In internal/claudecode/store.go, detectHardware() currently hardcodes 8 GB RAM and 0 GB VRAM (bumping to 4 if nvidia-smi exists). The model recommendation then uses these defaults, so most machines get the same phi3.5-mini/qwen result even if they have 32 GB RAM and an RTX 4090.
It would be helpful to query the real RAM/VRAM on Windows/Mac/Linux (e.g., parsing systeminfo, vmstat, nvidia-smi, rocm-smi, or /sys/class/drm/card?/modes). A contribution could wrap that detection in a helper, return the actual values from detectHardware, and that would enable smarter recommendModel choices.
In
internal/claudecode/store.go,detectHardware()currently hardcodes8 GBRAM and0 GBVRAM (bumping to 4 ifnvidia-smiexists). The model recommendation then uses these defaults, so most machines get the samephi3.5-mini/qwenresult even if they have 32 GB RAM and an RTX 4090.It would be helpful to query the real RAM/VRAM on Windows/Mac/Linux (e.g., parsing
systeminfo,vmstat,nvidia-smi,rocm-smi, or/sys/class/drm/card?/modes). A contribution could wrap that detection in a helper, return the actual values fromdetectHardware, and that would enable smarterrecommendModelchoices.