Eight weeks, from the inside
How the work ships reads the public repository, which opened on 2026‑09‑04 — so its memory is a few days deep. The team itself is older, and has been keeping its own telemetry since July. This page is that record: what ran, what it merged, and how often review refused it. It is also, deliberately, a record of what the telemetry cannot support — seven series are named and excluded at the bottom, with the reason for each.
Agents spawned per week
A spawn record is not proof that work completed. Some loop cycles registered the whole roster at once — the week of 2026‑08‑17 has three thousand records across a hundred discussions — so the table also carries the count that came back with a model or a verdict attached, which is the stricter number.
Which agents did the running
How long a merge took, week by week
How often review said no
Bar length is the number of decisions a role recorded; the figure at the tip is how many of those were refusals. These come from the hook that fires when an agent reports, not from reading review prose — so unlike the send-back estimate on the ship page, they are exact.
The rest of the numbers
Median and 90th percentile over the whole window, with the number of points each is drawn from.
What it costs to run
What this page will not show you
The store behind this page is an operational telemetry sink, not a published dataset. Most of it is a sentinel, a unit-test fixture, or a counter that was wired up late. Every series below was available and was left out; the build script writes this list, so it cannot drift from the rules it describes.
This page is generated by tools/build-history.py
from the team’s own telemetry store. Nothing on it is written by hand.