- Add INativeToolCall interface for native tool call format
- Add useNativeToolCalling option to IDualAgentOptions
- Add getToolsAsJsonSchema() to convert tools to Ollama JSON Schema format
- Add parseNativeToolCalls() to convert native tool calls to proposals
- Add startTaskWithNativeTools() and continueWithNativeTools() to DriverAgent
- Update DualAgentOrchestrator to support both XML parsing and native tool calling modes
Native tool calling is more efficient for models like GPT-OSS that use Harmony format,
as it activates Ollama's built-in tool parser instead of requiring XML generation.
The system message now clearly states that the <tool_call> XML tags MUST
be literally written in the response, not just described. Includes examples
of CORRECT vs WRONG usage to help smaller models understand.
When the LLM fails to emit a tool_call XML block, the feedback now includes
the exact XML format expected with a concrete example for json.validate.
This helps smaller models understand the exact output format required.
Each tool now provides comprehensive documentation including parameter
schemas and concrete <tool_call> XML examples. This helps smaller LLMs
understand the exact format required for tool invocation.
- Add onToken callback option to IDualAgentOptions interface
- Add onToken property and setOnToken method to DriverAgent
- Wire up streaming in startTask and continueWithMessage methods
- Pass source identifier ('driver'/'guardian') to onToken callback
