Skip to content

v1.0.6

Choose a tag to compare

View all tags
@AlbumenJ AlbumenJ released this 05 Jan 13:01
v1.0.6
0bae8dc

This release introduces significant enhancements to Human-in-the-Loop (HITL) interaction patterns and optimizes the streaming experience for tool usage. It also includes new integrations (Nacos), API simplifications, and various bug fixes.

🌟 Key Highlights

1. Comprehensive Human-in-the-Loop (HITL) Support

We have refactored the execution flow to support robust HITL scenarios. You can now suspend agent execution to await human input, confirmation, or modification, and then resume the process seamlessly.

  • Mechanism: HITL is achieved via stopAgent() within PostReasoningEvent or PostActingEvent.
  • External Tools: Added support for external tool execution with a suspension mechanism, allowing agents to "pause" while a tool (or human) performs an action asynchronously.
  • Documentation: For a deep dive on implementation, please refer to the HITL Documentation.
  • Example: Check out the new chat example with MCP tools and tool confirmation.

2. Enhanced Streaming: Incremental Tool Outputs

We have upgraded the streaming capabilities to support Real-time tool-use block emitting.

  • Impact: ReasoningChunkEvent now supports incremental returns for Tools.
  • For Developers: Please pay attention to the new output types in the stream. This feature allows for significant UI optimizations, such as displaying tool arguments or intermediate statuses to the user character-by-character as they are generated, rather than waiting for the full tool call to complete.

🚀 New Features

  • Nacos Integration: Added AgentScope Nacos Spring Boot starter and A2A integration (#446).
  • MCP Support: Added support for Model Context Protocol (MCP) builder configuration in the main config (#452).
  • Model API Simplification: Removed OpenAIConfig and simplified the OpenAIChatModel API for better usability (#427).
  • Customization:
  • Support for reasoningEffort field in OpenAIChatFormatter (#444).
  • Support for customized tool result converters (#413).
  • Support for customizing role numbers in the Werewolf example (#417).
  • HTTP Transport: Switched default HTTP transport to JDK and enhanced its capabilities (#445).

🛠️ Refactoring & Fixes

  • Tools: Added Toolkit deep copy to isolate agent state (#438).
  • JSON: Centralized JSON processing with JsonCodec abstraction (#439).
  • Formatters: Extracted provider-specific logic into dedicated formatters (#428) and fixed OpenAI streaming chunk type detection (#436).
  • Bug Fixes:
  • Fixed handling of TextBlock in the last REASONING event (#419).
  • Fixed output attribute after callTool response (#443).
  • Fixed ToolResultBlock transient field issues (#455).

❤️ New Contributors

Full Changelog: v1.0.5...v1.0.6

Contributors

hujun-w-2, TornadoDragon, and 3 other contributors
Assets 2
Loading