Community Template Fix Corrects GPT-OSS Bug That Dropped Answers in Multi-Turn Chats

A user on Hugging Face released an updated Jinja template for GPT-OSS to address a critical bug present in the Unsloth version of the template. The bug occurred when chat history included messages with both reasoning (thinking) and final answers (content); the template incorrectly rendered only the reasoning, dropping the answer. This caused significant confusion and incoherence in multi-turn conversations, particularly for the GPT-OSS 20B model, which could 'go completely off the rails.' The GPT-OSS 120B model appeared more robust, recovering context from reasoning traces alone, but the underlying template error remained. The new template corrects the rendering logic to include both reasoning and answers. It also introduces a preserve_thinking feature, which retains reasoning traces across turns to potentially improve inference speed via prefix caching, at the cost of higher token usage. The author notes that OpenAI's reference template does not contain this erroneous copy-paste logic.