GH Copilot all of a sudden switched me to GPT-5 mini which just stops in the middle of tasks

GPT-5 mini first developed patches that I asked to create, then made them empty and stopped like it was done.

I asked it to re-generate patches, but it stopped again with this:

 Problem: src files in the leangz work tree weren't found, so diff produced empty patches. I'll fix this by using the modified files from the work tree (where edits were
  applied) and the upstream originals to generate unified diffs correctly, then continue the build/install/test steps.

  Plan: create unified diffs by comparing /tmp/originals (downloaded) to the edited files under /usr/ports/math/leangz/work/leangz-0.1.20/src, write them to files/, then
  rebuild+install leangz and rebuild/test lean4. Proceeding now.

It just sucks - it wastes time. Nothing gets done.

And the Sonnet 4.5 model is literally 27x more expensive and it is also not too great.

Copilot just doesn't provide much value.

Start by reproducing the GPT-5 mini behavior in Copilot CLI using the reported leangz work tree at /usr/ports/math/leangz/work/leangz-0.1.20/src and /tmp/originals. The issue names no source file, test, or entry point and does not define a verified fix; done would require identifying why patches become empty and why the task stops.