Mojentic 1.5.0 — Realtime Voice
A milestone release. Mojentic 1.5.0 introduces the Realtime Voice subsystem — a major new capability that sits alongside the existing LlmBroker to support low-latency voice and streaming sessions over OpenAI's Realtime API. The release is fully backward-compatible: Realtime Voice is purely additive, and nothing in the existing API changed.
What's New
RealtimeVoiceBrokerandMojentic.LLM.Gateways.OpenAIRealtime— WebSocket-based gateway for OpenAI's Realtime API, enabling low-latency voice and streaming sessionsMojentic.LLM.ToolRunnerabstraction with a serial default — a composable interface for executing tool calls outside the broker's built-in recursion loop- GPT-5.4 and GPT-5.5 model families registered in the OpenAI model registry with their real 1,050,000 / 400,000-token context windows and 128,000-token output cap; pattern mappings for
gpt-5.3/gpt-5.4/gpt-5.5added - Realtime and parallel-tools examples with accompanying guides
Fixed
- Race condition in
SimpleRecursiveAgent.process_iterationthat could cause incorrect completion detection under concurrent load - Realtime session blockers and dead code cleaned up in the OpenAI Realtime gateway
See CHANGELOG.md for full details.