llama.cpp Release b10271 Introduces UI Improvements for Agent Working Directories
The llama.cpp project released version b10271, which introduces a new per-conversation Current Working Directory (CWD) picker for AI agents. This update includes path navigation, search scope enhancements, and support for Windows drive roots and UNC shares. As llama.cpp evolves to support more agentic workflows, providing a robust UI for managing directory access helps developers and users safely control where local AI agents execute tools and read or write files. The update treats path-like queries starting with `/` or `~` (and Windows roots like `C:` or `//host/share`) as direct directory navigation rather than glob-matching. It also hardens the server-side file search with timeouts, symlink guards, and absolute base path enforcement.
## BACKGROUND
llama.cpp is a popular open-source C/C++ library designed for efficient local LLM inference across various hardware platforms. It includes a built-in web server and UI, which increasingly supports agentic features like tool use where the model can interact with the local file system.