Skip to content
Nightshift Dev

The loop, in numbers

These are the real numbers from OpenWhen, a product I built and run alone. Claude Code agents write the code. A scheduled one runs the product every day: it reads the numbers, opens PRs, grades them with a second agent and merges what's safe. Counted on 1 November 2026 with loop-stats.sh from the agent-ops-kit, so you can run the same count on your own repo.

PRs merged on OpenWhen since July 2026
157
opened by the scheduled loop, the rest in sessions I directed
114
revert on main since July 2026
1
of 538 commits on main written with Claude Code
97%

Month by month

MonthCommits on mainPRs merged
July 20267512
August 20261040
September 202612246
October 202623194
November 202665

There are more commits than PRs because I also push straight to main from my own Claude Code sessions. In August the loop did too, which is why that month shows no PRs. Since September it opens a PR for nearly everything it changes.

The one revert

On 22 August 2026 a change moved the PRO badges in the letter editor. Two days of the product's own analytics showed people opening the still-badged features far less often, so the change was reverted, with the numbers in the commit message. A revert count undercounts mistakes, though: most bad changes since July were fixed forward with a new commit, and this count doesn't show those.

Grading

The graded merge policy has run since 31 October 2026. Before that, a list of file paths decided what could merge on its own, and I reviewed the rest. Grade X PRs, the ones a revert can't fully undo, get a founder label and wait for me. So far: 0. Grade counts per week go here once there is a full week of them.

The policy changed after a real miss. On 31 October a PR passed lint, types, build and tests while an envelope's artwork sat hidden behind its seal. Now a UI change can't grade A without a phone-width screenshot and a clean console. The PR grading guide has the details.

What each number means

What isn't here yet: cost per night. Each run has a budget cap, but I haven't been logging what it actually spends. When I do, it goes on this page.

This setup has only run on products with one human. If you run the kit on a team, tell me. I'd like to put a team's numbers here next to mine.

Get the kit on GitHubLearn to run it