@myassis/gateway 1.0.108 → 1.0.110

package/dist/config/index.js

@@ -146,7 +146,7 @@ exports.appConfig = {
      * 必定压缩，与内容多少无关 —— 这是「没聊几句就压缩」最直接的原因。
      * 条数只作为体积估算失真时的兜底，阈值应远高于正常对话轮数。
      */
-    summaryThreshold: parseInt(process.env.SUMMARY_THRESHOLD || '60', 10),
+    summaryThreshold: parseInt(process.env.SUMMARY_THRESHOLD || '10', 10),
     appName: '我的助手'
 };
 /**

package/dist/services/llm/LLMClient.js

@@ -12,7 +12,7 @@ const logger = (0, shared_1.getLogger)('LLMClient');
 /** 历史工具输出在上下文中的最大保留长度 */
 const HISTORY_TOOL_OUTPUT_LIMIT = index_js_2.appConfig.oldToolContentLimit;
 /** 请求体字符硬上限，超过则视为上下文超限 */
-const HARD_MAX_CHARS = 200000;
+const HARD_MAX_CHARS = index_js_2.appConfig.contextMaxChars;
 /**
  * 规范化工具调用参数。
  *
@@ -371,30 +371,16 @@ class LLMClient {
      */
     async buildRequest(stream, model) {
         // 系统模式：转发到 Server LLM 代理
-        if (this.systemMode) {
-            const { mapped } = this.mapMessages(this.messages);
-            return {
-                url: `${this.serverUrl}/api/v1/llm/${stream ? 'chat' : 'invoke'}`,
-                headers: {
-                    'Content-Type': 'application/json',
-                    'Authorization': `Bearer ${this.authToken}`,
-                },
-                body: {
-                    modelId: model.modelId,
-                    messages: mapped,
-                    temperature: 0.7,
-                    tools: this.tools,
-                },
-                isSystemMode: true,
-            };
-        }
         // 自定义模式：直接调用 LLM API
         if (!model) {
             throw new Error('No available model');
         }
-        const apiKey = index_js_1.persistStore.getModelApiKey(model.id);
-        if (!apiKey) {
-            throw new Error('No API Key');
+        var apiKey;
+        if (!this.systemMode) {
+            apiKey = index_js_1.persistStore.getModelApiKey(model.id);
+            if (!apiKey) {
+                throw new Error('No API Key');
+            }
         }
         // 使用统一的消息映射
         const { mapped: mappedMessages, existAttachments, existScreenShot } = this.mapMessages(this.messages);
@@ -410,17 +396,17 @@ class LLMClient {
         const bodyChars = JSON.stringify(body).length;
         logger.debug('context chars', bodyChars);
         // 附件/截图场景 base64 体积大但不可裁剪，跳过硬上限检查
-        if (bodyChars > HARD_MAX_CHARS && !existAttachments && !existScreenShot) {
-            throw Error('exceed max message tokens');
-        }
+        // if (bodyChars > HARD_MAX_CHARS && !existAttachments && !existScreenShot) {
+        //   throw Error('exceed max message tokens')
+        // }
         return {
-            url: `${model.baseUrl}/chat/completions`,
+            url: this.systemMode ? `${this.serverUrl}/api/v1/llm/${stream ? 'chat' : 'invoke'}` : `${model.baseUrl}/chat/completions`,
             headers: {
                 'Content-Type': 'application/json',
-                'Authorization': `Bearer ${apiKey}`,
+                'Authorization': `Bearer ${this.systemMode ? this.authToken : apiKey}`,
             },
             body,
-            isSystemMode: false,
+            isSystemMode: this.systemMode,
         };
     }
     /**

package/dist/services/memory/MemoryManager.js

@@ -37,6 +37,10 @@ const DEFAULT_CONFIG = {
     summaryTriggerChars: index_js_2.appConfig.summaryTriggerChars,
     enabled: true,
 };
+/** 单次摘要模型请求超时（毫秒） */
+const SUMMARY_REQUEST_TIMEOUT = 45000;
+/** 整个压缩流程的硬超时（毫秒），超过后放弃压缩、沿用原始历史 */
+const SUMMARY_OVERALL_TIMEOUT = 90000;
 /**
  * 会话级压缩任务表：sessionId -> 进行中的压缩任务。
  *
@@ -298,6 +302,7 @@ class MemoryManager {
             compressionTasks.delete(this.session.id);
         }
         if (!summary) {
+            this.emitSSE({ type: 'context_compress_done' });
             // 摘要失败/被中断：沿用旧摘要或退回原始历史，不写入残缺摘要
             if (plan.action === 'incremental') {
                 return [
@@ -307,6 +312,7 @@ class MemoryManager {
             }
             return plan.prevRetained;
         }
+        this.emitSSE({ type: 'context_compress_done' });
         return [
             this.toSummaryMessage(summary, plan.newBoundaryAt, '', plan.ledger),
             ...plan.retained,
@@ -351,7 +357,7 @@ class MemoryManager {
         const prompt = TURN_COMPACT_PROMPT_HEAD + transcript;
         try {
             const models = await this.getSummaryModels();
-            const llmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: prompt }], this.signal, [], 600000);
+            const llmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: prompt }], this.signal, [], SUMMARY_REQUEST_TIMEOUT);
             if (this.session.useSystemMode) {
                 const { getRequestToken, getServerBaseUrl } = await Promise.resolve().then(() => __importStar(require('../../api/index.js')));
                 llmClient.setSystemMode(true, this.session.selectModelId, getServerBaseUrl(), getRequestToken());
@@ -554,6 +560,26 @@ ${conversation}
         if (!this.childAgent) {
             this.emitSSE({ type: 'context_compressing' });
         }
+        try {
+            return await Promise.race([
+                this.doGenerateSummary(messages, lastSummary),
+                this.summaryTimeout(),
+            ]);
+        }
+        catch (error) {
+            logger.warn(`压缩流程异常，保留原始历史: ${error?.message}`);
+            return lastSummary || null;
+        }
+    }
+    summaryTimeout() {
+        return new Promise((resolve) => {
+            setTimeout(() => {
+                logger.warn(`压缩流程超时（${SUMMARY_OVERALL_TIMEOUT}ms），放弃压缩`);
+                resolve(null);
+            }, SUMMARY_OVERALL_TIMEOUT);
+        });
+    }
+    async doGenerateSummary(messages, lastSummary) {
         const MAX_INPUT_CHARS = 30000; // 单次摘要最大输入字符数
         // 格式化所有消息用于估算长度
         const formattedMessages = messages
@@ -565,7 +591,7 @@ ${conversation}
             return await this.hierarchicalSummary(formattedMessages, lastSummary);
         }
         const summaryPrompt = this.buildSummaryPrompt(messages, lastSummary);
-        const llmClient = new LLMClient_js_1.LLMClient(await this.getSummaryModels(), [{ role: 'user', content: summaryPrompt }], this.signal, [], 600000);
+        const llmClient = new LLMClient_js_1.LLMClient(await this.getSummaryModels(), [{ role: 'user', content: summaryPrompt }], this.signal, [], SUMMARY_REQUEST_TIMEOUT);
         if (this.session.useSystemMode) {
             const { getRequestToken, getServerBaseUrl } = await Promise.resolve().then(() => __importStar(require('../../api/index.js')));
             llmClient.setSystemMode(true, this.session.selectModelId, getServerBaseUrl(), getRequestToken());
@@ -598,7 +624,7 @@ ${conversation}
         const models = await this.getSummaryModels();
         const summarizeBatch = async (batch) => {
             const batchPrompt = `请简洁总结以下对话片段的关键信息，特别关注：项目结构、关键决策、错误及修复。\n\n${batch.join('\n\n')}`;
-            const llmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: batchPrompt }], this.signal, [], 600000);
+            const llmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: batchPrompt }], this.signal, [], SUMMARY_REQUEST_TIMEOUT);
             if (this.session.useSystemMode) {
                 const { getRequestToken, getServerBaseUrl } = await Promise.resolve().then(() => __importStar(require('../../api/index.js')));
                 llmClient.setSystemMode(true, this.session.selectModelId, getServerBaseUrl(), getRequestToken());
@@ -609,6 +635,7 @@ ${conversation}
         // 保持批次顺序，同时限制并发
         const subSummaries = new Array(batches.length).fill('');
         let aborted = false;
+        let consecutiveFailures = 0;
         for (let i = 0; i < batches.length; i += CONCURRENCY) {
             if (this.signal.aborted) {
                 aborted = true;
@@ -616,14 +643,26 @@ ${conversation}
             }
             const slice = batches.slice(i, i + CONCURRENCY);
             const results = await Promise.allSettled(slice.map((batch) => summarizeBatch(batch)));
+            let groupSuccess = 0;
             results.forEach((result, offset) => {
                 if (result.status === 'fulfilled' && result.value) {
                     subSummaries[i + offset] = result.value;
+                    groupSuccess++;
                 }
                 else if (result.status === 'rejected') {
                     logger.warn(`子摘要生成失败: ${result.reason?.message}`);
                 }
             });
+            if (groupSuccess === 0) {
+                consecutiveFailures++;
+                if (consecutiveFailures >= 2) {
+                    logger.warn(`分层摘要连续失败，提前终止`);
+                    break;
+                }
+            }
+            else {
+                consecutiveFailures = 0;
+            }
         }
         // 被中断时不写入残缺摘要，交由调用方回退到原始历史
         if (aborted || this.signal.aborted) {
@@ -638,7 +677,7 @@ ${conversation}
         }
         // 合并子摘要为最终摘要
         const mergePrompt = this.buildMergePrompt(valid, lastSummary);
-        const mergeLlmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: mergePrompt }], this.signal, [], 600000);
+        const mergeLlmClient = new LLMClient_js_1.LLMClient(models, [{ role: 'user', content: mergePrompt }], this.signal, [], SUMMARY_REQUEST_TIMEOUT);
         if (this.session.useSystemMode) {
             const { getRequestToken, getServerBaseUrl } = await Promise.resolve().then(() => __importStar(require('../../api/index.js')));
             mergeLlmClient.setSystemMode(true, this.session.selectModelId, getServerBaseUrl(), getRequestToken());

package/dist/services/session/Session.js

@@ -1031,6 +1031,7 @@ class Session {
                 if (built.trimmed) {
                     logger.debug(`上下文裁剪: level=${built.level} chars=${built.chars}`);
                 }
+                logger.debug(`调用模型${this.selectModelId},messages:`, built.messages);
                 // 收尾阶段不再下发工具定义：只要工具还在，模型大概率继续调用而不收尾
                 const llmClient = new LLMClient_js_1.LLMClient(models, built.messages, this.abortController.signal, forceFinalize ? [] : tools);
                 llmClient.setPreferredModel(this.selectModelId);
@@ -1370,12 +1371,6 @@ class Session {
                 };
                 await Promise.all(llmResult.toolCalls.map(x => startToolCall(x)));
                 toolCalls.push(toolCall);
-                // 重置 toolCall：下一轮工具调用必须开启新的分组。
-                // 否则模型连续多轮只返回工具调用（无正文）时，所有轮次的结果会
-                // 累积进同一个 toolCall 对象，并被重复 push 到 toolCalls 数组，
-                // 导致持久化消息出现重复的工具调用组，历史回放时模型看到重复
-                // 的 tool_calls/tool 结果对，进而重复执行工具。
-                toolCall = null;
                 // 本轮工具已执行完，增量落库一次。工具轮次可能很多且每轮都耗时，
                 // 这里保存能让崩溃后仍保留已完成的工具调用记录。
                 if (!childAgent) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@myassis/gateway",
-  "version": "1.0.108",
+  "version": "1.0.110",
   "description": "我的助手 Gateway Service - 本地 AI 网关服务，支持认证、WebSocket 实时通信和任务调度",
   "main": "dist/index.js",
   "bin": {