Feedback: Rethink the branch thing

[dlukt]

The way I work with this is, I ask GPT to write the commands by pointing at this repo and the describing the full product.

Last time it did a base skeleton and devops command,
a backend command, a frontend command and an admin ui command, and then some addon command that didn't fit the tech stack (main stack was Spring Boot Java and the addon was Go).

Then, as an experiment, I let qwen code cli do the job.
When it was done with branch 001, which qwen forgot to push to remote, I continued with 002 the backend, which it wrote and it commited and pushed.
The remote is here on github, I created a PR.
Because I have codex web PR scanning enabled, it found 3 issues.
But because I hadn't created an environment, it couldn't fix the issues (I have to add that eventually).
However now it's working on 003, and that will build on a 002 that has issues, which will lead to merge conflicts, because meanwhile codex will have fixed 002.

When I work with codex web, as an example, I can tell it to work on something, where it then creates a branch and PR, which then codex web review role scans and comments.

I think spec-kit should have a base repo mode and then a feature mode for smaller features.

Maybe my mistake was to have distinct branches for base+devops, backend, frontend, admin, edge,
but when I combined those into 1 in another experiment, it only did the backend, despite everything being defined in the commands.

So yeah there should be some kind of main repo mode and feature branch mode