Daita Documentation

#[0.3.0] - 2025-11-25

Universal Streaming Support for All LLM Providers
- Real-time text streaming across OpenAI, Anthropic, Grok, and Gemini providers
- Tool call streaming with complete argument streaming
- Unified LLMChunk format across all providers (types: "text", "tool_call_complete")
- Automatic token usage tracking during streaming
- Model metadata included in each chunk
- Multi-turn conversation support with streaming for all providers
Agent Streaming Events (on_event callback)
- Real-time execution monitoring via agent.run(prompt, on_event=callback)
- Six event types for comprehensive visibility:
  - ITERATION: Track multi-step reasoning iterations
  - THINKING: Stream LLM text generation in real-time
  - TOOL_CALL: Monitor tool invocations with arguments
  - TOOL_RESULT: Receive tool execution results
  - COMPLETE: Get final answer with cost and token metadata
  - ERROR: Handle execution errors gracefully
- Support for both sync and async event handlers
- Works with both run() and run_detailed() methods
- Zero-configuration - just add on_event parameter
Streaming Examples and Patterns
- Progress tracking with custom UI handlers
- Async event handling for database logging and notifications
- Buffered text updates for optimal UI performance
- Tool execution monitoring for debugging
- Cost and performance tracking during execution

Gemini Streaming Tool Calling
- Fixed critical issue where Gemini's streaming API would return empty tool calls in multi-turn conversations
- Added defensive validation to filter empty tool calls before they poison conversation history
- Ensures reliable multi-turn tool conversations with Gemini
- Applied validation to both async and sync streaming paths
- Gemini now fully supports text and tool streaming with multi-turn conversations