Architecture

Yunque Agent is a modular AI agent runtime with 40+ core modules, all wired in cmd/agent/main.go. It uses a "control plane + execution plane" dual-layer architecture, decoupling the cognitive engine from external integrations.

Request Flow

HTTP / WebSocket / IM Channel → Gateway (Auth · CORS · Rate-limit · RBAC)
  → Guardrails (Input: PII redaction · Injection detection · Shell Guard)
  → Session Manager (History load · Fork/Branch)
  → Smart Router (Model selection by complexity: fast / smart / expert)
  → Cost Tracker (Budget check · Real-time billing)
  → Context Manager (Token budget · Compression · Sliding window)
  → Planner (LLM planning loop + Memory + Knowledge RAG + Skill orchestration)
  → Guardrails (Output: PII redaction · Content filter)
  → Memory Pipeline (Async fact extraction → 5-layer memory)
  → Audit Trail (Merkle immutable log)
  → Reflection Loop (Experience accumulation → Strategy compilation)

Core Modules

Cognitive Layer (agentcore)

Module	Function
llm	LLM client, streaming, provider registry, circuit breaker
planner	Multi-step planning loop, skill orchestration, emotion-aware, Reverie
memory	5-layer memory (short/mid/long/editable/knowledge graph)
knowledge	Hybrid RAG: BM25 + vector + RRF + Rerank
skills	Skill registry, execution, parameter validation
skillmarket	Skill installation, ClawHub/ToriHub market, SkillFileLoader
emotion	Emotion analysis, history, sticker mapping
persona	Persona chain, feature switches, templates
task	Task lifecycle, thread management, working memory
workflow	DAG workflow engine, conditional branching, parallel execution

Security & Governance

Module	Function
guardrails	Content safety, PII redaction, injection detection, Shell Guard
audit	Merkle chain immutable audit log
trust	Trust score system, progressive permissions
approval	High-risk operation approval, 4-tier risk classification
rbac	Role-based access control, 4-level inheritance

Execution Layer

Module	Function
channel	16+ IM channel adapters with resilience wrappers
sandbox	Code sandbox (process/Docker/K8s/Wasm), E2B Desktop
scheduler	Task scheduler, Q-Learning optimization

Directory Structure

internal/
├── agentcore/        # 40+ cognitive modules
├── controlplane/     # HTTP/WS API gateway
├── execution/        # Sandbox, scheduler, 16+ channel adapters
├── experimental/     # Reflection, distillation, causal chains
├── storage/          # SQLite storage layer
├── ledger/           # KV storage + persistence bridge
cmd/agent/            # Entry point
plugins/              # Go skill plugins
pkg/                  # Shared libraries (plugin loader, skills)
data/
├── plugins/          # Script plugins (Python/Node/Shell)
├── skills/           # File-based skills (SKILL.md)
└── knowledge/        # Knowledge base documents
heroui-web/           # HeroUI v3 + Next.js dashboard
docs/                 # VitePress documentation