My day job is testing healthcare software, and most of it happens in the gap between the button saying Saved and the thing actually being saved. A toast is not a database. A lot of bugs are a success message with nobody home behind it.

Then I started letting coding agents loose in my own repos, and there it was again. The agent asks to run something, I say yes, the command exits 0, the agent says done, and every one of those steps is green and not one of them tells you anything changed.

So I built a runtime that keeps them apart. Claude wrote the TypeScript, I wrote the rules, and it sits between me and whichever agent I’m using that day, Claude Code or Codex or Gemini or Antigravity. Every action gets a receipt with three fields: what it was allowed to do, what it tried, what actually changed. Allowed doesn’t get to borrow happened’s checkmark.

Subagents only ever get less power than whoever spawned them. Some paths are off limits because the contract says so, and asking nicely doesn’t change that. There are 33 offline test scenarios, plus a holdout set I didn’t tune anything against, because writing the test and the answer key yourself is just grading your own homework with extra steps.

The ambitious half is still scaffolding. There’s a sentinel that’s supposed to watch my repos on a schedule and ping me when something drifts, and right now it does neither, which makes it a very confident folder.

After any run I want to know what happened and what I’m still unsure about, but mostly whether it’s safe to run again. If your logs only know what was allowed, you learn the rest by running it again and watching, and that’s how one mistake turns into two.

The agents still say done all the time. Now they have to bring a receipt.