This release candidate brings significant enhancements to the agent system, including a new skills system that allows more flexible tool usage. Agent semantics and user experience have been cleaned up, removing dead code and improving the handling of slash commands—commands are now completed before being submitted. The terminal interface now renders bold emphasis consistently across markdown, and redundant context-window refreshes in the event loop have been eliminated for better performance. Cloud model users will benefit from unlimited tool rounds enabled by default.
Backend improvements include updates to the llama.cpp inference engine and the MLX framework, ensuring better model compatibility and performance. The Anthropic integration now correctly closes text blocks before starting a thinking block, and Hermes integration has been refined. Support for Claude Code channels has been preserved to maintain compatibility.
On the build and infrastructure side, the Linux toolchain has been bumped to GCC 13, and CUDA v12 on Linux now includes compute capability 10.0, expanding GPU support. The server startup process now routes through a heartbeat check for more reliable initialization.
