gs-superadmin lets admins work in a live SaaS tenant through the vendor's CLI, with an approval prompt on every change. These pages show the system that builds and maintains it: how each vendor release is audited, how AI sessions build and check each other's changes, and where a person decides.
Most of this work lives in private repositories, so these pages show it instead: the problem, what I built, the principles behind it, and numbers that each come with the command that reproduces them. No code is reproduced here, and no employer, customer or tenant is named.
Tip
Start with the system page. It shows how four repos work together, and it's the one to read if you read one.
Keeping an AI admin co-pilot correct as its vendor's CLI changes underneath it. A public plugin lets an agent administer a live tenant behind approval prompts. A release watch audits every new vendor version, a builder/tester loop builds each change, and a small CLI keeps the loop's books. The page traces one vendor release through all four repos in 12 steps, with a person deciding at three of them.
What I've built since the 2026-10-06 publish and the problem each change solved, the lessons learned and the safeguards they added, and the measures that moved. Every number is re-measured with the same commands at each refresh.
| Page | In one line | Its steps in the system |
|---|---|---|
| gs-fortress → | Audits every vendor CLI release against a 13-item impact checklist and keeps the adopt/defer ledger | 2–4 and 12, plus the ledger behind the human decision at 5 |
| dev-utils → | Keeps the books for AI builder/tester sessions so the models spend their effort on judgment | 8 and 11 |
| superfriends → | The process skills: how AI sessions plan a build, hand it off, verify each other's work and release it | 7 and 9, plus the kickoff format at 4 and 6 |
The fourth repo, the plugin itself, is public: gs-admin-cli-docs. That's the place to read the code.
These repos, and these pages, were built with the help of Claude Code, Anthropic's AI coding tool. I set the direction, made the design and release decisions and reviewed the work; Claude Code sessions wrote much of the code and prose, inside the builder/tester process these pages describe.
Numbers last updated 2026-10-06; the learning page was measured on 2026-10-18. Each page says when its numbers were measured; counts such as commits and pull requests keep growing after that.
Every figure sits beside the command that produced it, pinned to the exact commit it
measured, in each page's proof/ folder. FACTS.md lists the key claims in one
place, with their evidence. For the public repo you can rerun the commands yourself; for
the private ones, I'm happy to rerun any of them live in a walkthrough.
© 2026 Bradley Bazhaw. All rights reserved; see LICENSE. You are welcome to read and link to these pages.
Gainsight is a trademark of Gainsight, Inc. This is independent work, not affiliated with or endorsed by Gainsight, Inc.