Phase 5 / Ep 24: Plugin Pipeline Architecture —— Onion Model for Message Processing

🎯 Learning Objective: Understand the 'Onion Model' of Plugins in the message pipeline.

1. Onion Model

Before a message reaches the Agent and before a response is returned to the user, it will sequentially pass through all registered Plugins:

sequenceDiagram
    participant User as 🧑 User Message
    participant P1 as Plugin A\n📝 Logging
    participant P2 as Plugin B\n🛡️ Sensitive Word Filtering
    participant P3 as Plugin C\n🏷️ Intent Classification
    participant Agent as 🤖 Agent
    participant P4 as Plugin D\n🎨 Response Formatting

    User->>P1: ① Message enters pipeline
    P1->>P2: ② Pass after logging
    P2->>P2: ③ Check sensitive words
    P2->>P3: ④ Pass after approval
    P3->>Agent: ⑤ Attach intent tag
    Agent->>P4: ⑥ Agent response
    P4->>User: ⑦ Return to user after formatting

2. Pipeline Execution Order

Plugins are executed sequentially according to their registration order (array order in openclaw.json):

{
  "plugins": {
    "pipeline": [
      "logger",           // 1st to execute
      "content-filter",   // 2nd
      "intent-classifier", // 3rd
      "response-formatter" // Last (response phase)
    ]
  }
}

3. Plugin Hook Types

4. Pipeline Control

Plugins can decide whether to continue passing the message:

• next()：Pass to the next Plugin
• block()：Intercept the message, no longer pass it
• modify(msg)：Modify the message and then pass it

Next Episode Preview: Ep 25, develop your first Plugin from scratch—the 'Message Counter'.