We profiled 199 CI jobs across 20 workflows in a 35-minute window on 2026-11-09. They used 170.9 runner-minutes. Of those, 44.5 minutes (26%) were actions/checkout, and 5.3 more were setup and teardown.
The shape matters more than the total. The big jobs (the site build and its tests) spend 14 to 16% of their time on checkout, because they do real work. The small guard jobs (conflict markers, einvoice-preflight, research) spend 73 to 75% of theirs cloning a large monorepo to run a check that reads a few files. Their median run is about 30 seconds, and about 22 of those are the clone.
Earlier passes took the algorithmic costs: a zlib level, a duplicated gallery search, a glob that scanned every file. Each was a setting or a library choice, never a slow language. What remains is a clone that is bigger than the job needs. A sparse checkout of the paths a guard reads would cut most of it.
The catch is not technical. Those workflow files are on the repository's control-plane list, so changing them waits on the owner. That is a reasonable trade for protection against rogue agents, and the profile now puts a number on its price: roughly 25 runner-minutes in every 35 minutes of pool time.
Reading and the ranked list of what was not changed: docs/history/ci-perf-2026-11-09-pool-reading-checkout-bound.md.