aws-runtime-bridge 1.9.78 → 1.9.80

package/package/aws-client-agent-mcp/dist/context-manager.js

@@ -1,477 +1,2 @@
-/**
- * 上下文管理器 — 会话级 Token 追踪、溢出检测与压缩
- *
- * 核心职责：
- * 1. 追踪每轮会话的累计 token 消耗
- * 2. 根据模型上下文窗口判断是否溢出
- * 3. 提供 prune（裁剪旧工具输出）/ summarize（摘要化）两种压缩策略
- * 4. 通过 MemoryStore(session_context) 持久化压缩后的摘要
- *
- * 设计思路（参考 OpenCode compaction.ts）：
- * - 保持 20K token 作为缓冲区，避免溢出前措手不及
- * - Prune 从最旧消息开始累计工具输出，超过保护线后标记 compacted
- * - Summarize 将早期对话结构化为摘要存入 session_context memory
- *
- * 关于模型上下文窗口大小：
- * - 当前版本通过模型名查找表获取（见 MODEL_CONTEXT_LIMITS）
- * - 查找不到时默认使用 128K（保守值）
- * - 后续可通过服务端 AiSourceTemplate 扩展 contextLimit/maxOutput 字段
- */
-import { logger } from "./logger.js";
-// ============ 常量 ============
-/**
- * 压缩缓冲区：保留这么多 token 的空闲空间，避免压缩过程中溢出。
- */
-const COMPACTION_BUFFER = 20_000;
-/**
- * Prune 保护线：从最新往旧遍历时，保留最近 PRUNE_PROTECT token 的工具输出完整。
- */
-const PRUNE_PROTECT = 40_000;
-/**
- * Prune 最小收益：累计可释放 token 不足此数时跳过修剪。
- */
-const PRUNE_MINIMUM = 20_000;
-/**
- * 默认保留的完整对话轮次（不被压缩）。
- */
-const DEFAULT_KEEP_TURNS = 3;
-/**
- * 默认上下文窗口上限（查找不到模型时使用）。保守取 128K。
- */
-const DEFAULT_CONTEXT_LIMIT = 128_000;
-/**
- * 默认最大输出 token。
- */
-const DEFAULT_MAX_OUTPUT = 8_192;
-/**
- * 知名模型的上下文窗口大小查找表。
- * 覆盖主流模型。未在此表中的模型回退到 DEFAULT_CONTEXT_LIMIT。
- * 数据来源：各模型官方文档。
- */
-const MODEL_CONTEXT_LIMITS = {
-    // Anthropic Claude
-    "claude-4": { contextLimit: 200_000, maxOutput: 8_192 },
-    "claude-sonnet-4": { contextLimit: 200_000, maxOutput: 8_192 },
-    "claude-3-5-sonnet": { contextLimit: 200_000, maxOutput: 8_192 },
-    "claude-3-opus": { contextLimit: 200_000, maxOutput: 8_192 },
-    "claude-3-haiku": { contextLimit: 200_000, maxOutput: 8_192 },
-    "claude-2": { contextLimit: 100_000, maxOutput: 4_096 },
-    // OpenAI GPT
-    "gpt-4o": { contextLimit: 128_000, maxOutput: 16_384 },
-    "gpt-4-turbo": { contextLimit: 128_000, maxOutput: 4_096 },
-    "gpt-4": { contextLimit: 8_192, maxOutput: 4_096 },
-    "gpt-3.5-turbo": { contextLimit: 16_384, maxOutput: 4_096 },
-    "o1": { contextLimit: 200_000, maxOutput: 100_000 },
-    "o3-mini": { contextLimit: 200_000, maxOutput: 100_000 },
-    // Google Gemini
-    "gemini-2.0-flash": { contextLimit: 1_048_576, maxOutput: 8_192 },
-    "gemini-2.5-pro": { contextLimit: 1_048_576, maxOutput: 8_192 },
-    "gemini-1.5-pro": { contextLimit: 2_097_152, maxOutput: 8_192 },
-    "gemini-1.5-flash": { contextLimit: 1_048_576, maxOutput: 8_192 },
-    // DeepSeek
-    "deepseek": { contextLimit: 64_000, maxOutput: 8_192 },
-    "deepseek-chat": { contextLimit: 64_000, maxOutput: 8_192 },
-    "deepseek-reasoner": { contextLimit: 64_000, maxOutput: 8_192 },
-    // Qwen
-    "qwen-max": { contextLimit: 32_000, maxOutput: 8_192 },
-    "qwen-plus": { contextLimit: 128_000, maxOutput: 8_192 },
-    "qwen-turbo": { contextLimit: 1_000_000, maxOutput: 8_192 },
-    // GLM
-    "glm-4": { contextLimit: 128_000, maxOutput: 4_096 },
-    // Yi
-    "yi-": { contextLimit: 200_000, maxOutput: 4_096 },
-    // Meta Llama
-    "llama-3": { contextLimit: 8_192, maxOutput: 4_096 },
-    "llama-2": { contextLimit: 4_096, maxOutput: 4_096 },
-    // Mistral
-    "mistral": { contextLimit: 32_000, maxOutput: 4_096 },
-    "mixtral": { contextLimit: 32_000, maxOutput: 4_096 },
-    "codestral": { contextLimit: 256_000, maxOutput: 8_192 },
-    // Grok
-    "grok": { contextLimit: 131_072, maxOutput: 4_096 },
-    // Amazon Nova
-    "nova-": { contextLimit: 200_000, maxOutput: 8_192 },
-};
-/**
- * 根据模型名查找上下文窗口大小。
- * 主流程：遍历查找表前缀匹配 → 找到返回 → 未找到返回默认值。
- */
-function lookupModelLimits(modelName) {
-    if (!modelName)
-        return { contextLimit: DEFAULT_CONTEXT_LIMIT, maxOutput: DEFAULT_MAX_OUTPUT };
-    const key = modelName.toLowerCase().trim();
-    // 精确匹配优先
-    if (MODEL_CONTEXT_LIMITS[key])
-        return MODEL_CONTEXT_LIMITS[key];
-    // 前缀匹配：按 key 长度降序排序，先匹配更具体的
-    const matches = Object.keys(MODEL_CONTEXT_LIMITS)
-        .filter((prefix) => key.startsWith(prefix))
-        .sort((a, b) => b.length - a.length);
-    if (matches.length > 0)
-        return MODEL_CONTEXT_LIMITS[matches[0]];
-    return { contextLimit: DEFAULT_CONTEXT_LIMIT, maxOutput: DEFAULT_MAX_OUTPUT };
-}
-/**
- * Prune 白名单：这些工具的输出永远不会被裁剪。
- */
-const PRUNE_PROTECTED_TOOLS = new Set(["skill"]);
-// ============ 上下文管理器 ============
-export class ContextManager {
-    /** 内存态会话上下文记录，避免每次查询都读磁盘 */
-    sessions = new Map();
-    memoryStore;
-    constructor(memoryStore) {
-        this.memoryStore = memoryStore;
-    }
-    /**
-     * 根据模型名查找上下文窗口大小。
-     * 公开方法，供外部（如 mcp-server）使用。
-     */
-    lookupModelLimits(modelName) {
-        return lookupModelLimits(modelName);
-    }
-    // ============ Token 估算 ============
-    /**
-     * 粗略估算文本的 token 数。
-     * 不使用真实 tokenizer（太慢），基于字符数乘经验系数。
-     * 对纯 ASCII 约 4 字符/token，对中文约 1.5 字符/token。
-     *
-     * x1.15 安全填充确保本地估算始终偏大（安全侧），
-     * 与 compaction.ts 的 estimateTokens 策略保持一致。
-     */
-    estimateTokens(text) {
-        if (!text)
-            return 0;
-        // 区分中文字符（多字节字符）和 ASCII
-        let chineseCount = 0;
-        let asciiCount = 0;
-        for (const ch of text) {
-            if (ch > '\u00ff') {
-                chineseCount++;
-            }
-            else {
-                asciiCount++;
-            }
-        }
-        // 中文约 1.5 char/token，英文约 4 char/token
-        const base = Math.ceil(chineseCount / 1.5 + asciiCount / 4);
-        return Math.ceil(base * 1.15);
-    }
-    /**
-     * 从 TokenUsage 估算总 token 数。
-     */
-    estimateTotalUsage(tokens) {
-        return tokens.input + tokens.output + (tokens.reasoning ?? 0)
-            + (tokens.cache?.read ?? 0) + (tokens.cache?.write ?? 0);
-    }
-    // ============ 会话记录管理 ============
-    /**
-     * 获取或创建会话上下文记录。
-     */
-    async getOrCreateSession(sessionId) {
-        const existing = this.sessions.get(sessionId);
-        if (existing)
-            return existing;
-        // 尝试从 MemoryStore 加载
-        const loaded = await this.loadFromMemory(sessionId);
-        if (loaded) {
-            this.sessions.set(sessionId, loaded);
-            return loaded;
-        }
-        // 创建新记录
-        const record = {
-            sessionId,
-            tokens: { totalInput: 0, totalOutput: 0, totalReasoning: 0 },
-            lastActivity: new Date().toISOString(),
-            compactions: [],
-            overflowWarned: false,
-            messageCount: 0,
-        };
-        this.sessions.set(sessionId, record);
-        return record;
-    }
-    /**
-     * 记录一次 token 消耗。
-     */
-    async recordUsage(sessionId, tokens, modelInfo) {
-        const session = await this.getOrCreateSession(sessionId);
-        const total = this.estimateTotalUsage(tokens);
-        // 按比例估算 input/output，不精确但够用
-        // 如果没有明确拆分，按 in/out 各半估算
-        const inputTokens = tokens.input ?? Math.floor(total * 0.7);
-        const outputTokens = tokens.output ?? Math.floor(total * 0.3);
-        session.tokens.totalInput += inputTokens;
-        session.tokens.totalOutput += outputTokens;
-        session.tokens.totalReasoning += tokens.reasoning ?? 0;
-        session.lastActivity = new Date().toISOString();
-        session.messageCount++;
-        if (modelInfo) {
-            session.modelInfo = modelInfo;
-        }
-        // 异步持久化，不阻塞
-        this.saveToMemory(session).catch((err) => {
-            logger.error("[ContextManager] saveToMemory failed", err);
-        });
-    }
-    /**
-     * 判断当前会话是否上下文溢出。
-     * 逻辑参考 OpenCode SessionCompaction.isOverflow：
-     * - 已用 token >= (contextLimit - maxOutput - COMPACTION_BUFFER)
-     */
-    async isOverflow(sessionId, modelInfo) {
-        const session = await this.getOrCreateSession(sessionId);
-        const info = modelInfo ?? session.modelInfo;
-        if (!info || !info.contextLimit || info.contextLimit === 0)
-            return false;
-        const totalUsed = session.tokens.totalInput + session.tokens.totalOutput
-            + session.tokens.totalReasoning;
-        const maxOutput = info.maxOutput ?? 4096;
-        const reserved = Math.min(COMPACTION_BUFFER, maxOutput);
-        const usable = info.maxInput
-            ? info.maxInput - reserved
-            : info.contextLimit - maxOutput;
-        return totalUsed >= usable;
-    }
-    /**
-     * 获取当前会话的上下文统计。
-     */
-    async getStats(sessionId, modelInfo) {
-        const session = await this.getOrCreateSession(sessionId);
-        const info = modelInfo ?? session.modelInfo;
-        const totalUsed = session.tokens.totalInput + session.tokens.totalOutput
-            + session.tokens.totalReasoning;
-        const contextLimit = info?.contextLimit ?? 128_000;
-        const remaining = Math.max(0, contextLimit - totalUsed);
-        const usagePercent = contextLimit > 0
-            ? Math.round((totalUsed / contextLimit) * 100)
-            : 0;
-        return {
-            totalUsed,
-            totalInput: session.tokens.totalInput,
-            totalOutput: session.tokens.totalOutput,
-            contextLimit,
-            remaining,
-            usagePercent,
-            overflowWarning: await this.isOverflow(sessionId, info),
-            lastChecked: new Date().toISOString(),
-        };
-    }
-    // ============ Prune 策略 ============
-    /**
-     * Prune：从消息列表中裁剪旧工具的输出内容。
-     *
-     * 算法（参考 OpenCode compaction.ts:50-99）：
-     * 1. 从最新消息往旧遍历
-     * 2. 跳过前 keepLastTurns 轮对话
-     * 3. 遇到已有 summary 标记的助手消息就停止
-     * 4. 累计已完成工具调用的输出 token
-     * 5. 超过 PRUNE_PROTECT 保护线的标记为 compacted
-     * 6. 只有收益超过 PRUNE_MINIMUM 才真正执行
-     *
-     * @param messages 消息列表（从旧到新）
-     * @param keepLastTurns 保留最近的完整轮次
-     * @returns 修剪结果
-     */
-    prune(messages, keepLastTurns = DEFAULT_KEEP_TURNS) {
-        if (messages.length === 0) {
-            return {
-                strategy: "prune",
-                freedTokens: 0,
-                preservedTurns: 0,
-                compactedMessages: 0,
-                remainingTokens: 0,
-                timestamp: new Date().toISOString(),
-            };
-        }
-        let totalFreed = 0;
-        let compactedCount = 0;
-        let turns = 0;
-        let protectedTokens = 0;
-        let toCompact = [];
-        // 从最新往旧遍历
-        for (let i = messages.length - 1; i >= 0; i--) {
-            const msg = messages[i];
-            // 简单计数：用户消息计为一轮
-            if (this.isUserMessage(msg))
-                turns++;
-            // 保留最近的 keepLastTurns 轮
-            if (turns <= keepLastTurns)
-                continue;
-            // 如果消息已经被 compacted，停止继续往旧遍历
-            if (msg.compacted)
-                break;
-            const estimatedTokens = this.estimateTokens(msg.content);
-            protectedTokens += estimatedTokens;
-            if (protectedTokens > PRUNE_PROTECT) {
-                totalFreed += estimatedTokens;
-                toCompact.push(msg);
-            }
-        }
-        // 收益不够大时不执行
-        if (totalFreed < PRUNE_MINIMUM) {
-            return {
-                strategy: "prune",
-                freedTokens: 0,
-                preservedTurns: Math.min(turns, keepLastTurns),
-                compactedMessages: 0,
-                remainingTokens: 0,
-                timestamp: new Date().toISOString(),
-            };
-        }
-        // 标记 compacted（实际调用者负责持久化）
-        for (const msg of toCompact) {
-            msg.compacted = true;
-            // 保留摘要信息
-            if (msg.content.length > 200) {
-                msg.compactionSummary = msg.content.slice(0, 200) + "...";
-            }
-            compactedCount++;
-        }
-        return {
-            strategy: "prune",
-            freedTokens: totalFreed,
-            preservedTurns: Math.min(turns, keepLastTurns),
-            compactedMessages: compactedCount,
-            remainingTokens: Math.max(0, protectedTokens - totalFreed),
-            timestamp: new Date().toISOString(),
-        };
-    }
-    /**
-     * 判断消息是否为"用户消息"（用于轮次计数）。
-     * 在 ACode 的上下文中，DirectMessage 我们都是"用户发给 Agent"或"Agent 间"的消息。
-     * 简单启发：判断是否有 requireReply 或消息较短（工具输出通常很长）。
-     */
-    isUserMessage(msg) {
-        // DirectMessage 如果没有 requireReply 且内容较短，可能是系统消息
-        if ("requireReply" in msg) {
-            // requireReply=true 通常是用户消息，或需要 Agent 回复的消息
-            // Agent 工具输出通常很长，这里用启发式判断
-            return msg.requireReply || msg.content.length < 500;
-        }
-        // GroupMessage 简单按内容长度判断
-        return msg.content.length < 500;
-    }
-    // ============ 获取压缩摘要 ============
-    /**
-     * 获取最后一次压缩的摘要（如有）。
-     */
-    async getCompactionSummary(sessionId) {
-        const session = await this.getOrCreateSession(sessionId);
-        if (session.compactions.length === 0)
-            return null;
-        // 读取最近一次摘要
-        const uri = `session://compacted/${sessionId}`;
-        const memory = await this.memoryStore.readMemory({ uri, memory_type: 'session_context' });
-        if (memory && typeof memory.content === 'string') {
-            try {
-                const parsed = JSON.parse(memory.content);
-                return parsed.summary ?? null;
-            }
-            catch {
-                return memory.content.length > 500
-                    ? memory.content.slice(0, 500) + "..."
-                    : memory.content;
-            }
-        }
-        return null;
-    }
-    // ============ 持久化 ============
-    /**
-     * 将会话上下文记录持久化到 MemoryStore。
-     */
-    async saveToMemory(session) {
-        const uri = `session://tokens/${session.sessionId}`;
-        const existing = await this.memoryStore.readMemory({ uri, memory_type: 'session_context' });
-        if (existing) {
-            await this.memoryStore.updateMemory({
-                uri,
-                memory_type: 'session_context',
-                old_string: existing.content,
-                new_string: JSON.stringify(session),
-                keywords: ['context', 'token-usage', session.sessionId],
-            });
-        }
-        else {
-            await this.memoryStore.createMemory({
-                uri,
-                memory_type: 'session_context',
-                content: JSON.stringify(session),
-                keywords: ['context', 'token-usage', session.sessionId],
-            });
-        }
-    }
-    /**
-     * 从 MemoryStore 加载会话上下文记录。
-     */
-    async loadFromMemory(sessionId) {
-        const uri = `session://tokens/${sessionId}`;
-        const memory = await this.memoryStore.readMemory({ uri, memory_type: 'session_context' });
-        if (!memory)
-            return null;
-        try {
-            return JSON.parse(memory.content);
-        }
-        catch {
-            logger.warn("[ContextManager] Failed to parse session context from memory", { sessionId });
-            return null;
-        }
-    }
-    /**
-     * 保存压缩结果到 MemoryStore。
-     */
-    async saveCompactionResult(sessionId, result) {
-        const session = await this.getOrCreateSession(sessionId);
-        session.compactions.push({
-            timestamp: result.timestamp,
-            strategy: result.strategy,
-            freedTokens: result.freedTokens,
-            messageCount: result.compactedMessages,
-        });
-        // 保存压缩摘要
-        if (result.summary) {
-            const uri = `session://compacted/${sessionId}`;
-            const existing = await this.memoryStore.readMemory({ uri, memory_type: 'session_context' });
-            const content = JSON.stringify({ summary: result.summary, timestamp: result.timestamp });
-            if (existing) {
-                await this.memoryStore.updateMemory({
-                    uri,
-                    memory_type: 'session_context',
-                    old_string: existing.content,
-                    new_string: content,
-                    keywords: ['context', 'compacted', sessionId],
-                });
-            }
-            else {
-                await this.memoryStore.createMemory({
-                    uri,
-                    memory_type: 'session_context',
-                    content,
-                    keywords: ['context', 'compacted', sessionId],
-                });
-            }
-        }
-        // 重置溢出警告
-        session.overflowWarned = false;
-        await this.saveToMemory(session);
-    }
-    /**
-     * 重置溢出警告状态。
-     */
-    async resetOverflowWarning(sessionId) {
-        const session = await this.getOrCreateSession(sessionId);
-        session.overflowWarned = false;
-        await this.saveToMemory(session);
-    }
-    /**
-     * 清空内存态记录（进程退出或会话切换时调用）。
-     */
-    clearSession(sessionId) {
-        if (sessionId) {
-            this.sessions.delete(sessionId);
-        }
-        else {
-            this.sessions.clear();
-        }
-    }
-}
-//# sourceMappingURL=context-manager.js.map
+import{logger as l}from"./logger.js";const g=2e4,y=4e4,O=2e4,f=3,_=128e3,x=8192,c={"claude-4":{contextLimit:2e5,maxOutput:8192},"claude-sonnet-4":{contextLimit:2e5,maxOutput:8192},"claude-3-5-sonnet":{contextLimit:2e5,maxOutput:8192},"claude-3-opus":{contextLimit:2e5,maxOutput:8192},"claude-3-haiku":{contextLimit:2e5,maxOutput:8192},"claude-2":{contextLimit:1e5,maxOutput:4096},"gpt-4o":{contextLimit:128e3,maxOutput:16384},"gpt-4-turbo":{contextLimit:128e3,maxOutput:4096},"gpt-4":{contextLimit:8192,maxOutput:4096},"gpt-3.5-turbo":{contextLimit:16384,maxOutput:4096},o1:{contextLimit:2e5,maxOutput:1e5},"o3-mini":{contextLimit:2e5,maxOutput:1e5},"gemini-2.0-flash":{contextLimit:1048576,maxOutput:8192},"gemini-2.5-pro":{contextLimit:1048576,maxOutput:8192},"gemini-1.5-pro":{contextLimit:2097152,maxOutput:8192},"gemini-1.5-flash":{contextLimit:1048576,maxOutput:8192},deepseek:{contextLimit:64e3,maxOutput:8192},"deepseek-chat":{contextLimit:64e3,maxOutput:8192},"deepseek-reasoner":{contextLimit:64e3,maxOutput:8192},"qwen-max":{contextLimit:32e3,maxOutput:8192},"qwen-plus":{contextLimit:128e3,maxOutput:8192},"qwen-turbo":{contextLimit:1e6,maxOutput:8192},"glm-4":{contextLimit:128e3,maxOutput:4096},"yi-":{contextLimit:2e5,maxOutput:4096},"llama-3":{contextLimit:8192,maxOutput:4096},"llama-2":{contextLimit:4096,maxOutput:4096},mistral:{contextLimit:32e3,maxOutput:4096},mixtral:{contextLimit:32e3,maxOutput:4096},codestral:{contextLimit:256e3,maxOutput:8192},grok:{contextLimit:131072,maxOutput:4096},"nova-":{contextLimit:2e5,maxOutput:8192}};function h(u){if(!u)return{contextLimit:_,maxOutput:x};const t=u.toLowerCase().trim();if(c[t])return c[t];const o=Object.keys(c).filter(e=>t.startsWith(e)).sort((e,n)=>n.length-e.length);return o.length>0?c[o[0]]:{contextLimit:_,maxOutput:x}}const L=new Set(["skill"]);class w{sessions=new Map;memoryStore;constructor(t){this.memoryStore=t}lookupModelLimits(t){return h(t)}estimateTokens(t){if(!t)return 0;let o=0,e=0;for(const s of t)s>"\xFF"?o++:e++;const n=Math.ceil(o/1.5+e/4);return Math.ceil(n*1.15)}estimateTotalUsage(t){return t.input+t.output+(t.reasoning??0)+(t.cache?.read??0)+(t.cache?.write??0)}async getOrCreateSession(t){const o=this.sessions.get(t);if(o)return o;const e=await this.loadFromMemory(t);if(e)return this.sessions.set(t,e),e;const n={sessionId:t,tokens:{totalInput:0,totalOutput:0,totalReasoning:0},lastActivity:new Date().toISOString(),compactions:[],overflowWarned:!1,messageCount:0};return this.sessions.set(t,n),n}async recordUsage(t,o,e){const n=await this.getOrCreateSession(t),s=this.estimateTotalUsage(o),i=o.input??Math.floor(s*.7),r=o.output??Math.floor(s*.3);n.tokens.totalInput+=i,n.tokens.totalOutput+=r,n.tokens.totalReasoning+=o.reasoning??0,n.lastActivity=new Date().toISOString(),n.messageCount++,e&&(n.modelInfo=e),this.saveToMemory(n).catch(a=>{l.error("[ContextManager] saveToMemory failed",a)})}async isOverflow(t,o){const e=await this.getOrCreateSession(t),n=o??e.modelInfo;if(!n||!n.contextLimit||n.contextLimit===0)return!1;const s=e.tokens.totalInput+e.tokens.totalOutput+e.tokens.totalReasoning,i=n.maxOutput??4096,r=Math.min(g,i),a=n.maxInput?n.maxInput-r:n.contextLimit-i;return s>=a}async getStats(t,o){const e=await this.getOrCreateSession(t),n=o??e.modelInfo,s=e.tokens.totalInput+e.tokens.totalOutput+e.tokens.totalReasoning,i=n?.contextLimit??128e3,r=Math.max(0,i-s),a=i>0?Math.round(s/i*100):0;return{totalUsed:s,totalInput:e.tokens.totalInput,totalOutput:e.tokens.totalOutput,contextLimit:i,remaining:r,usagePercent:a,overflowWarning:await this.isOverflow(t,n),lastChecked:new Date().toISOString()}}prune(t,o=f){if(t.length===0)return{strategy:"prune",freedTokens:0,preservedTurns:0,compactedMessages:0,remainingTokens:0,timestamp:new Date().toISOString()};let e=0,n=0,s=0,i=0,r=[];for(let a=t.length-1;a>=0;a--){const m=t[a];if(this.isUserMessage(m)&&s++,s<=o)continue;if(m.compacted)break;const p=this.estimateTokens(m.content);i+=p,i>y&&(e+=p,r.push(m))}if(e<O)return{strategy:"prune",freedTokens:0,preservedTurns:Math.min(s,o),compactedMessages:0,remainingTokens:0,timestamp:new Date().toISOString()};for(const a of r)a.compacted=!0,a.content.length>200&&(a.compactionSummary=a.content.slice(0,200)+"..."),n++;return{strategy:"prune",freedTokens:e,preservedTurns:Math.min(s,o),compactedMessages:n,remainingTokens:Math.max(0,i-e),timestamp:new Date().toISOString()}}isUserMessage(t){return"requireReply"in t&&t.requireReply||t.content.length<500}async getCompactionSummary(t){if((await this.getOrCreateSession(t)).compactions.length===0)return null;const e=`session://compacted/${t}`,n=await this.memoryStore.readMemory({uri:e,memory_type:"session_context"});if(n&&typeof n.content=="string")try{return JSON.parse(n.content).summary??null}catch{return n.content.length>500?n.content.slice(0,500)+"...":n.content}return null}async saveToMemory(t){const o=`session://tokens/${t.sessionId}`,e=await this.memoryStore.readMemory({uri:o,memory_type:"session_context"});e?await this.memoryStore.updateMemory({uri:o,memory_type:"session_context",old_string:e.content,new_string:JSON.stringify(t),keywords:["context","token-usage",t.sessionId]}):await this.memoryStore.createMemory({uri:o,memory_type:"session_context",content:JSON.stringify(t),keywords:["context","token-usage",t.sessionId]})}async loadFromMemory(t){const o=`session://tokens/${t}`,e=await this.memoryStore.readMemory({uri:o,memory_type:"session_context"});if(!e)return null;try{return JSON.parse(e.content)}catch{return l.warn("[ContextManager] Failed to parse session context from memory",{sessionId:t}),null}}async saveCompactionResult(t,o){const e=await this.getOrCreateSession(t);if(e.compactions.push({timestamp:o.timestamp,strategy:o.strategy,freedTokens:o.freedTokens,messageCount:o.compactedMessages}),o.summary){const n=`session://compacted/${t}`,s=await this.memoryStore.readMemory({uri:n,memory_type:"session_context"}),i=JSON.stringify({summary:o.summary,timestamp:o.timestamp});s?await this.memoryStore.updateMemory({uri:n,memory_type:"session_context",old_string:s.content,new_string:i,keywords:["context","compacted",t]}):await this.memoryStore.createMemory({uri:n,memory_type:"session_context",content:i,keywords:["context","compacted",t]})}e.overflowWarned=!1,await this.saveToMemory(e)}async resetOverflowWarning(t){const o=await this.getOrCreateSession(t);o.overflowWarned=!1,await this.saveToMemory(o)}clearSession(t){t?this.sessions.delete(t):this.sessions.clear()}}export{w as ContextManager};
+

package/package/aws-client-agent-mcp/dist/context-manager.test.d.ts

@@ -3,4 +3,3 @@
  * 验证 token 估算、溢出检测、prune 策略的核心逻辑
  */
 export {};
-//# sourceMappingURL=context-manager.test.d.ts.map

package/package/aws-client-agent-mcp/dist/context-manager.test.js

@@ -310,4 +310,3 @@ describe("ContextManager", () => {
         });
     });
 });
-//# sourceMappingURL=context-manager.test.js.map

package/package/aws-client-agent-mcp/dist/http-client.d.ts

@@ -83,4 +83,3 @@ export declare class HttpClient {
      */
     getMcpHttpUrl(): string;
 }
-//# sourceMappingURL=http-client.d.ts.map