Late-night vibe coding has its own context window

The routine is comfortable now. Around half nine, after the kids are in bed and the house is quiet, I open the editor and start what I half-jokingly call my vibe session. Opus is the opener. It is the one I think with: dense, expensive, unhurried. I throw a half-formed idea at it and let it talk back.

About an hour in, the session limit lands. The chat goes quiet, the rate-limit banner appears, and I switch to Codex on the high-fast tier. That gets me another two hours, which is usually enough to take whatever Opus and I sketched and turn it into something that runs.

By midnight both windows are out of road, and so am I. The interesting bit is how cleanly my own stamina maps to the same shape. The first hour is generative - I am still curious. The next two are executional - I tidy, refactor, rename. After that, I start compressing my own thoughts: skipping context, taking shortcuts I will regret, missing things I would normally catch on a first read.

That is the cue. When my context window is full, no amount of Codex tokens is going to save the next change. The model’s rate limit and my brain’s seem to be on roughly the same schedule. I have stopped fighting it.

The work is better when I close the laptop on time.