Blog

Our front page said 'AI features are disabled' for 40 seconds. The feature worked the whole time.

342 tiny plugin requests at 450ms round-trip starved the one promise the UI was waiting on — so a working composer wore a broken label. Fixing it meant redoing a change we had already tried and reverted.

The freeze that hid itself: 24 minutes of pinned window, and why we couldn't profile it

Our semantic index froze the editor for 1,457 seconds on a 1,341-file repo — and the freeze broke every tool we would normally use to find it. An await in a loop is not the same as yielding the thread.

We were charging every model up to 7× too much for memory it never used

Our context-window sizer used a formula from before grouped-query attention. It over-counted the KV cache by 3–7×, and over-counting is not the safe direction — it was handing a 4 GB laptop a quarter of the context it could hold.

Nothing to the cloud: how OIOXO's bridge gives a browser your machine's power

Archived experiment, not a shipped feature: the OIOXO bridge was a local relay on 127.0.0.1 meant to lend the web IDE the language servers on your machine, debugger, shell and git. It never shipped — see the docs for what OIOXO does today.

Note: this post describes an experiment that did not ship.

Night Shift: your computer builds while you sleep

Archived experiment from before 1.0: leaving an agent running overnight on your own hardware. Night Shift, git-worktree background agents and the `oioxo code` CLI did not ship — see the docs for what OIOXO does today.

Note: this post describes an experiment that did not ship.

The IDE that verifies its own work

Archived experiment from before 1.0. The verification loop described here — running your test runner, diverse candidate fixes, the `oioxo code` CLI — did not ship in this form. OIOXO does check its own work today; the docs describe what it actually does.

Note: this post describes an experiment that did not ship.