What changed

• Added a concrete roadmap for automatic llama.cpp runtime installation, Hugging Face GGUF search, model downloads, managed server startup, first-run setup, and recovery.
• Adopted accessibility-first defaults:
  • --cache-type-k q8_0
  • --cache-type-v q4_0
  • --parallel 1
  • --n-gpu-layers all
• Prefer official Vulkan builds on Windows and Linux where supported, with CPU fallback.
• Use official macOS builds with Metal rather than Vulkan.
• Added fallback order for insufficient VRAM and unsupported quantized KV cache.
• Defined acceptance criteria and a six-PR implementation sequence.

Why

The repository currently makes the proxy easier to configure, but users still need to install and operate a model server and find a compatible GGUF. This roadmap moves those tasks into the application and targets a terminal-free first run.

External interfaces reviewed

• Current llama.cpp releases provide official Vulkan assets for Windows and Linux and native macOS assets.
• llama-server supports quantized K/V cache types, --parallel, and GPU-layer offload.
• Hugging Face supports model search and filtering through huggingface_hub.HfApi and the Hub API.

Issues are disabled in this repository, so this plan is proposed as a versioned documentation PR instead.