Setting Up the Fork

Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.
It would be silly to start working without knowing what rules are already in place. Protections, builds, hooks. All of that lives in .github, which I have not opened yet.
But there is a bigger question sitting on top of it. Do I hard fork right now and never look back, or do I keep pulling from Microsoft?
My gut says hard fork. The history told me the team is less interested in the editor and more interested in the agent window. If I want something from a future release I can always pull that version down later and ask what is worth taking. For now I am calling the editor baked.
Then Git makes the decision urgent for me. I try to commit through the source control panel and it refuses: create a branch first. I should have expected that.
So I ask Claude, in plan mode, no work yet. What Git settings do I need to know about? I just tried to commit to main and it stopped me. And now I am wondering whether to forge my own path, because this is a fork and I do not plan on going back. I never want the agent code merged in. To me that means gutting the project. But I also do not want to code myself into a corner. I am trying to follow the Cursor model, where they keep the editor current while doing their own thing. Help me figure out what to do, and audit .github while you are at it.
Long prompt. That is fine. It is Opus, it will parse it, and it answers several questions at once.
While it runs I browse .github myself. Agents, commands, endgame, instructions, issue templates, prompts. Most of it is Copilot specific or VS Code team specific. The workflows are enterprise coding team workflows. I am not an enterprise coding team. I am one guy with a fork.
What comes back:
- The branch protection is not Git and not GitHub. It is a VS Code setting.
- No Git hooks.
- Origin is me. There is no upstream. I had no path to Microsoft's updates at all.
- Telemetry, Copilot payloads, code owners, notify settings. All deletion targets.
- The recommendation is the Cursor model: never diverge on main, keep it as a pristine mirror, live on your own branch.
That is the right call, so we take it. Create rob-main to branch from, add upstream as fetch only, and start deleting the Copilot infrastructure. Log the decision.
110 files gone from .github. Main untouched. rob-main is the trunk. The fork strategy goes into docs, and the branching strategy goes into CLAUDE.md so it survives a context clear.
One mistake, caught a little late. I meant to audit .github for skills and prompts worth keeping before deleting it. But it is all still in main, so nothing is lost.
That audit is next, and it matters more than it sounds like it does.
Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.