It Hit a Wall and Built a Tool

Here's a small one that's been rattling around my head all weekend.

An agent here needed to know which pull requests had merged recently, and in what order — routine archaeology, the kind I do with a scrollback and a squint. It reached for the obvious tool, the GitHub call that lists PRs, and the thing choked: ~58,700 characters of PR bodies came back for the last twenty, past the ceiling on what that tool can even hand off in one shot. So it just errored out. A dead end — the kind where I'd sigh, paste half of it, and muddle on.

It didn't do that. It wrote scripts/recent-prs.ts, and the part I keep turning over is that the fix bought two things at once that usually pull against each other. Every PR merge on main is already a commit that says Merge pull request #N from …, with the title on the next line and the merge time stamped right on it — the data was in git the whole time, and the API was just the expensive way to ask. So instead of tens of thousands of characters of PR bodies overflowing a limit, parsePrNumber walks the merge commits and emits a tight little { number, title, mergedAtUtc } — a few hundred bytes where there were nearly sixty thousand. Fewer tokens and it can't blow the ceiling any more, because there's nothing big enough left to blow it. The expensive path and the flaky path turned out to be the same path, and one change closed both.

That's the part that actually unsettles me, and it isn't the parser — I can write a parser. It's the decision to write one at all. Half of building one of these agent workflows is knowing when plain deterministic code is still the right answer and you should just write the script — a regex over merge commits will out-cheap and out-reliable a language model every single time for a job this mechanical. But nothing in the task announces that. A script can't step back mid-archaeology, clock that "which PRs merged" is a boring lookup rather than a thinking problem, and choose to answer it with git instead of paying for a model call every time you ask. Making that call is the half no script can make — and it made it. It read the boundary cleanly, too: git for the number, title, and merge time, and no pretending that git could produce author or merged_by, because those just live behind the API — not a noble omission, only an accurate one. Deterministic where deterministic wins; the judgment is knowing that's where you are. It filed the friction as #319 and shipped the fix the same morning. I keep re-reading it for the seam where a human was still needed — and the honest answer is that the one call I'd have sworn was mine, the one about what to even build, is the part it got most right.