#Agentic Streaming

Creatoria Agent supports advanced agentic streaming features, including thinking chains (Thinking), automatic planning (Planning), step execution (Step Execution), and tool calling (Tool Calling).

#Overview

When using streaming mode (stream: true) for chat, the system automatically enables the following advanced features based on the Agent's configuration and the hint parameter in the request:

CopyUser Message -> Thinking Chain -> Automatic Planning -> Step Execution (with Tool Calls) -> Final Reply

#Enabling Agentic Features

Control via the hint parameter in the Chat API request:

jsonCopy{
  "message": "Help me analyze the technical architecture of this project",
  "stream": true,
  "hint": {
    "planning": "auto",
    "onStepFailure": "continue"
  }
}

#hint Parameters

#SSE Event Types

#Thinking Chain Event (thinking)

Displays the LLM's reasoning process:

jsonCopy{
  "sessionId": "sess_xxx",
  "thinking": {
    "content": "The user is asking about technical architecture, I need to..."
  }
}

Thinking chain content is streamed incrementally. The content from multiple thinking events should be concatenated.

#Planning Event (plan)

When the LLM decides planning is needed, it generates an execution plan:

jsonCopy{
  "sessionId": "sess_xxx",
  "plan": {
    "steps": [
      {
        "id": "step_1",
        "title": "Search relevant documents",
        "description": "Search for technical architecture-related documents in the knowledge base"
      },
      {
        "id": "step_2",
        "title": "Analyze architecture patterns",
        "description": "Analyze the tech stack and design patterns used in the project"
      },
      {
        "id": "step_3",
        "title": "Output analysis report",
        "description": "Compile and output a complete technical architecture analysis"
      }
    ]
  }
}

#Step Start Event (stepStart)

jsonCopy{
  "sessionId": "sess_xxx",
  "stepStart": {
    "stepId": "step_1",
    "title": "Search relevant documents",
    "index": 0
  }
}

#Step Complete Event (stepComplete)

jsonCopy{
  "sessionId": "sess_xxx",
  "stepComplete": {
    "stepId": "step_1",
    "status": "completed",
    "summary": "Found 3 relevant documents"
  }
}

#Tool Call Event (toolCall)

When the Agent calls a tool:

jsonCopy{
  "sessionId": "sess_xxx",
  "toolCall": {
    "id": "tc_xxx",
    "name": "web-search",
    "arguments": {
      "query": "NestJS microservice architecture"
    }
  }
}

#Tool Result Event (toolResult)

jsonCopy{
  "sessionId": "sess_xxx",
  "toolResult": {
    "toolCallId": "tc_xxx",
    "name": "web-search",
    "success": true,
    "content": "Search result content...",
    "duration": 1200
  }
}

#Content Delta Event (delta)

Normal incremental text content:

jsonCopy{
  "sessionId": "sess_xxx",
  "delta": {
    "content": "Based on the analysis, this project uses"
  }
}

#Usage Event (usage)

jsonCopy{
  "sessionId": "sess_xxx",
  "usage": {
    "inputTokens": 500,
    "outputTokens": 1200,
    "totalTokens": 1700
  }
}

#Done Event (done)

jsonCopy{
  "sessionId": "sess_xxx",
  "done": true,
  "sources": [
    {
      "documentId": "doc_xxx",
      "title": "Architecture Document",
      "score": 0.95
    }
  ],
  "usage": {
    "inputTokens": 500,
    "outputTokens": 1200,
    "totalTokens": 1700
  }
}

#Error Event (error)

jsonCopy{
  "sessionId": "sess_xxx",
  "error": "Tool execution failed: timeout"
}

#Complete Event Stream Example

Copydata: {"sessionId":"sess_xxx","thinking":{"content":"The user wants to know about..."}}

data: {"sessionId":"sess_xxx","thinking":{"content":"I need to handle this step by step"}}

data: {"sessionId":"sess_xxx","plan":{"steps":[{"id":"step_1","title":"Search documents"},{"id":"step_2","title":"Analyze and summarize"}]}}

data: {"sessionId":"sess_xxx","stepStart":{"stepId":"step_1","title":"Search documents","index":0}}

data: {"sessionId":"sess_xxx","toolCall":{"id":"tc_1","name":"web-search","arguments":{"query":"..."}}}

data: {"sessionId":"sess_xxx","toolResult":{"toolCallId":"tc_1","name":"web-search","success":true,"content":"...","duration":800}}

data: {"sessionId":"sess_xxx","stepComplete":{"stepId":"step_1","status":"completed","summary":"Found relevant information"}}

data: {"sessionId":"sess_xxx","stepStart":{"stepId":"step_2","title":"Analyze and summarize","index":1}}

data: {"sessionId":"sess_xxx","delta":{"content":"Based on the search results, "}}

data: {"sessionId":"sess_xxx","delta":{"content":"the project's architecture..."}}

data: {"sessionId":"sess_xxx","stepComplete":{"stepId":"step_2","status":"completed"}}

data: {"sessionId":"sess_xxx","done":true,"usage":{"inputTokens":500,"outputTokens":800,"totalTokens":1300}}

data: [DONE]

#Agentic Data in History Messages

When reopening a session, the metadata field of history messages contains the full agentic data:

jsonCopy{
  "role": "assistant",
  "content": "The final reply content...",
  "metadata": {
    "thinking": "Full thinking chain content",
    "plan": {
      "steps": [...]
    },
    "stepResults": [
      {
        "stepId": "step_1",
        "title": "Search documents",
        "status": "completed",
        "summary": "Found relevant documents",
        "toolCalls": [
          {
            "id": "tc_1",
            "name": "web-search",
            "arguments": {"query": "..."},
            "status": "success",
            "result": "Search results...",
            "duration": 800
          }
        ]
      }
    ]
  }
}

#Tool Types

Agents can use the following types of tools:

#Built-in Tools

Controlled via the Agent's settings.toolsEnabled and settings.allowedToolCategories.

#Custom Tools

Associate custom code tools via settings.customToolIds. These tools execute user-defined code in a sandboxed environment.