Core Modules

These pages document the Python runtime modules that assemble a conversation session. The source is package-based: the public import normally comes from each package's __init__.py, while the implementation is split into focused mixins and provider workers.

Runtime map

Module	Current source	Responsibility
LLMSessionManager	main_logic/core/	Owns session lifecycle and coordinates input, model, memory, tools, and speech output
Realtime Client	main_logic/omni_realtime_client/	Runs native audio/realtime sessions over provider-specific transports
Offline Client	main_logic/omni_offline_client/	Runs streamed text and vision turns against chat-completion APIs
TTS Client	main_logic/tts_client/	Resolves an external TTS worker and exposes the worker queue contract
Config Manager	utils/config_manager/	Resolves runtime storage, migrations, character data, API profiles, and persisted settings

How they compose

LLMSessionManager reads normalized settings through ConfigManager, then selects one conversation client from the input mode:

• text input creates OmniOfflineClient;
• audio input creates OmniRealtimeClient;
• external speech output additionally starts the worker returned by get_tts_worker();
• native-audio realtime providers send their own audio and do not need the external TTS path.

The offline client is therefore not an automatic failure fallback for the realtime client. Switching between them is a session-mode decision made by the manager.

Execution boundaries

• Configuration and most persistence methods are synchronous filesystem operations. Async callers must use the provided a* wrappers where available or offload work explicitly.
• Both conversation clients expose async methods. Realtime keeps a persistent transport and background receive tasks; offline performs a streamed request per turn.
• External TTS runs in a dedicated worker thread. The session manager bridges async model output to that thread with request and response queues.

Start with LLMSessionManager for the call path, then use the client and TTS pages for provider-specific behavior.