@skynetxbt/flow-acp

ACP Agent Log Streaming

The ACP agents now stream LangChain execution logs to WebSocket clients in real-time.

How it works

• Log Streaming: Each agent (buyer/seller) has a LogStreamer that captures LangChain events
• Event Publishing: Logs are published to the swarm event bus as AGENT_LOG events
• WebSocket Forwarding: The WebSocket manager automatically forwards all events to connected clients

Log Event Format

WebSocket clients will receive log events with this structure:

{
  "type": "AGENT_LOG",
  "sessionId": "user-session-id",
  "userPublicKey": "0x...",
  "timestamp": "2024-01-01T12:00:00.000Z",
  "source": "acp_agent",
  "data": {
    "agentType": "seller",
    "level": "info",
    "message": "Chain started: AgentExecutor",
    "data": {
      "inputs": { "input": "..." }
    }
  }
}

Log Levels

• info: Chain starts/ends, tool starts/ends, job events
• error: Chain errors, tool errors
• debug: LLM starts/ends (verbose)

Example WebSocket Client

const ws = new WebSocket('ws://localhost:3000/ws?sessionId=your-session&apiKey=your-key&signature=your-signature&timestamp=123456');

ws.onmessage = (event) => {
  const data = JSON.parse(event.data);
  
  if (data.type === 'AGENT_LOG') {
    console.log(`[${data.data.agentType}] ${data.data.level}: ${data.data.message}`);
    if (data.data.data) {
      console.log('Additional data:', data.data.data);
    }
  }
};

What gets logged

Chain Events

• When LangChain agent executor starts/ends
• Input parameters and outputs
• Any chain-level errors

Tool Events

• When tools like respond_to_job, deliver_job start/end
• Tool inputs and outputs
• Tool execution errors

LLM Events (debug level)

• LLM API calls start/end
• Prompts sent to LLM
• LLM responses

Job Events

• New jobs received
• Job state changes
• Job completion

This provides real-time visibility into agent decision-making and execution.