- Implement dual-mode streaming: prioritize in-memory EventManager for running tasks with thinking_token support
- Add fallback to database polling for completed tasks without thinking_token replay capability
- Introduce SSE event formatter utility for consistent event serialization across streaming modes
- Add 10ms micro-delay for thinking_token events to ensure proper TCP packet separation and frontend incremental rendering
- Refactor stream_agent_with_thinking endpoint to support both runtime and historical event streaming
- Update event filtering logic to handle both in-memory and database event sources
- Improve logging with debug markers for thinking_token tracking and stream mode selection
- Optimize polling intervals: 0.3s for running tasks, 2.0s for completed tasks
- Reduce idle timeout from 10 minutes to 1 minute for completed task streams
- Update frontend useAgentStream hook to handle unified event format from dual-mode streaming
- Enhance AgentAudit UI to properly display streamed events from both sources
- Introduce streaming capabilities for agent events, allowing real-time updates during audits.
- Add new hooks for managing agent stream events in React components.
- Enhance the AgentAudit page to display LLM thinking processes and tool call details in real-time.
- Update API endpoints to support streaming event data and improve error handling.
- Refactor UI components for better organization and user experience during audits.
vinland100
lintsinghua