中文 | English
Bio-inspired swarm intelligence plugin for OpenClaw with 6-layer architecture, 20+ algorithms, skill-symbiosis scheduling, multi-type pheromones, structured DAG orchestration, speculative execution, work-stealing, state-convergence, runtime global-modulator, three feedback loops, governance triple metrics, and full observability.
Real-time monitoring dashboard with Gource-inspired swarm visualization (http://localhost:19100/?demo)
Claw-Swarm is a bio-inspired swarm intelligence plugin for OpenClaw that brings self-organizing multi-agent coordination to LLM-powered workflows. It models agent collaboration after real bee colonies — using pheromone trails, gossip protocols, and structured memory to let autonomous agents discover, negotiate, and complete tasks without centralized control.
Claw-Swarm 是 OpenClaw 的仿生蜂群智能插件，为 LLM 驱动的工作流带来自组织多智能体协调能力。它模拟真实蜂群的协作方式——通过信息素路径、Gossip 协议和结构化记忆，让自治智能体在无中央控制的情况下自主发现、协商并完成任务。
| Challenge / 挑战 | Symptom / 表现 | Claw-Swarm Solution / 解决方案 |
|---|---|---|
| Collaboration blind spots / 协作盲区 | Agents unaware of each other's progress / 智能体彼此不知道对方进度 | Indirect pheromone communication + Gossip protocol + StigmergicBoard / 信息素间接通信 + Gossip 协议 + 公告板 |
| Memory fragmentation / 记忆碎片 | Knowledge lost after context window reset / 上下文窗口重置后知识丢失 | 3-tier memory (working / episodic / semantic) / 三层记忆架构 |
| Scheduling inefficiency / 调度低效 | Manual task assignment, no adaptability / 手动分配任务，缺乏自适应 | DAG decomposition + Contract Net + ABC scheduling + Lotka-Volterra dynamics / DAG 分解 + 合同网 + ABC 调度 + 种群动力学 |
| Tool brittleness / 工具脆弱 | Single API failure cascades to full stop / 单个 API 失败全线崩溃 | AJV pre-validation + per-tool circuit breaker + retry injection + failure vaccination / AJV 预校验 + 断路器 + 重试注入 + 失败免疫 |
| Observability gap / 可观测性不足 | No visibility into swarm dynamics / 无法了解蜂群内部动态 | Real-time hex-hive dashboard + RED metrics + SSE + Jaeger-lite tracing / 实时仪表盘 + RED 指标 + SSE + Jaeger-lite 追踪 |
| Idle resources / 资源闲置 | Agents sitting idle while tasks queue up / Agent 空闲而任务排队 | Idle detection + automatic recruit pheromone emission / 空闲检测 + 自动招募信息素发射 |
The 4-layer V4.0 architecture became bloated after adding DAG orchestration, Contract Net, and knowledge graph modules. V5.0 re-divided responsibilities into 6 cleanly separated layers, with dependencies flowing strictly downward (L6 → L1). Only L5 couples to OpenClaw — L1–L4 can be independently reused in any Node.js 22+ environment.
V4.0 的 4 层架构在引入 DAG 编排、合同网协议和知识图谱模块后变得臃肿。V5.0 将职责重新划分为 6 个清晰解耦的层级，依赖严格向下流动（L6 → L1）。仅 L5 耦合 OpenClaw——L1–L4 可在任何 Node.js 22+ 环境中独立复用。
| Feature / 特性 | Description | 描述 |
|---|---|---|
| 6-Layer Architecture | Clean separation: infra, comm, agent, orchestration, app, monitoring | 六层解耦：基础设施、通信、智能体、编排、应用、监控 |
| 18+ Bio-Inspired Algorithms | MMAS, ACO, Ebbinghaus, BFS, PARL, GEP, CPM, Jaccard, MoE, FIPA CNP, ABC, k-means++, Lotka-Volterra, FRTM, cosine similarity, PI controller + more | 18+ 种仿生/经典算法融合 |
| 3-Tier Memory | Working (focus/context/scratchpad), Episodic (forgetting curve), Semantic (knowledge graph) | 三级记忆：工作记忆、情景记忆、语义知识图谱 |
| Multi-Type Pheromones | MMAS-bounded signals with typed decay (trail/alarm/recruit/food/danger), pressure gradient auto-escalation | 多类型信息素 + 压力梯度自动升级 |
| Stigmergic Board | Persistent bulletin board for indirect coordination beyond short-lived pheromones | 持久公告板，补充短寿信息素的间接协调 |
| Failure Vaccination | Pattern-based immunization with effectiveness feedback loop | 基于模式的免疫记忆 + 效果反馈循环 |
| DAG Orchestration | Task decomposition, critical path, contract-net, work-stealing, DLQ | DAG 任务分解、关键路径、合同网、工作窃取、死信队列 |
| Tool Resilience | AJV pre-validation, per-tool circuit breaker, retry injection, adaptive repair memory | AJV 预校验、断路器、重试注入、自适应修复记忆 |
| Hierarchical Swarm | Agents can spawn sub-agents within governance bounds (depth + concurrency limits) | 层级蜂群：Agent 可在治理边界内派生子 Agent |
| Lotka-Volterra Ecology | Species competition dynamics with carrying capacity and predator-prey equations | 种群竞争动力学 + 环境容量 + 捕食方程 |
| Response Threshold (FRTM) | Per-agent adaptive thresholds with PI controller for homeostatic task allocation | 固定响应阈值 + PI 控制器自适应任务分配 |
| Skill Symbiosis | Cosine-similarity based complementarity tracking for optimal agent pairing | 余弦相似度互补性追踪，最优 Agent 配对 |
| Swarm Decision Empowerment | V5.3: 9-signal composite aggregation + PI-controller adaptive advisory context injection | V5.3: 9 信号源聚合 + PI 控制器自适应赋能上下文 |
| Adaptive Arbiter (4-State) | V5.4: DIRECT/BIAS_SWARM/PREPLAN/BRAKE mode routing with environment-aware escalation | V5.4: 四态仲裁 + 环境感知升级 |
| Evidence Discipline | V5.4: 3-tier evidence gate (PRIMARY/CORROBORATION/INFERENCE) with weighted scoring | V5.4: 三层证据纪律 + 加权评分 |
| Protocol Semantics | V5.4: 9 typed messages (REQUEST/COMMIT/ACK/...) with conversation tracking and validation | V5.4: 9 种语义消息 + 会话追踪和协议验证 |
| Collaboration Tax | V5.4: 5-dimension budget tracking with per-turn tax computation and per-mode ROI | V5.4: 五维预算追踪 + 协作税 + 按模式 ROI |
| Unified Observability | V5.4: 4-category observation (decision/execution/repair/strategy) with ring buffer | V5.4: 四类观测数据 + 环形缓冲区 |
| State Convergence | V5.5: SWIM failure detection (alive→suspect→dead) + anti-entropy sync with DB as truth source | V5.5: SWIM 故障探测 + 反熵同步 |
| Global Modulator | V5.5: Runtime work-point controller (EXPLORE/EXPLOIT/RELIABLE/URGENT) with hysteresis switching | V5.5: 运行时工作点控制器，滞后切换 |
| Governance Metrics | V5.5: Audit + Policy + ROI triple metrics for swarm governance | V5.5: 审计 + 策略 + ROI 三联治理指标 |
| Three Feedback Loops | V5.5: Strategy/Repair/Environment reflux chains for continuous self-improvement | V5.5: 策略/修复/环境三条回流链 |
| Speculative Execution | V5.6: Parallel candidate execution for critical-path tasks; first completion wins | V5.6: 临界路径任务并行候选执行 |
| DAG-Orchestrator Bridge | V5.6: Shadow plans as DAGs with CPM analysis and bottleneck detection | V5.6: 计划影子化为 DAG + CPM 分析 |
| Work-Stealing | V5.6: Idle agents auto-steal tasks with modulator-aware cooldown | V5.6: 空闲 Agent 自动窃取 + 调节器感知冷却 |
| Skill-Symbiosis Scheduling | V5.7: SkillSymbiosisTracker integrated into ContractNet (5th weight), ExecutionPlanner (4th MoE expert), SwarmAdvisor (6th signal) | V5.7: 共生技能接入 ContractNet/ExecutionPlanner/SwarmAdvisor |
| Multi-Type Pheromones | V5.7: food (linear decay) + danger (step decay) types; typed decay routing via computeTypedDecay(); DB-backed type config |
V5.7: food/danger 类型 + 类型化衰减路由 + DB 配置 |
| 5 Bee Personas | scout, worker, guard, queen-messenger, designer — signal-driven behavior | 5 种蜜蜂人格：侦察蜂、工蜂、守卫蜂、女王信使、设计蜂 |
| Real-Time Dashboard | Fastify + SSE, hex hive view, DAG graph, pheromone particles, RED metrics, breaker status, trace timeline | 实时仪表盘：六边形蜂巢、DAG 图、信息素粒子、RED 指标、断路器状态、追踪时间线 |
| Jaeger-lite Tracing | Lightweight distributed tracing with trace span collector and startup diagnostics | 轻量分布式追踪 span 收集器 + 启动诊断 |
| Plugin SDK Integration | 16 OpenClaw hooks, 8 agent tools, { id, register(api) } pattern |
16 个钩子、8 个工具，标准 Plugin SDK 模式 |
┌─────────────────────────────────────────────────────────────┐
│ L6 Monitoring 监控层 │
│ StateBroadcaster · MetricsCollector · DashboardService │
│ HealthChecker · ObservabilityCore · dashboard-v2.html │
│ TraceCollector · StartupDiagnostics (SSE, port 19100) │
├─────────────────────────────────────────────────────────────┤
│ L5 Application 应用层 │
│ PluginAdapter · ContextService · CircuitBreaker │
│ ToolResilience · SkillGovernor · TokenBudgetTracker │
│ 7 Tool Factories (spawn/query/pheromone/gate/ │
│ memory/plan/zone) │
├─────────────────────────────────────────────────────────────┤
│ L4 Orchestration 编排层 │
│ Orchestrator · CriticalPathAnalyzer · QualityController│
│ PipelineBreaker · ResultSynthesizer · ExecutionPlanner │
│ ContractNet · ReplanEngine · ABCScheduler │
│ RoleDiscovery · RoleManager · ZoneManager │
│ HierarchicalCoordinator · TaskDAGEngine │
│ SpeciesEvolver · SwarmAdvisor · BudgetTracker │
│ GlobalModulator · GovernanceMetrics · SpeculativeExecutor│
├─────────────────────────────────────────────────────────────┤
│ L3 Agent 智能体层 │
│ WorkingMemory · EpisodicMemory · SemanticMemory │
│ ContextCompressor · CapabilityEngine · PersonaEvolution│
│ ReputationLedger · SoulDesigner · SwarmContextEngine │
│ ResponseThreshold · FailureVaccination · EvidenceGate │
│ SkillSymbiosisTracker │
├─────────────────────────────────────────────────────────────┤
│ L2 Communication 通信层 │
│ MessageBus · PheromoneEngine · GossipProtocol │
│ PheromoneTypeRegistry · PheromoneResponseMatrix │
│ StigmergicBoard · ProtocolSemantics · StateConvergence │
├─────────────────────────────────────────────────────────────┤
│ L1 Infrastructure 基础设施层 │
│ DatabaseManager (SQLite, 44 tables) · ConfigManager │
│ MigrationRunner · 8 Repositories · 3 Schema modules │
│ Logger · Types · MonotonicClock │
└─────────────────────────────────────────────────────────────┘
Only L5 couples to OpenClaw via Plugin SDK. Layers L1--L4 and L6 are reusable in any Node.js 22+ environment.
仅 L5 通过 Plugin SDK 与 OpenClaw 耦合。L1--L4 及 L6 可在任何 Node.js 22+ 环境中独立复用。
Note: SwarmContextEngine (L3) currently operates via hook fallback (
buildSwarmContextFallback()). Feature flagcontextEngineis disabled by default.注意： SwarmContextEngine (L3) 当前通过 hook fallback 降级使用。Feature flag
contextEngine默认 disabled。
- Node.js >= 22.0.0 (required for
node:sqliteDatabaseSync) - OpenClaw with Plugin SDK support
npm (Recommended / 推荐):
npm install openclaw-swarm
cd node_modules/openclaw-swarm
node install.js # Register plugin in OpenClaw config / 注册插件到 OpenClaw 配置
openclaw gateway restart # Load the plugin / 加载插件Git clone:
git clone https://github.com/DEEP-IOS/claw-swarm.git
cd claw-swarm
node install.js # One-click setup / 一键安装
openclaw gateway restart # Load the plugin / 加载插件The installer automatically registers the plugin path in ~/.openclaw/openclaw.json and enables it with default configuration.
安装脚本自动在 ~/.openclaw/openclaw.json 中注册插件路径并启用默认配置。
See docs/installation.md for manual installation and advanced options. / 手动安装和高级选项见安装文档。
Plugin-specific settings must be nested under the config key in ~/.openclaw/openclaw.json. The api.pluginConfig receives this object directly.
插件配置必须嵌套在 ~/.openclaw/openclaw.json 的 config 键内。api.pluginConfig 直接接收此对象。
{
"plugins": {
"entries": {
"claw-swarm": {
"enabled": true,
"config": {
"memory": { "enabled": true, "maxPrependChars": 4000 },
"pheromone": { "enabled": true, "decayIntervalMs": 60000 },
"orchestration": { "enabled": true, "maxWorkers": 16 },
"dashboard": { "enabled": false, "port": 19100 }
}
}
}
}
}See docs/installation.md for full option reference. / 完整配置参考见安装文档。
Claw-Swarm requires models with strong tool-calling capabilities. See docs/model-compatibility.md for the full guide.
Claw-Swarm 要求模型具备强工具调用能力。完整指南见 docs/model-compatibility.md。
| Tier | Models / 模型 | Notes / 说明 |
|---|---|---|
| S | Opus 4.6, Sonnet 4.6, GPT-5.4, GPT-5.3-Codex, Gemini 2.5 Pro | Best tool calling + reasoning / 最佳工具调用 + 推理 |
| A | Kimi K2.5, Qwen3.5-Plus/Max, DeepSeek-V3, Gemini 2.5 Flash, o4-mini | Strong with minor trade-offs / 强，少量取舍 |
| B | DeepSeek-Reasoner, GLM-5, Qwen3-Coder-Next, MiniMax-M2.5, Llama 4 Maverick | Usable for specific roles / 特定角色可用 |
| # | Algorithm / 算法 | Layer | Module | Purpose / 用途 |
|---|---|---|---|---|
| 1 | MMAS (Max-Min Ant System) | L2 | PheromoneEngine | Pheromone intensity bounding / 信息素强度上下界 |
| 2 | ACO Roulette (Ant Colony Optimization) | L2 | PheromoneEngine | Probabilistic path selection / 概率路径选择 |
| 3 | Pheromone Pressure Gradient | L2 | PheromoneResponseMatrix | V5.2: Auto-escalation when threshold exceeded / 压力梯度自动升级 |
| 4 | Multi-Type Pheromone Decay | L2 | PheromoneEngine | V5.2: trail(linear), alarm(step), recruit(exp) typed decay / 类型化衰减 |
| 5 | Stigmergic Coordination | L2 | StigmergicBoard | V5.2: Persistent indirect coordination board / 持久间接协调公告板 |
| 6 | Ebbinghaus Forgetting | L3 | EpisodicMemory | Memory decay curve / 记忆遗忘曲线 |
| 7 | BFS Knowledge Graph | L3 | SemanticMemory | Relation traversal / 知识图谱关系遍历 |
| 8 | PARL (Persona A/B) | L3 | PersonaEvolution | Persona evolution via A/B testing / 人格 A/B 进化 |
| 9 | FRTM (Fixed Response Threshold) | L3 | ResponseThreshold | V5.2: Per-agent adaptive thresholds + PI controller / 自适应响应阈值 |
| 10 | Failure Vaccination | L3 | FailureVaccination | V5.2: Pattern immunization with effectiveness tracking / 免疫记忆库 |
| 11 | Cosine Similarity Symbiosis | L3 | SkillSymbiosisTracker | V5.2: Agent complementarity matching / 技能互补配对 |
| 12 | GEP (Gene Expression) | L4 | ExecutionPlanner | Execution plan generation / 执行计划生成 |
| 13 | CPM (Critical Path Method) | L4 | CriticalPathAnalyzer | Task dependency scheduling / 关键路径调度 |
| 14 | Jaccard Dedup | L4 | ResultSynthesizer | Result deduplication / 结果去重 |
| 15 | MoE (Mixture of Experts) | L4 | RoleManager | Expert role routing / 专家角色路由 |
| 16 | FIPA CNP (Contract-Net Protocol) | L4 | ContractNet | Task negotiation / 合同网任务协商 |
| 17 | ABC (Artificial Bee Colony) | L4 | ABCScheduler + SpeciesEvolver | V5.2: Three-stage evolution (employed/onlooker/scout) / 三阶段进化 |
| 18 | k-means++ | L4 | RoleDiscovery | Automatic role clustering / 角色自动发现 |
| 19 | Lotka-Volterra | L4 | SpeciesEvolver | V5.2: Population dynamics dN/dt = rN(1-N/K) - αNP / 种群竞争动力学 |
16 hooks registered via Plugin SDK / 通过 Plugin SDK 注册 16 个钩子：
| Hook | Trigger | Internal Mapping / 内部映射 |
|---|---|---|
gateway_start |
Gateway starting | Engine initialization + config validation / 引擎初始化 + 配置校验 |
before_model_resolve |
Model selection | Model capability auto-detection / 模型能力自动检测 |
before_tool_call |
Tool invocation | ToolResilience AJV validation + circuit breaker / 工具韧性校验 + 断路器 |
before_prompt_build |
Prompt assembly | Tool failure injection + swarm context / 工具失败注入 + 蜂群上下文 |
before_agent_start |
Agent begins | SOUL injection + context prepend (memory, knowledge, pheromone) / SOUL 注入 + 上下文注入 |
agent_end |
Agent finishes | Quality gate + pheromone reinforcement + memory consolidation / 质量门控 + 信息素 + 记忆固化 |
after_tool_call |
Tool completes | Tool resilience + health check + working memory / 工具韧性 + 健康检查 + 工作记忆 |
before_reset |
Conversation reset | Memory consolidation (working → episodic) / 记忆固化 |
gateway_stop |
Gateway shutting down | Engine cleanup + PID file removal / 引擎关闭 + PID 清理 |
message_sending |
Message routed | Agent-to-agent message routing via MessageBus / 消息路由 |
subagent_spawning |
Sub-agent creating | Hierarchical coordinator validation / 层级协调器校验 |
subagent_spawned |
Sub-agent created | Hierarchy tracking / 层级追踪 |
subagent_ended |
Sub-agent finished | Result collection + pheromone update / 结果收集 + 信息素更新 |
llm_output |
LLM response | SOUL.md dual-stage migration / SOUL.md 双阶段迁移 |
Sub-agent lifecycle is managed by the hierarchical coordinator: depth limits, concurrency control, and governance gates are enforced automatically.
子 Agent 生命周期由层级协调器管理：深度限制、并发控制和治理门控自动执行。
8 agent tools registered via Plugin SDK / 通过 Plugin SDK 注册的 8 个智能体工具：
| Tool | Purpose | 用途 |
|---|---|---|
swarm_spawn |
Create and dispatch sub-agents | 创建并调度子智能体 |
swarm_query |
Query swarm state and agent status | 查询蜂群状态 |
swarm_pheromone |
Deposit and read pheromone signals | 发布和读取信息素信号 |
swarm_gate |
Governance gating and capability checks | 治理门控与能力检查 |
swarm_memory |
Read/write agent memory (working/episodic/semantic) | 读写智能体记忆 |
swarm_plan |
Create and manage execution plans | 创建和管理执行计划 |
swarm_zone |
Manage work zones and auto-assignment | 管理工作区与自动分配 |
swarm_run |
V5.3: One-click execution (plan + spawn combined) | V5.3: 一键执行（计划+派生合一） |
| Requirement | Version |
|---|---|
| Node.js | >= 22.0.0 |
| Runtime deps | eventemitter3, fastify, nanoid, pino, zod |
| Dev deps | vitest |
Claw-Swarm maintains a rigorous multi-level testing strategy to ensure production readiness:
Claw-Swarm 采用多层次测试策略确保生产可用性：
| Level | Type | Coverage | 覆盖范围 |
|---|---|---|---|
| Unit | 902 tests across 49 files (vitest) | All 6 layers, every module | 6 层全覆盖 |
| Integration | End-to-end pipeline | Multi-tool workflows, memory persistence, zone governance | 跨工具流程、记忆持久化、Zone 治理 |
| Stress | High-frequency & boundary | 20+ rapid calls, WAL concurrency, edge cases | 高频调用、并发写入、边界值 |
| Production | 20 tests in live OpenClaw Gateway | Plugin load, tool invocation, MMAS, memory, quality gate, MoE, integration scenarios, stress | 真实 Gateway 环境全链路验证 |
| Install | Clean-environment install test | Clone → install.js → gateway restart → tool invocation on Linux | 干净 Linux 环境一键安装全流程验证 |
The production test suite validates the plugin end-to-end in a live OpenClaw Gateway environment — not mocked, not simulated. All 20 production tests passed with 7 bugs discovered and fixed during testing. See the full report: Production Test Report
The install test was independently conducted on a clean Linux (Node.js v22, OpenClaw 2026.2.13) environment — from git clone to swarm_query in under 3 minutes, 100% pass rate with 0 blocking issues. See: Install Test Report
生产测试套件在真实 OpenClaw Gateway 环境中端到端验证插件 — 非 mock、非模拟。20 项生产测试全部通过，测试过程中发现并修复了 7 个 bug。完整报告见：生产测试报告
安装测试在干净 Linux 环境（Node.js v22, OpenClaw 2026.2.13）中独立执行 — 从 git clone 到 swarm_query 调用成功仅需 3 分钟，100% 通过率，零阻断性问题。报告见：安装测试报告
# All tests (902 tests, 49 files) / 全部测试
npm test
# By category / 按类别
npm run test:unit
npm run test:integration
npm run test:stress
# By layer / 按层级
npm run test:L1
npm run test:L2
npm run test:L3
npm run test:L4
npm run test:L5
npm run test:L6
# Watch mode / 监听模式
npm run test:watch
# Coverage report / 覆盖率
npm run test:coveragesrc/
├── index.js # Plugin entry { id, register(api) }
│ # 插件入口
├── L1-infrastructure/ # 基础设施层 (18 files)
│ ├── types.js # Type definitions / 类型定义
│ ├── logger.js # Pino-based logging / 日志
│ ├── monotonic-clock.js # V5.1: hrtime monotonic timing
│ ├── config/
│ │ └── config-manager.js # Zod-validated config / 配置管理
│ ├── database/
│ │ ├── database-manager.js # SQLite DatabaseSync (44 tables)
│ │ ├── migration-runner.js # Schema migrations / 迁移
│ │ ├── sqlite-binding.js # node:sqlite binding
│ │ └── repositories/ # 8 data repositories
│ │ ├── agent-repo.js
│ │ ├── episodic-repo.js
│ │ ├── knowledge-repo.js
│ │ ├── pheromone-repo.js
│ │ ├── pheromone-type-repo.js
│ │ ├── plan-repo.js
│ │ ├── task-repo.js
│ │ └── zone-repo.js
│ └── schemas/
│ ├── config-schemas.js # Config Zod schemas
│ ├── database-schemas.js # DB table schemas
│ └── message-schemas.js # Message format schemas
│
├── L2-communication/ # 通信层 (7 files)
│ ├── message-bus.js # Pub/sub + wildcards + DLQ
│ ├── pheromone-engine.js # MMAS bounds, typed decay
│ ├── gossip-protocol.js # Epidemic broadcast + heartbeat
│ ├── pheromone-type-registry.js # Custom pheromone types
│ ├── pheromone-response-matrix.js # V5.2: Pressure gradient + auto-escalation
│ ├── stigmergic-board.js # V5.2: Persistent bulletin board
│ └── protocol-semantics.js # V5.4: 9 semantic message types
│
├── L3-agent/ # 智能体层 (13 files)
│ ├── memory/
│ │ ├── working-memory.js # 3-tier: focus/context/scratchpad
│ │ ├── episodic-memory.js # Ebbinghaus forgetting curve
│ │ ├── semantic-memory.js # BFS knowledge graph
│ │ └── context-compressor.js # LLM context compression
│ ├── capability-engine.js # 4D capability scoring
│ ├── persona-evolution.js # PARL A/B testing
│ ├── reputation-ledger.js # Agent reputation tracking
│ ├── soul-designer.js # 5 bee persona templates
│ ├── swarm-context-engine.js # V5.1: Rich context builder
│ ├── response-threshold.js # V5.2: FRTM + PI controller
│ ├── failure-vaccination.js # V5.2: Pattern immunization
│ ├── skill-symbiosis.js # V5.2: Cosine complementarity
│ └── evidence-gate.js # V5.4: 3-tier evidence discipline
│
├── L4-orchestration/ # 编排层 (18 files)
│ ├── orchestrator.js # DAG task decomposition
│ ├── critical-path.js # CPM scheduling
│ ├── quality-controller.js # Multi-rubric quality gate
│ ├── pipeline-breaker.js # State machine breaker
│ ├── result-synthesizer.js # Jaccard deduplication
│ ├── execution-planner.js # GEP plan generation
│ ├── contract-net.js # FIPA CNP negotiation
│ ├── replan-engine.js # Pheromone-triggered replan
│ ├── abc-scheduler.js # Artificial Bee Colony
│ ├── role-discovery.js # k-means++ clustering
│ ├── role-manager.js # MoE expert routing
│ ├── zone-manager.js # Jaccard auto-assign
│ ├── hierarchical-coordinator.js # V5.1: Hierarchical swarm
│ ├── task-dag-engine.js # V5.1: DAG + work-stealing + DLQ
│ ├── species-evolver.js # V5.1+V5.2: Species evolution + GEP + LV + ABC
│ ├── swarm-advisor.js # V5.3+V5.4: Decision empowerment + 4-state arbiter
│ └── budget-tracker.js # V5.4: 5-dimension collaboration tax
│
├── L5-application/ # 应用层 (14 files)
│ ├── plugin-adapter.js # Engine lifecycle manager
│ ├── context-service.js # Rich LLM context builder
│ ├── circuit-breaker.js # 3-state circuit breaker
│ ├── tool-resilience.js # V5.1: AJV + per-tool breaker
│ ├── skill-governor.js # V5.1: Skill inventory + tracking
│ ├── token-budget-tracker.js # V5.1: 800-token budget coord
│ └── tools/
│ ├── swarm-spawn-tool.js
│ ├── swarm-query-tool.js
│ ├── swarm-pheromone-tool.js
│ ├── swarm-gate-tool.js
│ ├── swarm-memory-tool.js
│ ├── swarm-plan-tool.js
│ ├── swarm-zone-tool.js
│ └── swarm-run-tool.js # V5.3: One-click execution
│
├── event-catalog.js # V5.1-V5.4: 46 EventTopics + schema
│
└── L6-monitoring/ # 监控层 (7 files)
├── state-broadcaster.js # SSE push to clients
├── metrics-collector.js # RED metrics (Rate/Errors/Duration)
├── dashboard-service.js # Fastify HTTP + /v2 API + trace spans
├── dashboard.html # Dark theme web dashboard
├── dashboard-v2.html # V5.1: Hex hive + DAG + particles
├── health-checker.js # V5.1+V5.2: Health + idle detection
└── observability-core.js # V5.4: 4-category unified observability
tests/
├── unit/
│ ├── L1/ (4 files) # Infrastructure tests
│ ├── L2/ (6 files) # Communication tests (+V5.4)
│ ├── L3/ (9 files) # Agent tests (+V5.4)
│ ├── L4/ (15 files) # Orchestration tests (+V5.4)
│ ├── L5/ (5 files) # Application tests (+V5.3)
│ └── L6/ (5 files) # Monitoring tests (+V5.4)
├── integration/ (1 file) # Full pipeline tests
└── stress/ (legacy) # Stress/edge-case tests
MIT License. Copyright 2025-2026 DEEP-IOS.
See LICENSE for full text.
