BossAgents × WeChat Intelligent Agent Ecosystem: Learning, Integration, and Evolution Plan (Optimized Version v2.0)

Version: v2.0 | Date: 2026-06-24 | Based on In-depth Audit of Current Codebase + Latest 2025-2026 WeChat/Tencent AI Agent Ecosystem Research

0. Key Revisions from v1 to v2

I. Current State Benchmark Analysis (Revised Version)

1.1 BossAgents Current Capability Matrix vs. WeChat Ecosystem

2. Option 1: Learning from WeChat Agent — Six Upgrade Directions (Revised Version)

2.1 ? P0 Priority: Memory System Upgrade (Benchmarked against Hy-Memory)

Current Issues (Revised):

• BossAgents has no MEMORY.md (v1 plan incorrectly assumed existence)
• ConversationManager only handles session-level context, lacks cross-session long-term memory
• RuleEvolution has an evolution loop framework, but lacks independent evolution log storage
• Audit logs are written to DB but the audit-logs/ directory is empty, and query/replay APIs are missing

Hy-Memory Reusable Architecture (Revised Version):

┌─────────────────────────────────────────────────────────────┐
│              BossAgents 记忆系统升级方案 v2                    │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Layer 1: Atomic Traces (原子痕迹)                           │
│  └─ 复用现有 audit-log.js 的 start/complete 两阶段日志       │
│     → 修正：audit-logs/ 目录空置，需确认DB写入是否生效        │
│     → 新增：查询/回放 API（/api/memory/traces）              │
│                                                             │
│  Layer 2: Atomic Facts (原子事实) ★ 核心新增                 │
│  └─ 从交互中提取事实：老板偏好、供应商偏好、常用操作          │
│     → 新增 memory/facts.json 结构化存储                      │
│     → 集成点：CapabilityDispatcher.execute() finally块       │
│     → 与现有自进化学习器（runWorker finally块）复用模式        │
│                                                             │
│  Layer 3: Identity Profile (用户画像) ★ 核心新增             │
│  └─ 企业信息、老板角色、常用对象类型、审批偏好                │
│     → 新增 memory/profiles/{tenantId}.json                   │
│     → 与租户中间件(tenant-middleware.js)集成                  │
│     → 与小程序Dashboard数据对接                               │
│                                                             │
│  Layer 4: Session Summaries (会话摘要)                       │
│  └─ 增强现有 ConversationManager                             │
│     → 按"业务主题"聚合（BOM/采购/变更）                      │
│     → 新增 staff_session_summaries 表                        │
│                                                             │
│  Layer 5: Mental Models (心智模型)                            │
│  └─ 深度理解老板决策模式：风险偏好、成本敏感度、创新意愿      │
│     → 新增 memory/mental-models/{tenantId}.json              │
│     → 由 System 2 异步构建                                   │
│                                                             │
│  Layer 6: Forward-looking Intentions (前瞻意图)               │
│  └─ 预判老板下一步需求                                       │
│     → 新增 memory/intentions/{tenantId}.json                 │
│     → 心跳任务定期更新（复用LiteScheduler cron机制）          │
│                                                             │
│  System 1 (Day Shift): 实时写入                               │
│  └─ CapabilityDispatcher.execute() 每次执行后提取关键事实     │
│     → 复用现有自进化学习器的 finally 块模式                   │
│                                                             │
│  System 2 (Night Shift): 异步分析                             │
│  └─ 新增 cron job，分析 audit logs → 更新 Mental Models       │
│     → 复用 LiteScheduler 的 cron 调度机制                    │
│                                                             │
│  演化链 (Evolution Chain):                                    │
│  └─ supersedes 指针链接旧/新记忆                              │
│     → 复用 RuleEvolution 的版本管理思路                       │
│     → 新增 memory/evolution-chain.json                       │
│                                                             │
└─────────────────────────────────────────────────────────────┘

具体实施方案（修正版）：

| Step | Content | Dependencies | Effort |

|------|------|------|--------|

| 0 | Prerequisite: Confirm audit log DB writes are effective, fix empty audit-logs directory | None | 0.5 days |

| 1 | Create server/services/memory-service.js six-layer memory core | Step 0 | 3 days |

| 2 | Layer 1 Fix: Audit log query/playback API | Step 1 | 1 day |

| 3 | Layer 2-3 Implementation: Fact extraction + profile construction | Step 1 | 5 days |

| 4 | System 1 Integration: Dispatcher finally block extracts facts | Step 3 | 2 days |

| 5 | System 2 cron job: Nightly analysis + profile update | Step 3, LiteScheduler | 3 days |

| 6 | Layer 4 Enhancement: ConversationManager session summary | Step 1, ConversationManager | 2 days |

| 7 | Layer 5-6 Implementation: Mental Model + Prospective Intent | Step 5 | 5 days |

| 8 | Evolution Chain: supersedes pointer + version management | Step 7, RuleEvolution | 3 days |

| Total | | | 24.5 days |

2.2 ? P1 Priority: Multi-Agent Parallel Execution (Benchmarking WorkBuddy + ADP 3.0)

Current Issues (Revised):

• LiteScheduler already has a collaboration chain (triggering next agent via onComplete), describing it as "serial scheduling" in v1 is inaccurate.
• However, it indeed lacks: parallel execution, DAG orchestration, and result aggregation.
• The mutex lock _withStaffMutex is an in-memory Map, lacking distributed locking, timeout release, and deadlock detection.

Upgrade Plan (Revised Version):

现有架构（v1低估了）:
  cron(15min) → Staff A → runOnce → _runWorker → onComplete → Staff B
  协作链已支持串行编排，Goal调度有框架

升级后:
  老板: "比较三家供应商报价并生成采购建议"
       │
       ▼
  CapabilityDispatcher
       │
       ▼
  TaskDecomposer (新增)  ← LLM 拆解
       │
       ├──→ SubTask 1: 查询供应商A报价  ──→ Staff A (并行)
       ├──→ SubTask 2: 查询供应商B报价  ──→ Staff B (并行)
       ├──→ SubTask 3: 查询供应商C报价  ──→ Staff C (并行)
       │
       ▼
  ResultAggregator (新增)  ← LLM 汇总
       │
       ▼
  返回："推荐供应商B，综合评分最高"

新增模块：

• server/core/task-decomposer.js — Task decomposition + parallel scheduling
• server/core/result-aggregator.js — Result aggregation + conflict resolution
• Add a runParallel(staffIds[], subTasks[]) method to LiteScheduler.
• Mutex lock upgrade: Timeout release (30s) + deadlock detection.

Revised effort estimate: 5 days → 7 days (requires additional mutex lock upgrade).

2.3 ? P1 Priority: Enhanced LLM Router (Revised Version)

Current Status (Revised): SmartLLMRouter already has a 5-level fallback chain; describing it as "two-way routing" in v1 was a severe underestimation.

现有5级降级链路（v1未识别）:
  MTCLAW(Worker/Solver) → AscendNPU → Ollama → DeepSeek → 规则引擎 → simulatedThinking
  + 熔断机制 + 响应缓存 + YAML配置驱动 + 统计监控

升级方向（修正）:
  1. 分类升级：关键词匹配 → 语义级分类（嵌入向量）
  2. 动态路由：基于历史准确率调整模型选择
  3. 混元接入：新增 provider（非替换，作为降级链路一环）
  4. A/B测试：同一任务对比不同模型效果

混元模型接入方案：

// smart-llm-router.js 新增 provider
providers: {
  // 现有5级保持不变
  mtclaw: { ... },
  ascend: { ... },
  ollama: { ... },
  deepseek: { ... },
  // 新增混元
  hunyuan: {
    baseURL: 'https://api.hunyuan.cloud.tencent.com/v1',
    apiKey: process.env.HUNYUAN_API_KEY,
    models: {
      fast: 'hunyuan-2.0-instruct-20251109',
      think: 'hunyuan-2.0-thinking-20251109',
    },
    // 插入降级链路：DeepSeek之后、规则引擎之前
    priority: 4.5
  }
}

工作量修正： 3天 → 5天（需增加语义分类+动态路由）

2.4 ? P0 Priority (Promoted in v2): Knowledge Base Frontend-Backend Integration

This issue was not mentioned in v1, but it is a more urgent foundational problem than Agentic RAG.

Current Status:

• The backend knowledge-search.js has FTS5 full-text search + LIKE fallback, but FTS5 is unavailable in the current SQLite version.
• The frontend KnowledgeService uses hardcoded data (design patterns/best practices/AML knowledge) and is not connected to the backend search API.
• Chinese segmentation npm packages like jieba fail to compile in the Windows/Node v22 environment.

Upgrade Plan:

Phase 1 (紧急): 前端接入后端搜索API
  KnowledgeService → fetch('/api/knowledge/search?q=...')
  → 解决硬编码问题，数据来源统一

Phase 2 (增强): 搜索质量提升
  → FTS5不可用，改用LIKE+简单中文分词
  → 或引入 better-sqlite3 的 FTS5 支持

Phase 3 (高级): Agentic RAG
  → Agent 自评检索质量
  → 重新检索/换策略

工作量： Phase 1: 2天 | Phase 2: 3天 | Phase 3: 5天

2.5 ? P2 Priority: User Profile System (Benchmarking ima.copilot memory)

Merge with the Layer 3 memory system from 2.1; do not implement separately.

2.6 ? P3 Priority: Self-Evolution Closed-Loop Enhancement (Benchmarking evolution chain)

Current Status (Revised): RuleEvolution already has a complete closed-loop framework (414 lines), but the approval and deployment stage is not integrated with the UI.

Upgrade Direction:

• Independently persist evolution logs (add new sciot_evolution_logs table)
• UI for approval and deployment stage (frontend management interface)
• Retain supersedes pointer to preserve complete evolution history.

Effort Estimate: 3 days

III. Option 2: Integration with WeChat Agent — Four Integration Paths (Revised Version)

3.1 ? P0→? P1 (v2 Downgrade): Integrate ClawBot — Enable Native WeChat Entry Point

v2 Revision: Feasibility is questionable; requires preliminary research.

v1 Issues:

• OpenClaw/iLink is part of Tencent's internal ecosystem; external project access paths are not publicly documented.
• npx -y @tencent-weixin/openclaw-weixin-cli@latest install — Is this package publicly available?
• "WeChat App: Me → Settings → Plugins → ClawBot → Enable" — Is this feature available to all developers?
• BossAgins are currently running on OpenClaw (as per v1 documentation), but no OpenClaw-related configurations or dependencies were found in the codebase.

v2 Suggestion:

Step 0 (调研, 1天):
  → 确认 OpenClaw CLI 是否公开可用
  → 确认 ClawBot 插件是否对第三方开发者开放
  → 确认 iLink 协议接入门槛

Step 1 (如果可行, 1-2天):
  → 安装 OpenClaw CLI
  → 配置 wechat channel
  → CapabilityDispatcher 已有 source: 'wechat' 路由支持

Step 2 (如果不可行, 备选方案):
  → 企业微信自建应用（标准API，完全可控）
  → WeChat+客服消息接口
  → 微信小程序内嵌对话（已有bossagents-miniapp）

工作量修正： 1-2天 → 1天调研 + 1-2天实施（如果可行）

3.2 ? P1 Priority: Integrate bossagents-miniapp into the WeChat AI ecosystem

Current Status (Corrected):

• bossagents-miniapp is a Vue 3 + uni-app project with 7 existing pages and 10 API modules.
• It already has WeChat login (miniapp-auth.js), Dashboard data, and notification list functionalities.
• A hardcoded JWT secret fallback 'change-me-in-production' exists — a security risk that needs to be fixed first.

Two modes (keeping v1 approach):

A. Automatic Mode (Quick Win):

• Authorize WeChat to read the mini-program source code during submission for review.
• Requires zero additional development.
• ⚠️ Needs confirmation: The specific policy for WeChat opening mini-program AI ecosystem access on June 8, 2026.

B. Development Mode (Deep Integration):

• Build custom integration based on the WeChat Mini-Program AI SDK.
• Embed a "Brain" chat entry within the mini-program (existing ai-chat page).
• Call WeChat speech recognition → Backend CapabilityDispatcher.

v2 New: Pre-fix

• JWT secret key read from environment variable, remove hardcoded fallback
• Token cache changed to Redis/file persistence (multi-instance safe)

Workload correction: 0.5 days → 1.5 days (includes JWT fix)

3.3 ? P1 Priority: Integrate Hunyuan model (domestic compliance + cost optimization)

Merge with 2.3, not implemented separately.

3.4 ? P2 Priority: Integrate Hy-Memory plugin

v2 Correction: Hy-Memory is an OpenClaw plugin, need to confirm if BossAgents runs on OpenClaw first (relevant config not found in codebase)

v2 Suggestion:

• If confirmed to run on OpenClaw → install with one command
• If running independently → build own memory system (i.e., 2.1 solution), do not rely on Hy-Memory plugin
• Prioritize self-built to avoid external dependencies

IV. Solution Three: Learn from WorkBuddy's Best Practices (Revised Version)

4.1 "Zero Deployment · Install and Use" — Lower BossAgents Deployment Barrier

Current Status (Revised):

• start.bat already has basic startup logic (Line 36: Check Node→Check Port→Kill Process→Start)
• docker-compose.yml orchestrates 3 services (Node.js + nginx + certbot)
• Missing: No environment checks, no dependency installation, no DB initialization, no configuration wizard

Upgrade plan (revised version):

Phase 1 (start.bat 增强):
  → 检查 Node.js 版本
  → 检查 pnpm 是否安装
  → 自动 pnpm install（如果 node_modules 不存在）
  → 检查 .env 文件（不存在则从 .env.example Copy）
  → 自动 DB 初始化（调用 db-schema-init.js）
  → 启动服务器

Phase 2 (Docker Optimize):
  → 单命令 docker run（包含 Node.js + SQLite + nginx）
  → 健康检查依赖顺序
  → 数据卷持久化

Phase 3 (Windows 安装包):
  → pkg/nexe 打包为 .exe
  → 首次启动配置向导

工作量： Phase 1: 2天 | Phase 2: 3天 | Phase 3: 5天

4.2 "Send tasks on mobile, execute automatically on PC" — Remote control enhancement

Current status (revised): Feishu → Dispatcher → Digital Employee → results back to Feishu is implemented. WebSocket service (ws-push-service.js) is implemented.

Upgrade direction:

• After integrating WeChat channel → Send commands via WeChat → Server executes → Results return to WeChat
• WebSocket pushes execution progress in real-time (ws-push-service.js + useWebSocket.js already exist)

Workload: Depends on WeChat channel integration (see 3.1)

4.3-4.5 Keep v1 plan, no revisions

Five, Priority Implementation Roadmap (Revised Version)

Phase 0 — Prerequisite Fixes (1-2 days) ★ New in v2

Phase 1 — Quick Wins (1-2 weeks)

| # | Task | Type | Value | Effort | v2 Changes |

|---|------|------|------|--------|--------|

| 1 | Integrate WeChat channel (ClawBot or self-built on WeCom) | Integration | ? Very High | 1-3 days | Depends on completing 0d study |

| 2 | Integrate bossagents-miniapp auto mode with WeChat AI | Integration | ? High | 1.5 days | Includes JWT fix |

| 3 | Implement Memory system Layer 1-3 | Learning | ? Very High | 9 days | Includes audit log fix |

| 4 | One-click deployment package optimization (start.bat enhancement) | Learning | ? Medium | 2 days | No change |

Phase 1 Milestone: The boss can directly command digital employees within WeChat, and the system begins to remember the boss's preferences.

Phase 2 — Deep Enhancement (3-4 weeks)

Phase 3 — Ecosystem Expansion (5-8 weeks)

Six. Architecture Evolution Panorama (Revised Version)

                         ┌──────────────────────────┐
                         │   WeChat/企微/飞书/元宝     │
                         │  (ClawBot或企微自建入口)   │
                         └────────────┬─────────────┘
                                      │
┌─────────────────────────────────────┼──────────────────────────────────┐
│                             OpenClaw Gateway (待确认)                   │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────────────┐      │
│  │ WeChat   │  │ Feishu   │  │ MiniApp  │  │ Web Dashboard    │      │
│  │ Channel  │  │ Channel  │  │ Channel  │  │ Channel          │      │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘  └───────┬──────────┘      │
│       └──────────────┼────────────┼─────────────────┘                 │
│                      ▼                                                 │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │              CapabilityDispatcher (统一调度)                   │    │
│  │  ┌──────────┐  ┌──────────┐  ┌──────────┐                   │    │
│  │  │ 参数归一化│  │ 审计日志  │  │ 权限校验  │                   │    │
│  │  └──────────┘  └──────────┘  └──────────┘                   │    │
│  └──────────────────────────┬───────────────────────────────────┘    │
│                             ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │                   TaskDecomposer (★新增)                      │    │
│  │        LLM 拆解复杂任务 → 并行调度多员工                       │    │
│  └──────────────────────────┬───────────────────────────────────┘    │
│                             ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │         SmartLLMRouter (★现有5级+增强语义分类+混元)           │    │
│  │  L1:规则引擎  L2:Worker  L3:Solver  L4:混元  L5:降级链路     │    │
│  │  (0ms,0cost)  (~70ms)    (~500ms)   (新增)   (现有5级)       │    │
│  └──────────────────────────┬───────────────────────────────────┘    │
│                             ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │              CapabilityRuntime (六大能力)                      │    │
│  │  identify │ create │ repair │ optimize │ compare │ generate   │    │
│  │              ★ Agentic RAG (Phase 3)                          │    │
│  └──────────────────────────┬───────────────────────────────────┘    │
│                             ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │                   Memory Service (★新增)                       │    │
│  │  Traces → Facts → Profile → Summaries → Models → Intentions  │    │
│  │  System 1 (实时写入)  +  System 2 (异步分析)                  │    │
│  └──────────────────────────┬───────────────────────────────────┘    │
│                             ▼                                         │
│  ┌──────────────────────────────────────────────────────────────┐    │
│  │              Knowledge base (★前后端打通 + Agentic RAG)               │    │
│  │  SCIOT规则引擎 + FTS5/LIKE检索 + 向量检索(远期)              │    │
│  └──────────────────────────────────────────────────────────────┘    │
└──────────────────────────────────────────────────────────────────────┘

七、总结（修正版）

核心结论

1. 最大的学习机会：记忆系统。 Hy-Memory 的六层记忆+双系统+演化链是 BossAgents 最需要学习的架构。当前 BossAgents 在这块几乎空白（MEMORY.md不存在，仅有会话级上下文）。

1. 最紧迫的前置修复：安全+基础。 JWT密钥硬编码、审计日志空置、知识库前后端断裂 — 这些是比新功能更紧迫的技术债务。

1. ClawBot接入可行性存疑。 OpenClaw/iLink是Tencent内部生态，需先调研确认外部项目接入路径。备选方案：企业微信自建应用。

1. LLM调度能力被低估。 SmartLLMRouter已有5级降级+熔断+缓存，升级方向是语义分类+动态路由，而非从零构建。

1. 知识库前后端断裂是隐藏的P0。 前端硬编码数据、后端FTS5不可用、jieba编译失败 — 这些问题比Agentic RAG更基础。

工作量总估算（v2修正）

| Phase | v1估算 | v2估算 | 差异原因 |

|-------|--------|--------|----------|

| Phase 0 (Newly Added) | 0 | 4 days | Security Fixes + Foundation Confirmation + Knowledge Base Integration + Research |

| Phase 1 | 8.5 days | 15.5 days | JWT Fixes + Audit Log Fixes + Preliminary Research |

| Phase 2 | 29 days | 32 days | Mutex Lock Upgrade + Semantic Classification + Dynamic Routing |

| Phase 3 | 15 days | 18 days | Self-Evolving UI + Hy-Memory Self-Building |

| Total | 52.5 days | 69.5 days | Version 2 is more pragmatic, includes preliminary fixes and research |

Summary in One Sentence

What BossAgents aims to do is: first repair the foundation (security + basics + knowledge base), then understand the boss with Hy-Memory-level memory, remain on standby via the WeChat channel, leverage the Hunyuan model for cost reduction and compliance, and with a WorkBuddy-style zero-barrier experience, truly make industrial digital employees the AI companion in the boss's pocket.