foliko 1.0.80 → 1.0.81

package/src/core/agent-chat.js

@@ -5,7 +5,7 @@
 
 const { EventEmitter } = require('../utils/event-emitter');
 const { logger } = require('../utils/logger');
-const { generateText, pruneMessages } = require('ai');
+const { generateText, pruneMessages, stepCountIs, convertToModelMessages } = require('ai');
 
 // 模型上下文限制表（留 15-20% 余量给 system prompt 和输出）
 const MODEL_CONTEXT_LIMITS = {
@@ -97,7 +97,7 @@ class AgentChatHandler extends EventEmitter {
     this._systemPrompt = config.systemPrompt || 'You are a helpful assistant.';
     this._messages = [];
     this._tools = new Map();
-    this._maxSteps = 5; // 降低默认步骤数，减少上下文消耗
+    this._maxSteps = 20; // 降低默认步骤数，减少上下文消耗
 
     // 上下文压缩配置：根据模型自动设置限制
     const modelKey = Object.keys(MODEL_CONTEXT_LIMITS).find((k) =>
@@ -163,7 +163,7 @@ class AgentChatHandler extends EventEmitter {
       const existingCount = existingHistory.length;
 
       // 只添加新消息（避免重复添加已加载的历史）
-      const simpleMessages = messages.filter((m) => m.role === 'user' || m.role === 'assistant');
+      const simpleMessages = messages;
 
       const newMessages = simpleMessages.slice(existingCount);
       for (const msg of newMessages) {
@@ -366,13 +366,14 @@ class AgentChatHandler extends EventEmitter {
     }
 
     const summary = {
-      role: 'system',
+      role: 'assistant',
       content: summaryContent,
     };
 
-    this._messages = [...systemMessages, summary, ...recentMessages];
+    this._messages = [...systemMessages, summary, ...recentMessages].filter(
+      (item) => item.role !== 'tool'
+    );
     this._compressionCount++;
-
     const totalTokens = this._countMessagesTokens(this._messages);
     logger.info(
       `Context compressed (${this._compressionCount} times). Messages: ${this._messages.length}, Est. tokens: ${totalTokens}`
@@ -639,80 +640,77 @@ ${truncatedContent}${truncatedNote}
     const context = { sessionId, isStream: false };
     const framework = this.agent.framework;
     const self = this; // 保存引用用于回调
-
-    // 从 session 加载聊天历史
-    if (sessionId) {
-      const savedMessages = this._loadHistoryFromSession(sessionId);
-      if (savedMessages.length > 0) {
-        this._messages = savedMessages;
-        logger.info(`Loaded ${savedMessages.length} messages from session ${sessionId}`);
+    try {
+      // 从 session 加载聊天历史
+      if (sessionId) {
+        const savedMessages = this._loadHistoryFromSession(sessionId);
+        if (savedMessages.length > 0) {
+          this._messages = savedMessages;
+          logger.info(`Loaded ${savedMessages.length} messages from session ${sessionId}`);
+        }
       }
-    }
-
-    // 关键：每次 chat 调用时刷新系统提示词，确保包含最新的工具/技能描述
-    // 解决上下文过长时 LLM 不调用工具的问题
-    this._systemPrompt = this.agent._buildSystemPrompt();
-    // 动态导入 AI SDK
-    const { tool, ToolLoopAgent } = await this._importAI();
-
-    const userMessage = typeof message === 'string' ? { role: 'user', content: message } : message;
-
-    this._messages.push(userMessage);
-
-    // 检查是否需要压缩上下文（包括工具定义）
-    const messagesTokens = this._countMessagesTokens(this._messages);
-    const toolsTokens = this._countToolsTokens();
-    const systemPromptTokens = this._countTokens(this._systemPrompt);
-    const totalTokens = messagesTokens + toolsTokens + systemPromptTokens;
-    const limit = this._maxContextTokens * 0.7; // 降低到 70%，更早压缩
-
-    if (totalTokens > limit) {
-      logger.info(
-        `Context large (${totalTokens}/${this._maxContextTokens} tokens = msgs:${messagesTokens} + tools:${toolsTokens} + sys:${systemPromptTokens}), compressing...`
-      );
-      // 使用带超时控制的压缩
-      await this._compressContext();
-    } else {
-      logger.info(`Context OK: ${totalTokens}/${this._maxContextTokens} tokens`);
-    }
 
-    const maxSteps = options.maxSteps || this._maxSteps;
-    const tools = this._getAITools(tool);
+      // 关键：每次 chat 调用时刷新系统提示词，确保包含最新的工具/技能描述
+      // 解决上下文过长时 LLM 不调用工具的问题
+      this._systemPrompt = this.agent._buildSystemPrompt();
+      // 动态导入 AI SDK
+      const { tool, ToolLoopAgent } = await this._importAI();
+
+      const userMessage =
+        typeof message === 'string' ? { role: 'user', content: message } : message;
+      //console.log('System Prompt:', this._systemPrompt);
+      this._messages.push(userMessage);
+
+      // 检查是否需要压缩上下文（包括工具定义）
+      const messagesTokens = this._countMessagesTokens(this._messages);
+      const toolsTokens = this._countToolsTokens();
+      const systemPromptTokens = this._countTokens(this._systemPrompt);
+      const totalTokens = messagesTokens + toolsTokens + systemPromptTokens;
+      const limit = this._maxContextTokens * 0.7; // 降低到 70%，更早压缩
+
+      if (totalTokens > limit) {
+        logger.info(
+          `Context large (${totalTokens}/${this._maxContextTokens} tokens = msgs:${messagesTokens} + tools:${toolsTokens} + sys:${systemPromptTokens}), compressing...`
+        );
+        // 使用带超时控制的压缩
+        await this._compressContext();
+      } else {
+        logger.info(`Context OK: ${totalTokens}/${this._maxContextTokens} tokens`);
+      }
 
-    if (!this._aiClient) {
-      throw new Error('AI client not configured.');
-    }
+      const maxSteps = options.maxSteps || this._maxSteps;
+      const tools = this._getAITools(tool);
 
-    const agent = new ToolLoopAgent({
-      model: this._aiClient,
-      instructions: this._systemPrompt,
-      tools: tools,
-      stopWhen: (step) => step.stepCount >= maxSteps,
-    });
+      if (!this._aiClient) {
+        throw new Error('AI client not configured.');
+      }
+      // const prunedMessages = pruneMessages({
+      //   messages: this._messages,
+      //   reasoning: 'all', // Remove all reasoning parts
+      //   toolCalls: 'before-last-6-messages', // Remove tool calls except those in the last message
+      // });
+      const agent = new ToolLoopAgent({
+        model: this._aiClient,
+        instructions: this._systemPrompt,
+        tools: tools,
+        stopWhen: stepCountIs(maxSteps),
+        prepareStep: async ({ stepNumber, steps, messages }) => {
+          if (messages.length > 20) {
+            return {
+              messages: messages.slice(-10),
+            };
+          }
 
-    const messages = this._cleanMessages(this._messages);
-    const prunedMessages = pruneMessages({
-      messages,
-      reasoning: 'all', // Remove all reasoning parts
-      toolCalls: 'before-last-2-messages', // Remove tool calls except those in the last message
-    });
+          return {};
+        },
+      });
 
-    try {
       // 使用 runWithContext 让工具执行时能获取 sessionId
       const result = await framework.runWithContext(context, async () => {
-        return agent.generate({ messages: prunedMessages, ...this.providerOptions });
+        return agent.generate({ messages: this._messages, ...this.providerOptions });
       });
 
-      // 处理返回的消息历史：只保留 user 和 assistant 消息，让 SDK 自动处理 tool 消息
-      if (result.messages && Array.isArray(result.messages)) {
-        // 只保留 user 和 assistant 消息，SDK 会自动维护 tool 消息
-        this._messages = result.messages.filter((m) => m.role === 'user' || m.role === 'assistant');
-      } else if (result.text) {
-        this._messages.push({
-          role: 'assistant',
-          content: result.text,
-        });
-      }
+      this._messages.push(...result.response.messages);
 
       // 生成后检查：如果消息太长，下次需要压缩
       const afterTokens = this._countMessagesTokens(this._messages);
@@ -746,66 +744,77 @@ ${truncatedContent}${truncatedNote}
     const sessionId = options.sessionId || null;
     const context = { sessionId, isStream: true };
     const framework = this.agent.framework;
-
-    // 从 session 加载聊天历史
-    if (sessionId) {
-      const savedMessages = this._loadHistoryFromSession(sessionId);
-      if (savedMessages.length > 0) {
-        this._messages = savedMessages;
-        logger.info(`Loaded ${savedMessages.length} messages from session ${sessionId}`);
+    try {
+      //从 session 加载聊天历史
+      if (sessionId) {
+        const savedMessages = this._loadHistoryFromSession(sessionId);
+        if (savedMessages.length > 0) {
+          this._messages = savedMessages;
+          logger.info(`Loaded ${savedMessages.length} messages from session ${sessionId}`);
+        }
       }
-    }
-
-    // 关键：每次 chat 调用时刷新系统提示词，确保包含最新的工具/技能描述
-    // 解决上下文过长时 LLM 不调用工具的问题
-    this._systemPrompt = this.agent._buildSystemPrompt();
-    // 动态导入 AI SDK
-    const { tool, ToolLoopAgent } = await this._importAI();
-
-    const userMessage = typeof message === 'string' ? { role: 'user', content: message } : message;
-    //console.log('System Prompt:', this._systemPrompt);
-    this._messages.push(userMessage);
-    // 检查是否需要压缩上下文（包括工具定义）
-    const messagesTokens = this._countMessagesTokens(this._messages);
-    const toolsTokens = this._countToolsTokens();
-    const systemPromptTokens = this._countTokens(this._systemPrompt);
-    const totalTokens = messagesTokens + toolsTokens + systemPromptTokens;
-    const limit = this._maxContextTokens * 0.7; // 降低到 70%
-
-    // 对于流式调用，如果上下文太大，先压缩再开始
-    if (totalTokens > limit) {
-      logger.info(
-        `Context large (${totalTokens}/${this._maxContextTokens} tokens), compressing...`
-      );
-      // 流式调用时等待压缩完成（使用带超时控制的压缩）
-      await this._compressContext();
-    }
 
-    const maxSteps = options.maxSteps || this._maxSteps;
-    const tools = this._getAITools(tool);
-    const self = this; // 保存引用用于回调
+      // 关键：每次 chat 调用时刷新系统提示词，确保包含最新的工具/技能描述
+      // 解决上下文过长时 LLM 不调用工具的问题
+      this._systemPrompt = this.agent._buildSystemPrompt();
+      // 动态导入 AI SDK
+      const { tool, ToolLoopAgent } = await this._importAI();
+
+      const userMessage =
+        typeof message === 'string' ? { role: 'user', content: message } : message;
+      //console.log('System Prompt:', this._systemPrompt);
+      this._messages.push(userMessage);
+
+      // 检查是否需要压缩上下文（包括工具定义）
+      const messagesTokens = this._countMessagesTokens(this._messages);
+      const toolsTokens = this._countToolsTokens();
+      const systemPromptTokens = this._countTokens(this._systemPrompt);
+      const totalTokens = messagesTokens + toolsTokens + systemPromptTokens;
+      const limit = this._maxContextTokens * 0.7; // 降低到 70%
+      // 对于流式调用，如果上下文太大，先压缩再开始
+      if (totalTokens > limit) {
+        logger.info(
+          `Context large (${totalTokens}/${this._maxContextTokens} tokens), compressing...`
+        );
+        // 流式调用时等待压缩完成（使用带超时控制的压缩）
+        await this._compressContext();
+      }
 
-    if (!this._aiClient) {
-      throw new Error('AI client not configured.');
-    }
+      const maxSteps = options.maxSteps || this._maxSteps;
+      const tools = this._getAITools(tool);
+      const self = this; // 保存引用用于回调
+      if (!this._aiClient) {
+        throw new Error('AI client not configured.');
+      }
+      console.log(JSON.stringify(this._messages, null, 2));
+      // const prunedMessages = pruneMessages({
+      //   messages: this._messages,
+      //   reasoning: 'all', // Remove all reasoning parts
+      //   toolCalls: 'before-last-6-messages', // Remove tool calls except those in the last message
+      // });
+
+      const agent = new ToolLoopAgent({
+        model: this._aiClient,
+        instructions: this._systemPrompt,
+        tools: tools,
+        stopWhen: stepCountIs(maxSteps),
+        prepareStep: async ({ stepNumber, steps, messages }) => {
+          if (messages.length > 20) {
+            return {
+              messages: messages.slice(-10),
+            };
+          }
 
-    const agent = new ToolLoopAgent({
-      model: this._aiClient,
-      instructions: this._systemPrompt,
-      tools: tools,
-      stopWhen: (step) => step.stepCount >= maxSteps,
-    });
+          return {};
+        },
+      });
 
-    const messages = this._cleanMessages(this._messages);
-    const prunedMessages = pruneMessages({
-      messages,
-      reasoning: 'all', // Remove all reasoning parts
-      toolCalls: 'before-last-2-messages', // Remove tool calls except those in the last message
-    });
-    try {
       // 使用 runWithContext 让工具执行时能获取 sessionId（支持并行）
       const result = await framework.runWithContext(context, async () => {
-        return agent.stream({ messages: prunedMessages, ...this.providerOptions });
+        return agent.stream({
+          messages: this._messages,
+          ...this.providerOptions,
+        });
       });
 
       const stream = result.fullStream;
@@ -813,7 +822,6 @@ ${truncatedContent}${truncatedNote}
 
       // 流式迭代器
       const iterator = stream[Symbol.asyncIterator] ? stream : stream.fullStream;
-      const finalMessages = [];
 
       for await (const part of iterator || stream) {
         if (part.type === 'text-delta') {
@@ -821,25 +829,24 @@ ${truncatedContent}${truncatedNote}
           fullText += text;
           yield { type: 'text', text };
         } else if (part.type === 'reasoning') {
-          finalMessages.push(part);
           this.emit('thinking', { content: part.text });
         } else if (part.type === 'tool-call') {
-          finalMessages.push(part);
-          yield { type: 'tool-call', toolName: part.toolName, args: part.input };
+          yield { type: 'tool-call', toolName: part.toolName, input: part.input };
         } else if (part.type === 'tool-result') {
-          // AI SDK 6.x 要求 tool 消息格式为:
-          // { role: 'tool', content: [{ type: 'tool-result', toolCallId, toolName, output }] }
-          finalMessages.push(part); // 先保存到 finalMessages，等生成结束后统一添加到历史
           yield { type: 'tool-result', toolName: part.toolName, result: part.output };
         } else if (part.type === 'error') {
           yield { type: 'error', error: part.error };
         }
       }
 
-      this._messages.push({ role: 'tool', content: finalMessages });
-      // 添加 assistant 消息
-      const assistantMsg = { role: 'assistant', content: fullText };
-      this._messages.push(assistantMsg);
+      const finish_messages = (await result.response).messages;
+      this._messages.push(...finish_messages);
+      // console.log(JSON.stringify(this._messages, null, 2));
+      // 生成后检查：如果消息太长，下次需要压缩
+      const afterTokens = this._countMessagesTokens(this._messages);
+      if (afterTokens > this._maxContextTokens * 0.8) {
+        logger.info(`After generation: ${afterTokens} tokens, will compress on next turn`);
+      }
 
       // 保存聊天历史到 session
       if (sessionId) {