Repository navigation
Description
What
ci.yml's two runner resolvers are asymmetric. Pick the Linux runner label
has three branches:
- fleet-eligible -> the CodeBuild label
IN_REPObut not fleet-admitted ->kirocrew-linux-4-core- anywhere else (a fork running its own CI) ->
ubuntu-latest
Pick the Windows runner label has only the first and the last:
if [ "$ELIGIBLE" = "true" ]; then
echo "windows_runner=$LABEL instance-size:large" >> "$GITHUB_OUTPUT"
else
echo "windows_runner=windows-latest" >> "$GITHUB_OUTPUT"
fiSo every run that executes in this repository but misses the fleet -- a fork
pull request against this repository, an actor outside CODEBUILD_ACTOR_IDS, or
a pull_request action other than opened/synchronize -- takes the
enterprise-shared standard Windows pool, even though it can reach
kirocrew-runners-windows.
This was a decision, not an oversight. docs/ci/ci-and-reviews.md records
it: the non-fleet path is the high-volume one rather than the exception, so
pointing it at a capped, billed group moves the largest population of Windows
runs there. That objection still describes the trade accurately. The question
this issue raises is whether to make the trade anyway, as the Linux resolver
already does and as #26809 does for fifteen inline expressions.
Add the middle branch, mirroring the Linux step:
elif [ "$IN_REPO" = "true" ]; then
echo "windows_runner=kirocrew-windows-4-core" >> "$GITHUB_OUTPUT"By LABEL rather than group:, for the reason the Linux step already records: an
organization runner group is unreachable from a fork, and branch 3 has to stay on
the hosted image because a label no runner carries leaves every consumer queued
until GitHub cancels it, which is worse than the shared pool. build-windows.yml
already uses this label under the same repository conditional.
Why
Same reason as the lanes before it: nothing gets faster, because
kirocrew-windows-4-core is the same 4 vCPU / 16 GB as windows-latest. Only
queue position moves, and the capacity being protected is the enterprise-shared
standard pool rather than the invoice.
This is the last place a Windows job reaches a standard image by omission. The
remaining two sites are deliberate: #26178 owns the mixed-matrix legs, and
security-scope-review.yml's adjudicate leg is held for the reason in
What is deliberately still on ubuntu-latest in docs/ci/ci-and-reviews.md.
What has to be decided with it
kirocrew-windows-4-core is capped at maximum_runners: 32 and raising it goes
through GitHub Support. The measurement, from the jobs API over the last 60
ci.yml runs:
| Reading | Value |
|---|---|
| Windows jobs per fleet-admitted run | 11 (8 backend-test-windows shards, backend-test-windows-fail-closed, electron-test-windows, e2e-boot-matrix's Windows leg) |
| Peak concurrent on the hosted path | 7 -- max-parallel holds the shards to 4 off CodeBuild |
Windows jobs that took windows-latest across the sample |
89 |
| Group cap | 32 |
So roughly four or five concurrent fallback-taking runs saturate the whole
group, which already serves pod-boot-windows, build-windows-installer,
verify-windows and build-windows.yml's two jobs. The fallback is a recurring
path rather than an occasional one, so this is PR-time occupancy on that cap.
The decision this needs: whether 7 concurrent jobs per run belongs on a
32-slot cap today, or whether the cap is raised through Support first.
Not in scope
- The two mixed-matrix Windows legs in
ci.yml.#26178owns thoseruns-on
lines. main-ratchet-audit.yml's hosted fallback, which
test_the_ratchet_audit_route_stays_off_the_larger_runner_grouppins
deliberately until the merge-group leg is settled.- The macOS definition. GitHub shares the maximum concurrent macOS jobs across
standard hosted runners and larger runners, so moving a macOS job buys no
additional slot. - The
max-parallelshard cap. It stays at 4 off CodeBuild; this change does not
lift it.
Activity
tim-finnigan commentedon Nov 1, 2026
Taking the trade. Every Windows run that executes in this repository and misses the CodeBuild fleet should queue in our own group rather than the pool the whole company shares, on the same rule I've applied across this migration: nothing stays on a default runner.
The doc's reasoning is still correct and that is what I'm accepting, not disputing. The non-fleet path is the high-volume one, so this does move the largest population of Windows runs onto a capped, billed group.
kirocrew-windows-4-coreis the same 4 vCPU and 16 GB aswindows-latest, so no job gets faster or slower; what changes is whose queue we sit in, and that we pay for the slot.Raising the cap goes through GitHub Support. I'll open that separately rather than hold the change for it, and I'm accepting the smaller private queue until it is raised.
The
docs/ci/ci-and-reviews.mdsection that still calls the hosted fallback deliberate gets reconciled in #26425 with the rest of the runner inventory, not in #26828. A dozen in-flight runner PRs all append to those same paragraphs, and #26303 already removed the running totals from that file for the same reason.Recorded by my agent (Kiro) on my instruction.