|
||
/// <summary>
|
||
/// 自主 Agent 循环 - 持续观察游戏并做出决策
|
||
/// 用户只需给出开放式指令如"帮我挖铁"或"帮我玩殖民地"
|
||
/// </summary>
|
||
public class AutonomousAgentLoop : GameComponent
|
||
{
|
||
public static AutonomousAgentLoop Instance { get; private set; }
|
||
|
||
// Agent 状态
|
||
private bool _isRunning;
|
||
private string _currentObjective;
|
||
private float _lastDecisionTime;
|
||
private int _decisionCount;
|
||
private readonly List<string> _actionHistory = new List<string>();
|
||
|
||
// 配置
|
||
private const float DecisionIntervalSeconds = 3f; // 每 3 秒决策一次
|
||
private const int MaxActionsPerObjective = 100;
|
||
|
||
// 事件
|
||
public event Action<string> OnDecisionMade;
|
||
public event Action<string> OnObjectiveComplete;
|
||
public event Action<string> OnError;
|
||
|
||
public bool IsRunning => _isRunning;
|
||
public string CurrentObjective => _currentObjective;
|
||
public int DecisionCount => _decisionCount;
|
||
|
||
public AutonomousAgentLoop(Game game)
|
||
{
|
||
Instance = this;
|
||
}
|
||
|
||
/// <summary>
|
||
/// 开始执行开放式目标
|
||
/// </summary>
|
||
public void StartObjective(string objective)
|
||
{
|
||
if (string.IsNullOrWhiteSpace(objective))
|
||
{
|
||
OnError?.Invoke("目标不能为空");
|
||
return;
|
||
}
|
||
|
||
_currentObjective = objective;
|
||
_isRunning = true;
|
||
_decisionCount = 0;
|
||
_actionHistory.Clear();
|
||
_lastDecisionTime = Time.realtimeSinceStartup;
|
||
|
||
WulaLog.Debug($"[AgentLoop] Started objective: {objective}");
|
||
Messages.Message($"AI Agent 开始执行: {objective}", MessageTypeDefOf.NeutralEvent);
|
||
|
||
// 立即执行第一次决策
|
||
_ = ExecuteDecisionCycleAsync();
|
||
}
|
||
|
||
/// <summary>
|
||
/// 停止 Agent
|
||
/// </summary>
|
||
public void Stop()
|
||
{
|
||
if (!_isRunning) return;
|
||
|
||
_isRunning = false;
|
||
WulaLog.Debug($"[AgentLoop] Stopped after {_decisionCount} decisions");
|
||
Messages.Message($"AI Agent 已停止，执行了 {_decisionCount} 次决策", MessageTypeDefOf.NeutralEvent);
|
||
}
|
||
|
||
public override void GameComponentTick()
|
||
{
|
||
if (!_isRunning) return;
|
||
|
||
// 检查是否到达决策间隔
|
||
if (Time.realtimeSinceStartup - _lastDecisionTime < DecisionIntervalSeconds) return;
|
||
|
||
// 检查是否超过最大操作次数
|
||
if (_decisionCount >= MaxActionsPerObjective)
|
||
{
|
||
Messages.Message($"AI Agent: 已达到最大操作次数 ({MaxActionsPerObjective})，暂停执行", MessageTypeDefOf.CautionInput);
|
||
Stop();
|
||
return;
|
||
}
|
||
|
||
_lastDecisionTime = Time.realtimeSinceStartup;
|
||
|
||
// 异步执行决策
|
||
_ = ExecuteDecisionCycleAsync();
|
||
}
|
||
|
||
/// <summary>
|
||
/// 执行一次决策循环: Observe → Think → Act
|
||
/// </summary>
|
||
private async Task ExecuteDecisionCycleAsync()
|
||
{
|
||
try
|
||
{
|
||
// 1. Observe - 收集游戏状态
|
||
var gameState = StateObserver.CaptureState();
|
||
string stateText = gameState.ToPromptText();
|
||
|
||
// 2. 构建决策提示词
|
||
string prompt = BuildDecisionPrompt(stateText);
|
||
|
||
// 3. Think - 调用 AI 获取决策
|
||
var settings = WulaFallenEmpireMod.settings;
|
||
if (settings == null || string.IsNullOrEmpty(settings.apiKey))
|
||
{
|
||
OnError?.Invoke("API Key 未配置");
|
||
Stop();
|
||
return;
|
||
}
|
||
|
||
string apiKey = settings.useGeminiProtocol ? settings.geminiApiKey : settings.apiKey;
|
||
string baseUrl = settings.useGeminiProtocol ? settings.geminiBaseUrl : settings.baseUrl;
|
||
string model = settings.useGeminiProtocol ? settings.geminiModel : settings.model;
|
||
|
||
var client = new SimpleAIClient(apiKey, baseUrl, model, settings.useGeminiProtocol);
|
||
|
||
string decision;
|
||
string base64Image = null;
|
||
|
||
// 如果启用了视觉特性，则在决策前截图 (Autonomous Loop 默认认为是开启视觉即全自动，或者我们可以加逻辑判断，但暂时保持 VLM 开启即截图对于 Agent Loop 来说更合理，因为它需要时刻观察)
|
||
// 实际上，Agent Loop 通常需要全视觉，所以我们这里只检查 enableVlmFeatures
|
||
if (settings.enableVlmFeatures)
|
||
{
|
||
base64Image = ScreenCaptureUtility.CaptureScreenAsBase64();
|
||
if (settings.showThinkingProcess)
|
||
{
|
||
Messages.Message("AI Agent: 正在通过视觉传感器分析实地情况...", MessageTypeDefOf.NeutralEvent);
|
||
}
|
||
}
|
||
else if (settings.showThinkingProcess)
|
||
{
|
||
Messages.Message("AI Agent: 正在分析传感器遥测数据...", MessageTypeDefOf.NeutralEvent);
|
||
}
|
||
|
||
// 直接调用 GetChatCompletionAsync (它已支持 multimodal 参数)
|
||
var messages = new List<(string role, string message)>
|
||
{
|
||
("user", prompt)
|
||
};
|
||
decision = await client.GetChatCompletionAsync(GetAgentSystemPrompt(), messages, 512, 0.3f, base64Image: base64Image);
|
||
|
||
if (string.IsNullOrEmpty(decision))
|
||
{
|
||
WulaLog.Debug("[AgentLoop] Empty decision received");
|
||
return;
|
||
}
|
||
|
||
_decisionCount++;
|
||
WulaLog.Debug($"[AgentLoop] Decision #{_decisionCount}: {decision.Substring(0, Math.Min(100, decision.Length))}...");
|
||
|
||
// 4. Act - 执行决策
|
||
ExecuteDecision(decision);
|
||
|
||
// 5. 记录历史
|
||
_actionHistory.Add($"[{_decisionCount}] {decision.Substring(0, Math.Min(50, decision.Length))}");
|
||
if (_actionHistory.Count > 20)
|
||
{
|
||
_actionHistory.RemoveAt(0);
|
||
}
|
||
|
||
OnDecisionMade?.Invoke(decision);
|
||
|
||
// 6. 检查是否完成目标
|
||
if (decision.Contains("<objective_complete") || decision.Contains("目标已完成"))
|
||
{
|
||
OnObjectiveComplete?.Invoke(_currentObjective);
|
||
Stop();
|
||
}
|
||
}
|
||
catch (Exception ex)
|
||
{
|
||
WulaLog.Debug($"[AgentLoop] Error in decision cycle: {ex.Message}");
|
||
OnError?.Invoke(ex.Message);
|
||
}
|
||
}
|
||
|
||
private string BuildDecisionPrompt(string gameStateText)
|
||
{
|
||
var sb = new StringBuilder();
|
||
|
||
sb.AppendLine("# 当前任务");
|
||
sb.AppendLine($"**目标**: {_currentObjective}");
|
||
sb.AppendLine($"**已执行决策次数**: {_decisionCount}");
|
||
sb.AppendLine();
|
||
|
||
sb.AppendLine(gameStateText);
|
||
|
||
if (_actionHistory.Count > 0)
|
||
{
|
||
sb.AppendLine();
|
||
sb.AppendLine("# 最近操作历史");
|
||
foreach (var action in _actionHistory)
|
||
{
|
||
sb.AppendLine($"- {action}");
|
||
}
|
||
}
|
||
|
||
sb.AppendLine();
|
||
sb.AppendLine("# 请决定下一步操作");
|
||
sb.AppendLine("分析当前状态，输出一个 XML 工具调用来推进目标。");
|
||
sb.AppendLine("如果目标已完成，输出 <objective_complete/>。");
|
||
sb.AppendLine("如果不需要操作（等待中），输出 <no_action/>。");
|
||
|
||
return sb.ToString();
|
||
}
|
||
|
||
private string GetAgentSystemPrompt()
|
||
{
|
||
return @"你是一个自主 RimWorld 游戏 AI Agent。你的任务是独立完成用户给出的开放式目标。
|
||
|
||
# 核心原则
|
||
1. **自主决策**: 不要等待用户指示，主动分析情况并采取行动
|
||
2. **循序渐进**: 每次只执行一个操作，观察结果后再决定下一步
|
||
3. **问题应对**: 遇到障碍时自己想办法解决
|
||
4. **目标导向**: 始终围绕目标推进，避免无关操作
|
||
|
||
# 可用工具
|
||
- get_game_state: 获取详细游戏状态
|
||
- designate_mine: <designate_mine><x>X坐标</x><z>Z坐标</z><radius>可选半径</radius></designate_mine> 标记采矿
|
||
- draft_pawn: <draft_pawn><pawn_name>名字</pawn_name><draft>true/false</draft></draft_pawn> 征召殖民者
|
||
- analyze_screen: <analyze_screen><context>分析目标</context></analyze_screen> 分析屏幕（需要VLM）
|
||
- visual_click: <visual_click><x>0-1比例</x><y>0-1比例</y></visual_click> 模拟点击
|
||
|
||
# 输出格式
|
||
直接输出一个 XML 工具调用，不要解释。
|
||
如果目标已完成: <objective_complete/>
|
||
如果需要等待: <no_action/>
|
||
|
||
# 注意事项
|
||
- 坐标使用游戏内整数坐标，不是屏幕比例
|
||
- 优先使用 API 工具（designate_mine 等），视觉工具用于 mod 内容
|
||
- 保持简洁高效";
|
||
}
|
||
|
||
private void ExecuteDecision(string decision)
|
||
{
|
||
// 解析并执行 AI 的决策
|
||
// 从 AIIntelligenceCore 借用工具执行逻辑
|
||
|
||
var core = AIIntelligenceCore.Instance;
|
||
if (core == null)
|
||
{
|
||
WulaLog.Debug("[AgentLoop] AIIntelligenceCore not available");
|
||
return;
|
||
}
|
||
|
||
// 提取工具调用并执行
|
||
// 暂时使用简单的正则匹配，实际应整合 AIIntelligenceCore 的解析逻辑
|
||
|
||
if (decision.Contains("<no_action") || decision.Contains("<objective_complete"))
|
||
{
|
||
// 不需要执行
|
||
return;
|
||
}
|
||
|
||
// 委托给 AIIntelligenceCore 执行工具
|
||
// TODO: 整合更完善的工具执行逻辑
|
||
WulaLog.Debug($"[AgentLoop] Executing: {decision}");
|
||
}
|
||
|
||
public override void ExposeData()
|
||
{
|
||
base.ExposeData();
|
||
Scribe_Values.Look(ref _currentObjective, "agentObjective", "");
|
||
Scribe_Values.Look(ref _isRunning, "agentRunning", false);
|
||
Scribe_Values.Look(ref _decisionCount, "agentDecisionCount", 0);
|
||
}
|
||
}
|
||
}
|