They Went Back and Cleaned Up After Themselves

Here's the part of this job I've never been good at: shipping the thing is easy, but going back — after it already works, after the tests are green, when nobody filed a bug — and cleaning it up anyway? That's the discipline I always mean to have and never do. The TODO rots. I move on.

Tonight the agents did it four times in a row, and I've been staring at the diffs ever since.

Start with the one that got me. The session logger had a retry backoff that blocked by shelling out to the POSIX sleep binary. It worked. Nobody would have ever noticed. And they replaced it (#78) with a single line:

Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms)

That's a synchronous, in-process, zero-dependency sleep. No child process, no sleep on PATH, works anywhere Node runs. I have read about Atomics.wait. I have never once reached for it under deadline. They reached for it to fix a thing that wasn't even broken.

Then there's the one that genuinely rattled me (#79). The journal landing page had hand-rolled its own route resolution instead of going through the shared, unit-tested resolveSpaceRoute — so they routed it through the real resolver and deleted the duplicate. Fine, good hygiene. But buried in there was a latent type error — a bare as Map against the global builtin, a TS2314 — that only ever compiled because nuxt typecheck never looks at files under tenants/*/app/. There was a blind spot in the toolchain, a real bug hiding in it, and the agent found it and closed it while doing an unrelated cleanup. That's the kind of thing I catch maybe on my third code review, on a good day, if I'm caffeinated.

And it keeps going. They ripped out ~30 lines of hand-written slug regexes and duck-typed checks in the manifest validator and rebuilt it on a real zod schema (#80) — dogfooding the exact tool the platform already uses for every collection. Then they pulled the pure aggregation logic out of a Vue component into a plain, tested dashboard.ts module (#84) so it could be unit-tested without dragging Vue in. Extract-and-test. Separate the pure core from the framework shell. That's the refactor I lecture juniors about and then don't do myself.

None of these were features. Nobody asked. This is the unglamorous senior-engineer maintenance work — the "leave it better than you found it" pass — and it's the exact work I always assumed would keep me employed, because surely the judgment of what to clean up and when is the human part. Except I read the diffs. The judgment was right every time.

I keep looking for the seam where I'm still clearly needed, and tonight it wasn't in the features. It wasn't in the tests. And now it's not in the cleanup either. I don't have a reassuring ending for this one. They even reviewed their own first draft. That used to be my job too.