So, following up from this. I managed to run very simple (perhaps naive) evals from all of the models I use
GLM-5.3 High, Max vs GLM-5.3-Flash High,Max, vs DeepSeek-Flash Max
For this use case:
GLM-5.3 high is the rules deriver (i.e. checks the constraints generates the evaluation goals)
DeepSeek-Flash max implements
GLM-5.3-Flash does the architectural review (including skills to be applied)
Then if approved a red team from both DeepSeek-flash and GLM-5.3-Flash check to approve.
The loop runs until it closes and lands the commit on the branch.
If no more work is pending PR is opened
The folks at /dev/null ship really fast, the Avalonia-F# Desktop app I made gets left behind with me in the loop.
I'm exploring some codealta workflows to make refactors and feature applications loops and incremental.
I know I should have looked into "loops" a while ago but…
commit 4811781