- Update smartai to v0.13.2 (native tool calling support)
- Update smartagent to v1.5.1 (useNativeToolCalling option)
- Enable think: true for GPT-OSS reasoning mode in Ollama config
- Enable useNativeToolCalling: true in DualAgentOrchestrator
- Simplify driver system message (native tools don't need XML instructions)
Native tool calling uses Ollama's built-in Harmony format parser
instead of requiring XML generation, which is more efficient for GPT-OSS models.
- Add tryExtractJson function to validate JSON before accepting
- Use orchestrator.continueTask() to request correction if JSON is invalid
- Retry up to 2 times for malformed JSON responses
- Remove duplicate parseJsonToInvoice function
- Update test.invoices.nanonets.ts to use DualAgentOrchestrator for JSON extraction
- Enable streaming token callback for real-time progress visibility
- Add markdown caching to avoid re-running Nanonets OCR for cached files
- Update test.bankstatements.minicpm.ts and test.invoices.minicpm.ts with streaming
- Update dependencies to @push.rocks/smartai@0.11.1 and @push.rocks/smartagent@1.2.8
