The core idea: You want to do software development in an environment that matches your CI.
And in the age of agentic AI, you want to align your CI and AI.
The idea behind this repository is simple: it's a playground that can be force-pushed to arbitrarily and acts as the "dev space" for one particular human (in this case, cgwalters).
There's a few bits of code you might find useful in this repository, but the key thing is just the ideas. See below.
Testing changes to CI by force pushing to a PR spams repository notifications.
Instead of that, force reset the main of your devspace to the target repository
you're working on, and then submit PRs against that instead.
Once that works, then do the PR against the actual repository.
GitHub supports workflow_dispatch to start a runner outside of an event (like a PR).
A generalization of this is interactive login to work on code. GitHub codespaces are a first class thing on GitHub, but have their own costs and semantics.
In this repository, I have a workflow where the repository has an oauth token that allows access to my personal Tailnet; the runner is accessible via ssh from my personal machines.
So then it's easy to do any hybrid you want - test some GHA YAML and then listen on SSH, or just have it be "listen on SSH".
The next obvious generalization of this is: Give agentic AI to do all of the above; the agent runs e.g. on your local machine, but can do the heavy lifting on an ephemeral remote environment. It works really well.
A stricter evolution of the above is running the agent in the remote environment. For more on this, see e.g. agentic-job.
There's a small CLI in repository which adds a CLI usable by humans or agents that abstracts over how this works, but is not required at all.