Ubicloud reports that Kimi K2.6 served through vLLM 0.28 achieved 92% tool-call schema accuracy, rising to 100% in its K2VV checks after constrained decoding was enabled. A server-side option in vLLM 0.31 restricts invalid call structures during generation. Documentation distinguishes the function level, which constrains the call envelope and function name, from parameter, which also enforces argument schemas. Operators must check parser support and configuration; valid formatting does not guarantee appropriate tool selection, meaningful argument values or task completion.