This repo contains a single Jinja chat template file that fixes rendering errors, KV cache invalidation, and agentic stalling in official Qwen 3.5, 3.6, and 3.8 templates.

It works across LM Studio, llama.cpp, vLLM, MLX, oMLX, and KoboldCPP, adding reasoning effort kwargs, tool argument parsing, and inline per-turn steering tags. All filters are rewritten to be minijinja-safe for C++ inference engines, replacing Python-specific logic from the official templates.