Coding agent progress: I thought I was there with the OpenAI integration.
I had model reasoning integrated, and this time I was correctly calling the system itself, so it’s no longer burning quota and falling back on every call.
I really did use OAuth correctly. But then I ran the end-to-end eval, only to find that these incredibly smart Codex agents were failing on my projects.
Why?
Because the whole architecture depends on shadow repos and shadow artifacts separated by project ID inside the system. That keeps memory and work deltas coherent, isolated, and recoverable.
The shadow system is nice because it makes the workflow lossless and non-destructive. I’ve basically built an elaborate containment system to make sure an agent can’t actually hurt your project, even if it melts down completely, rebases like crazy, and fries the repo.
This happens sometimes, and at this point I think it’s a reality of the models we use more than a pure skill issue, because even people who know better keep reporting the same class of failures.
So it has to be addressed systematically.
To their credit, Cursor does attempt something in this direction, but what I’m building is more comprehensive.
The collision with Node.js
Anyway, I thought the shadow artifacts were safely tucked away. Turns out Node.js tried to use them.
That sent the Codex models into a minor existential crisis. They were trying to reason about shadow assets they could not actually see, when they should have been running tests.
Meanwhile, my business manager Samuel is asking for a dashboard where he can track all of our happy little agents across multiple running projects.
Honestly, that makes sense.
But the collision of concerns here prompted a small architecture change.
I’m still tracking everything on project ID, but I moved the shadow data, Git repo, missions database, and context stores into a centralized directory outside the project folder.
And now we test again.