b8656

The open-source llama.cpp project, maintained by ggml-org, has pushed a significant technical update with commit b8656. This release focuses on fixing a core component of AI agent functionality: tool calling. Specifically, it resolves bugs in how the system detects and handles data types for function arguments when those arguments are defined as 'nullable' (can be a value or null) or 'enum' (must be one of a specific set of values). The fix ensures the underlying grammar parser correctly interprets complex JSON schemas, which is essential for models that act as agents and need to call external functions or APIs reliably.

The update includes targeted fixes for the 'TAG_WITH_TAGGED' format used by models like Qwen3-Coder, allowing them to process raw text arguments correctly instead of requiring strictly formatted JSON strings. The commit also adds comprehensive test cases to validate the fixes for scenarios like nullable strings, nullable integers, and typeless enums. This is a foundational improvement for developers using llama.cpp to run local AI models in agentic workflows, where precise schema parsing is non-negotiable for successful tool execution across various hardware backends, from Apple Silicon to CUDA.