llama.cpp b10331 fixes a working directory bug in the server’s get_info endpoint that returned the wrong path when tools ran inside an isolate.
What changed:
get_infonow asks the isolate for its working directory when a tools runtime is configured, instead of falling back to the server process cwd- Previously it reported a host path no tool would ever run in — isolates start in their own directory
- The process cwd is still returned when tools run on the host directly (no isolate)
- Co-authored by Xuan-Son Nguyen
This matters for anyone running the llama.cpp server with tool-calling agents. If your function call or MCP tool resolved file paths from get_info, it was reading a directory the tool never executed in. The reported cwd now matches where tools actually run.
Prebuilt binaries ship across the usual matrix:
| Platform | Builds |
|---|---|
| macOS | arm64, x64, iOS XCFramework (KleidiAI disabled) |
| Linux | x64/arm64/s390x CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL |
| Windows | x64 CPU, arm64 CPU, arm64 Vulkan |
| Android | arm64 CPU |
NOTE
The KleidiAI-enabled macOS Apple Silicon build is disabled in this release. Fall back to the standard arm64 build until it’s re-enabled.
Available now in release b10331. The fix is server-side — no model file changes required.