However, this model lacks image interpretation ability, so we will have Codex take on the auxiliary role of directly checking and interpreting asset images
Codex will store asset paths, pixel sizes, purposes, duplication status, and cautions in English in assets/ASSET_LABELS.md
The plan is for Qwen to refer to the documented asset descriptions rather than opening PNGs directly when selecting assets during Godot coding
Experimenting with attaching the Reddit search/investigation flow to MCP to make it easier to find community examples or Godot‑related problem‑solving material
May try it later if we need to adapt/tune the model based on collected Godot work patterns or failure cases
qwen2.5-coder showed different response behavior depending on vLLM vs. llama.cpp execution
In llama.cpp, Qwen Code’s tool calling was not properly linked, resulting in output like <function=glob>
vLLM has a higher likelihood of tool calling due to its execution method, but in our current experiments neither produced stable code fixes
The current model/serving combo lacked the ability to grasp the overall project context and fix it, but we will treat this as a fine‑tuning/environment‑alignment task rather than giving up
The goal is to make Qwen a coding agent that performs actual code modifications, not just a limited assistant tool
Will explore additional insights on AlphaEvolve as suggested by the coach and decide whether to incorporate it into the project’s experiment/optimization flow if needed