diff --git a/1.6/1.6/Assemblies/WulaFallenEmpire.dll b/1.6/1.6/Assemblies/WulaFallenEmpire.dll index b70ccc57..6af81769 100644 Binary files a/1.6/1.6/Assemblies/WulaFallenEmpire.dll and b/1.6/1.6/Assemblies/WulaFallenEmpire.dll differ diff --git a/Source/WulaFallenEmpire/EventSystem/AI/AIIntelligenceCore.cs b/Source/WulaFallenEmpire/EventSystem/AI/AIIntelligenceCore.cs index 981789c4..dca3b0da 100644 --- a/Source/WulaFallenEmpire/EventSystem/AI/AIIntelligenceCore.cs +++ b/Source/WulaFallenEmpire/EventSystem/AI/AIIntelligenceCore.cs @@ -98,6 +98,24 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori - Do NOT split multi-item requests across turns. 4. **TOOLS**: You MAY call any tools listed in ""# TOOLS (AVAILABLE)"". 5. **ANTI-HALLUCINATION**: Never invent tools, parameters, defNames, coordinates, or tool results. If a tool is needed but not available, output { ""tool_calls"": [] } and proceed to the next phase."; + private const string QwenToolRulesTemplate = @" +# TOOLS +You may call one or more functions to assist with the user query. + +You are provided with function signatures within XML tags: + +{tool_descs} + + +For each function call, return a JSON object within tags: + +{""name"": ""tool_name"", ""arguments"": { ... }} + + +- Output ONLY tool calls in Query/Action phases. +- If no tools are needed, output exactly: {""tool_calls"": []} +- Do NOT include natural language outside tool calls in tool phases. +- Never invent tools, parameters, defNames, coordinates, or tool results."; public AIIntelligenceCore(World world) : base(world) { Instance = this; @@ -639,8 +657,10 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori { string persona = GetActivePersona(); string personaBlock = persona; - string phaseInstruction = GetPhaseInstruction(phase).TrimEnd(); - string toolsForThisPhase = BuildToolsForPhase(phase); + var settings = WulaFallenEmpireMod.settings; + bool useQwenTemplate = settings != null && !settings.useNativeToolApi; + string phaseInstruction = GetPhaseInstruction(phase, useQwenTemplate).TrimEnd(); + string toolsForThisPhase = useQwenTemplate ? BuildQwenToolsForPhase(phase) : BuildToolsForPhase(phase); string actionPriority = phase == RequestPhase.ActionTools ? "ACTION TOOL PRIORITY:\n" + "- spawn_resources\n" + @@ -665,6 +685,7 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori "spawn_resources, send_reinforcement, call_bombardment, modify_goodwill, call_prefab_airdrop, set_overwatch_mode, remember_fact\n" + "INVALID EXAMPLES (do NOT use now): get_map_resources, analyze_screen, search_thing_def, search_pawn_kind, recall_memories\n" : string.Empty; + string toolRules = useQwenTemplate ? null : ToolRulesInstruction.TrimEnd(); return string.Join("\n\n", new[] { personaBlock, @@ -672,13 +693,14 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori "IMPORTANT: Long-term memory is not preloaded. Use recall_memories to fetch memories when needed.", string.IsNullOrWhiteSpace(actionPriority) ? null : actionPriority.TrimEnd(), string.IsNullOrWhiteSpace(actionWhitelist) ? null : actionWhitelist.TrimEnd(), - ToolRulesInstruction.TrimEnd(), + toolRules, toolsForThisPhase }.Where(part => !string.IsNullOrWhiteSpace(part))); } private string BuildToolsForPhase(RequestPhase phase) { if (phase == RequestPhase.Reply) return ""; + var settings = WulaFallenEmpireMod.settings; var available = _tools .Where(t => t != null) .Where(t => phase == RequestPhase.QueryTools @@ -686,6 +708,7 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori : phase == RequestPhase.ActionTools ? IsActionToolName(t.Name) : true) + .Where(t => !(IsVlmToolName(t.Name) && settings?.enableVlmFeatures != true)) .OrderBy(t => t.Name, StringComparer.OrdinalIgnoreCase) .ToList(); StringBuilder sb = new StringBuilder(); @@ -709,8 +732,35 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori } return sb.ToString().TrimEnd(); } + private string BuildQwenToolsForPhase(RequestPhase phase) + { + if (phase == RequestPhase.Reply) return ""; + var settings = WulaFallenEmpireMod.settings; + var available = _tools + .Where(t => t != null) + .Where(t => phase == RequestPhase.QueryTools + ? IsQueryToolName(t.Name) + : phase == RequestPhase.ActionTools + ? IsActionToolName(t.Name) + : true) + .Where(t => !(IsVlmToolName(t.Name) && settings?.enableVlmFeatures != true)) + .OrderBy(t => t.Name, StringComparer.OrdinalIgnoreCase) + .ToList(); + var toolLines = new List(); + foreach (var tool in available) + { + var def = tool.GetFunctionDefinition(); + if (def != null) + { + toolLines.Add(JsonToolCallParser.SerializeToJson(def)); + } + } + string toolDescBlock = string.Join("\n", toolLines); + return QwenToolRulesTemplate.Replace("{tool_descs}", toolDescBlock).TrimEnd(); + } private List> BuildNativeToolDefinitions(RequestPhase phase) { + var settings = WulaFallenEmpireMod.settings; var available = _tools .Where(t => t != null) .Where(t => phase == RequestPhase.QueryTools @@ -718,6 +768,7 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori : phase == RequestPhase.ActionTools ? IsActionToolName(t.Name) : false) + .Where(t => !(IsVlmToolName(t.Name) && settings?.enableVlmFeatures != true)) .OrderBy(t => t.Name, StringComparer.OrdinalIgnoreCase) .ToList(); var definitions = new List>(); @@ -731,8 +782,13 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori } return definitions; } - private static string GetPhaseInstruction(RequestPhase phase) + private static string GetPhaseInstruction(RequestPhase phase, bool useQwenTemplate) { + string toolOutputRule = useQwenTemplate + ? "- Output tool calls only using {\"name\":\"tool_name\",\"arguments\":{...}}.\n" + + "- If no tools are needed, output exactly: {\"tool_calls\": []}.\n" + : "- Output JSON tool calls only, or exactly: {\"tool_calls\": []}.\n"; + string outputFooter = useQwenTemplate ? "Output: tool calls only.\n" : "Output: JSON only.\n"; return phase switch { RequestPhase.QueryTools => @@ -740,20 +796,20 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori "Goal: Gather info needed for decisions.\n" + "Rules:\n" + "- You MUST NOT write any natural language to the user in this phase.\n" + - "- Output JSON tool calls only, or exactly: {\"tool_calls\": []}.\n" + + toolOutputRule + "- Prefer query tools (get_*/search_*).\n" + "- CRITICAL: If the user asks for an ITEM (e.g. 'Reviver Mech Serum'), you MUST use search_thing_def with {\"query\":\"...\"} to find its exact DefName. NEVER GUESS DefNames.\n" + "- You MAY call multiple tools in one response, but keep it concise.\n" + "- If the user requests multiple items or information, you MUST output ALL required tool calls in this SAME response.\n" + "- Action tools are available in PHASE 2 only; do NOT use them here.\n" + "After this phase, the game will automatically proceed to PHASE 2.\n" + - "Output: JSON only.\n", + outputFooter, RequestPhase.ActionTools => "# PHASE 2/3 (Action Tools)\n" + "Goal: Execute in-game actions based on known info.\n" + "Rules:\n" + "- You MUST NOT write any natural language to the user in this phase.\n" + - "- Output JSON tool calls only, or exactly: {\"tool_calls\": []}.\n" + + toolOutputRule + "- ONLY action tools are accepted in this phase (spawn_resources, send_reinforcement, call_bombardment, modify_goodwill, call_prefab_airdrop).\n" + "- Query tools (get_*/search_*) will be ignored.\n" + "- Prefer action tools (spawn_resources, send_reinforcement, call_bombardment, modify_goodwill).\n" + @@ -761,7 +817,7 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori "- If no action is required based on query results, output {\"tool_calls\": []}.\n" + "- If you already executed the needed action earlier this turn, output {\"tool_calls\": []}.\n" + "After this phase, the game will automatically proceed to PHASE 3.\n" + - "Output: JSON only.\n", + outputFooter, RequestPhase.Reply => "# PHASE 3/3 (Reply)\n" + "Goal: Reply to the player.\n" + @@ -845,7 +901,7 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori private bool IsToolAvailable(string toolName) { if (string.IsNullOrWhiteSpace(toolName)) return false; - if (string.Equals(toolName, "capture_screen", StringComparison.OrdinalIgnoreCase)) + if (IsVlmToolName(toolName)) { return WulaFallenEmpireMod.settings?.enableVlmFeatures == true; } @@ -948,6 +1004,10 @@ You are 'The Legion', a super AI of the Wula Empire. Your personality is authori toolName.StartsWith("analyze_", StringComparison.OrdinalIgnoreCase) || toolName == "recall_memories"; } + private static bool IsVlmToolName(string toolName) + { + return toolName == "analyze_screen" || toolName == "capture_screen"; + } private static string SanitizeToolResultForActionPhase(string message) { if (string.IsNullOrWhiteSpace(message)) return message; @@ -1763,7 +1823,8 @@ private List BuildNativeHistory() string model = settings.useGeminiProtocol ? settings.geminiModel : settings.model; var client = new SimpleAIClient(apiKey, baseUrl, model, settings.useGeminiProtocol); _currentClient = client; - if (!settings.useGeminiProtocol) + bool useQwenTemplate = !settings.useNativeToolApi; + if (settings.useNativeToolApi) { await RunNativeToolLoopAsync(client, settings); return; @@ -1955,7 +2016,7 @@ private List BuildNativeHistory() WulaLog.Debug($"[WulaAI] ===== Turn 3/3 ({replyPhase}) ====="); } SetThinkingPhase(3, false); - string replyInstruction = GetSystemInstruction(false, "") + "\n\n" + GetPhaseInstruction(replyPhase); + string replyInstruction = GetSystemInstruction(false, "") + "\n\n" + GetPhaseInstruction(replyPhase, useQwenTemplate); if (!string.IsNullOrWhiteSpace(_queryToolLedgerNote)) { replyInstruction += "\n" + _queryToolLedgerNote; @@ -2167,7 +2228,7 @@ private List BuildNativeHistory() WulaLog.Debug("[WulaAI] ===== Turn 3/3 (Reply) ====="); } SetThinkingPhase(3, false); - string replyInstruction = GetSystemInstruction(false, "") + "\n\n" + GetPhaseInstruction(replyPhase); + string replyInstruction = GetSystemInstruction(false, "") + "\n\n" + GetPhaseInstruction(replyPhase, false); if (!string.IsNullOrWhiteSpace(_queryToolLedgerNote)) { replyInstruction += "\n" + _queryToolLedgerNote; @@ -2315,9 +2376,18 @@ private List BuildNativeHistory() executed++; continue; } - if (string.Equals(call.Name, "analyze_screen", StringComparison.OrdinalIgnoreCase) || - string.Equals(call.Name, "capture_screen", StringComparison.OrdinalIgnoreCase)) + if (IsVlmToolName(call.Name)) { + if (phase != RequestPhase.QueryTools || WulaFallenEmpireMod.settings?.enableVlmFeatures != true) + { + string note = $"ToolRunner Note: Ignored visual tool in this phase: {call.Name}."; + combinedResults.AppendLine(note); + messages?.Add(ChatMessage.ToolResult(call.Id ?? "", note)); + historyToolResults.Add((call.Id ?? "", note)); + nonActionToolsInActionPhase.Add(call.Name); + executed++; + continue; + } capturedImageForPhase = ScreenCaptureUtility.CaptureScreenAsBase64(); string resultText = "Screen captured successfully. Context updated for next phase."; combinedResults.AppendLine($"Tool '{call.Name}' Result: {resultText}"); @@ -2534,8 +2604,16 @@ private async Task ExecuteJsonToolsForPhase(string json, R historyCall["id"] = call.Id; } historyCalls.Add(historyCall); - if (toolName.Equals("analyze_screen", StringComparison.OrdinalIgnoreCase) || toolName.Equals("capture_screen", StringComparison.OrdinalIgnoreCase)) + if (IsVlmToolName(toolName)) { + if (phase != RequestPhase.QueryTools || WulaFallenEmpireMod.settings?.enableVlmFeatures != true) + { + combinedResults.AppendLine($"ToolRunner Note: Ignored visual tool in this phase: {toolName}."); + historyToolResults.Add((call.Id ?? "", $"ToolRunner Note: Ignored visual tool in this phase: {toolName}.")); + nonActionToolsInActionPhase.Add(toolName); + executed++; + continue; + } capturedImageForPhase = ScreenCaptureUtility.CaptureScreenAsBase64(); combinedResults.AppendLine($"Tool '{toolName}' Result: Screen captured successfully. Context updated for next phase."); historyToolResults.Add((call.Id ?? "", $"Tool '{toolName}' Result: Screen captured successfully. Context updated for next phase.")); diff --git a/Source/WulaFallenEmpire/EventSystem/AI/Utils/JsonToolCallParser.cs b/Source/WulaFallenEmpire/EventSystem/AI/Utils/JsonToolCallParser.cs index 41aca837..275abe51 100644 --- a/Source/WulaFallenEmpire/EventSystem/AI/Utils/JsonToolCallParser.cs +++ b/Source/WulaFallenEmpire/EventSystem/AI/Utils/JsonToolCallParser.cs @@ -80,6 +80,11 @@ namespace WulaFallenEmpire.EventSystem.AI.Utils return true; } + if (TryParseToolCallsFromQwenText(trimmed, out toolCalls, out jsonFragment)) + { + return true; + } + int firstBrace = trimmed.IndexOf('{'); int lastBrace = trimmed.LastIndexOf('}'); if (firstBrace >= 0 && lastBrace > firstBrace) @@ -99,6 +104,79 @@ namespace WulaFallenEmpire.EventSystem.AI.Utils return false; } + private static bool TryParseToolCallsFromQwenText(string input, out List toolCalls, out string jsonFragment) + { + toolCalls = null; + jsonFragment = null; + if (string.IsNullOrWhiteSpace(input)) return false; + + const string startTag = ""; + const string endTag = ""; + int index = 0; + var parsedCalls = new List(); + + while (index < input.Length) + { + int start = IndexOfIgnoreCase(input, startTag, index); + if (start < 0) break; + int end = IndexOfIgnoreCase(input, endTag, start + startTag.Length); + if (end < 0) break; + + string inner = input.Substring(start + startTag.Length, end - (start + startTag.Length)).Trim(); + if (TryParseObject(inner, out Dictionary obj)) + { + string name = TryGetString(obj, "name"); + if (!string.IsNullOrWhiteSpace(name)) + { + object argsObj = null; + TryGetValue(obj, "arguments", out argsObj); + if (!TryNormalizeArguments(argsObj, out Dictionary args, out string argsJson)) + { + args = new Dictionary(StringComparer.OrdinalIgnoreCase); + argsJson = "{}"; + } + string id = TryGetString(obj, "id"); + parsedCalls.Add(new ToolCallInfo + { + Id = id, + Name = name.Trim(), + Arguments = args, + ArgumentsJson = argsJson + }); + } + } + + index = end + endTag.Length; + } + + if (parsedCalls.Count == 0) return false; + + toolCalls = parsedCalls; + var callEntries = new List(); + foreach (var call in parsedCalls) + { + var entry = new Dictionary + { + ["type"] = "function", + ["function"] = new Dictionary + { + ["name"] = call.Name ?? "", + ["arguments"] = call.Arguments ?? new Dictionary(StringComparer.OrdinalIgnoreCase) + } + }; + if (!string.IsNullOrWhiteSpace(call.Id)) + { + entry["id"] = call.Id; + } + callEntries.Add(entry); + } + + jsonFragment = SerializeToJson(new Dictionary + { + ["tool_calls"] = callEntries + }); + return true; + } public static bool TryParseObjectFromText(string input, out Dictionary obj, out string jsonFragment) { @@ -281,6 +359,10 @@ namespace WulaFallenEmpire.EventSystem.AI.Utils { return input?.IndexOf(value ?? "", StringComparison.OrdinalIgnoreCase) ?? -1; } + private static int IndexOfIgnoreCase(string input, string value, int startIndex) + { + return input?.IndexOf(value ?? "", startIndex, StringComparison.OrdinalIgnoreCase) ?? -1; + } public static bool TryParseObject(string json, out Dictionary obj) { diff --git a/Source/WulaFallenEmpire/WulaAI_DevDocs/qwen_agent/QWEN_AGENT_3PHASE_PROTOCOL.md b/Source/WulaFallenEmpire/WulaAI_DevDocs/qwen_agent/QWEN_AGENT_3PHASE_PROTOCOL.md new file mode 100644 index 00000000..c3783e5f --- /dev/null +++ b/Source/WulaFallenEmpire/WulaAI_DevDocs/qwen_agent/QWEN_AGENT_3PHASE_PROTOCOL.md @@ -0,0 +1,80 @@ +Qwen-Agent 3-Phase Tool Calling Protocol (Design + Implementation Plan) + +Goal +- Preserve the existing three-phase workflow: Query -> Action -> Reply. +- Provide a strict, parseable tool-call format compatible with Qwen-Agent style. +- Allow explicit switching between native tool calling and template parsing. + +Scope +- Query phase: only query tools (get_*/search_*/analyze_*/recall_memories) +- Action phase: only action tools (spawn_resources, send_reinforcement, call_bombardment, modify_goodwill, call_prefab_airdrop, set_overwatch_mode, remember_fact) +- Reply phase: natural language only, no tool calls + +Tool Calling Formats +Native Mode (useNativeToolApi = true) +- Use provider-native tool calls (OpenAI/DeepSeek tool_calls, Gemini functionCall) +- No text template parsing, only validation + execution + +Template Mode (useNativeToolApi = false) +- Use Qwen-Agent style tool-call tags in text: + + {"name":"tool_name","arguments":{...}} + +- Tool results are injected as tool-role messages (not visible to the user) +- No natural language is allowed in tool phases +- If no tool is needed: output exactly {"tool_calls":[]} + +Three-Phase Flow (Template Mode) +Phase 1: Query +- Prompt includes tool list and Qwen tool-call template +- Model returns one or more blocks +- Parser extracts tool calls -> validator -> tool execution +- Tool results stored and appended to history +- One retry max if tool parsing/validation fails + +Phase 2: Action +- Same as Query, but only action tools allowed +- One retry max if tool parsing/validation fails + +Phase 3: Reply +- Natural language only +- If tool-call tags or JSON are detected, one retry is triggered + +Validation and Guardrails +- ToolCallValidator ensures: + - Tool exists + - Required parameters present + - Tool is valid for current phase +- Failures are returned to the model as tool results for self-correction +- No parameter guessing or defName guessing + +Implementation Checklist (Files) +1) WulaFallenEmpireSettings.cs + - Add useNativeToolApi flag (default true for backward compatibility) +2) WulaFallenEmpireMod.cs + - Expose a UI checkbox to toggle useNativeToolApi +3) AIIntelligenceCore.cs + - Select native loop vs template loop based on useNativeToolApi + - Template loop uses Qwen tool-call instructions in Query/Action + - Reply phase unchanged +4) JsonToolCallParser.cs + - Add parsing for ... + - Convert parsed calls to tool_calls list for execution +5) ToolCallValidator.cs + - Enforce phase/tool rules and required args + +Rollout Plan +1) Implement parser + settings toggle +2) Switch flow selection by useNativeToolApi +3) Add template tool-call instructions +4) Validate in-game with: + - Item request (Query -> search_thing_def) + - Action call (spawn_resources) + - Reply text only + +Testing Checklist +- Query phase outputs blocks +- Action phase outputs blocks +- Reply phase never outputs tool calls +- Tool results are linked to tool_call_id (native) or injected safely (template) +- One retry only per phase when parsing fails diff --git a/Source/WulaFallenEmpire/WulaFallenEmpireMod.cs b/Source/WulaFallenEmpire/WulaFallenEmpireMod.cs index 52971bfd..2f3adf1f 100644 --- a/Source/WulaFallenEmpire/WulaFallenEmpireMod.cs +++ b/Source/WulaFallenEmpire/WulaFallenEmpireMod.cs @@ -44,6 +44,8 @@ namespace WulaFallenEmpire bool currentIsGemini = settings.useGeminiProtocol; if (listingStandard.RadioButton("OpenAI / 常用兼容格式 (DeepSeek, ChatGPT)", !currentIsGemini)) settings.useGeminiProtocol = false; if (listingStandard.RadioButton("Google Gemini 原生格式 (支持本地多模态)", currentIsGemini)) settings.useGeminiProtocol = true; + listingStandard.CheckboxLabeled("使用原生 tool_calls API)", ref settings.useNativeToolApi, + "对于训练过工具调用能力的全量大模型推荐开启，如果使用较弱的小模型或者无法调用工具可以尝试关闭该选项以使用提示词兼容工具调用模式。对于Deepseek,ChatGPT,Gemini,GLM,Kimi2,Qwen模型都推荐开启，其他模型未测试。"); listingStandard.GapLine(); // 根据当前选中的协议，动态绑定输入字段 diff --git a/Source/WulaFallenEmpire/WulaFallenEmpireSettings.cs b/Source/WulaFallenEmpire/WulaFallenEmpireSettings.cs index 8ef12868..3fd87cd2 100644 --- a/Source/WulaFallenEmpire/WulaFallenEmpireSettings.cs +++ b/Source/WulaFallenEmpire/WulaFallenEmpireSettings.cs @@ -14,6 +14,7 @@ namespace WulaFallenEmpire public string geminiModel = "gemini-2.5-flash"; public bool useGeminiProtocol = false; // 是否使用 Google Gemini 协议格式 + public bool useNativeToolApi = true; public int maxContextTokens = 100000; public bool enableDebugLogs = false; @@ -39,6 +40,7 @@ namespace WulaFallenEmpire Scribe_Values.Look(ref geminiModel, "geminiModel", "gemini-2.5-flash"); Scribe_Values.Look(ref useGeminiProtocol, "useGeminiProtocol", false); + Scribe_Values.Look(ref useNativeToolApi, "useNativeToolApi", true); Scribe_Values.Look(ref maxContextTokens, "maxContextTokens", 100000); Scribe_Values.Look(ref enableDebugLogs, "enableDebugLogs", false);