Request Flow in AgentDock Core

This document outlines the typical sequence of events when a request is processed by AgentDock Core.

Request Flow Diagram

Detailed Steps

1. Request Initiation: A client sends a request (e.g., a user message, potentially including a sessionId) to an API endpoint.
2. API Endpoint Handling: Extracts the sessionId and message payload, determining the target agentId.
3. Agent Instantiation (AgentNode): Creates an instance of AgentNode, passing the agent configuration, API keys, and core managers.
4. State Retrieval / Initialization: Loads or creates orchestration state for the session.
5. Orchestration & Tool Filtering: Evaluates conditions to determine the activeStep and filters available tools.
6. LLM Interaction (CoreLLM.streamText): Prepares the prompt, sets callbacks, and calls the LLM with filtered tools.
7. Response Streaming & Tool Handling:
   • LLM streams response (text chunks or tool call requests)
   • Tool calls are executed and results returned to the LLM
   • Text chunks are streamed back to the client
8. State Updates: Update token usage, tool history, sequence index, and timestamps.
9. Response Completion: Stream ends and response is finalized.
10. Cleanup: AgentNode instance is discarded while session state persists.

For implementation details, see agentdock-core/src/nodes/agent-node.ts.

Enhanced Streaming Flow

AgentDock extends standard streaming with AgentDockStreamResult:

This enhanced flow provides:

1. Automatic State Management: Updates token usage and tracks tools used
2. Error Handling: Propagates errors from LLM providers with better context
3. Tool Orchestration: Manages tool execution and updates session state

For details on streaming implementation, see Response Streaming.