feat(realtime): configurable pipeline.max_history_items

[localai-bot]

Problem

Composed realtime pipelines (VAD+STT+LLM+TTS) get MaxHistoryItems = 0 (unlimited) by default — only self-contained any-to-any models (liquid-audio) get the 6-item cap. So a long-running pipeline session grows every turn and feeds the entire conversation to the LLM until its context window fills. There's no way to cap it from config.

Fix

Add an optional pipeline.max_history_items to the model config. resolveMaxHistoryItems uses the explicit value when set (including 0 = unlimited), otherwise falls back to defaultMaxHistoryItems (6 for realtime-audio, 0 for legacy pipelines). Wired into session init; the existing trimRealtimeItems does the trimming.

name: gpt-realtime
pipeline:
  vad: silero-vad-ggml
  transcription: parakeet-cpp-tdt-0.6b-v3
  llm: gemma-4-e2b-it-qat-q4_0
  tts: vits-piper-it_IT-paola-sherpa
  max_history_items: 20   # cap trailing items fed to the LLM per turn

Tests cover the resolver (explicit value, explicit 0, fallback-to-default, nil). Existing defaultMaxHistoryItems/trimRealtimeItems tests unchanged.

🤖 Generated with Claude Code