I like this idea. In fact, I've been doing something manually that I figured I should try to automate (or try to see if I could get added as a feature directly to gsd). My approach (which is not elegant for sure) is to copy my codebase right after I built my CONTEXT.MD file for a phase. I then have a separate AI (like in Codex or using GLM through Claude Code) to create the RESEARCH.MD and PLAN.MD file. I then have the original directory with Claude Code + Anthropic models also do an independent planning and research. I then have copied the GLM and/or Codex PLAN.md (sometimes I've done more than 1 alternative) into an "/externals" folder I have in my codebase. I tell my original Claude Code to compare the plan and research files and see if there is anything it can use that is done better in that version. It usually finds at least one thing.

I've been doing this for a few phases and it helps fill things out more. Your approach is way more streamlined. Part of what I was trying to achieve was independent producers of the files, sort of like having two people look at the same directions and see if they come to different conclusions.

Oh, and I also more recently even tried doing the split again, where I took the merged PLAN.md files (with input from all runs merged) and then let two independent directories execute on the same plan. Then I made a custom skill that produces a diff. I then feed the diff to my original instance of Claude Code so it can compare the changes it made to the one the other model made when executing the same PLAN.md. This also allowed for some better execution ideas to flow back into the original instance.