Skip to content

The cluster suite runs on a pull request that changes what it deploys, into main and into a chunk alike, and applies the CRDs and admission policy before it builds anything

Decision 889 · accepted · 2026-10-12 · issue #679 · supersedes 676

This replaces 676's sentence about the cluster suite and nothing else of it.

  • When. A pull request that changes what the cluster runs: charts/** (the CRDs and the admission policy template among it), docker/**, dev/kind/**, scripts/kind-up, scripts/remote-up, scripts/build-images, apps/troupe_operator/** (the operator and the suite) and cluster.yml itself. Into main through ci.yml's changes (output cluster), and into a development-* chunk through dev-check.yml, with the same list. The nightly and a release still run it whatever changed. Not on a push to main, whose pull request ran it, and not for any change under .github/workflows/, which runs every other job: twenty-five minutes for a change to the GUI's workflow would be a cost with nothing behind it.
  • Into a chunk too. A chunk's fixer may have no cluster, and the chunk's pull request into main comes once every fix is in, which is where #678 would have been found. The light check stays light everywhere else.
  • One job, two callers. The job moved into cluster.yml, a reusable workflow both call, so the chunk's and main's are the same steps. Each caller runs it after the checks that see the same change for less (lint, the test legs and chart in ci.yml; the compile in dev-check.yml), counting one skipped because its part did not change as passed, so a change they refuse spends no cluster.
  • The CRDs first. Right after scripts/kind-up, before the images are built (nine of the job's twenty-five minutes), the CRDs are applied and the admission policy and the kind values' TroupePolicy go through a server-side dry run, so the chart still owns them when it installs. A refused rule fails in about a minute.
  • What the suite proves of a worker's isolation (#670). A worker reaches no other namespace's pod and no pod but the ingress's reaches it; it cannot reach 169.254.169.254 even where another policy admits it, against a stand-in the job puts on the runner (an address on the node is the node's to Cilium, which no egress rule speaks of); the IPv6 and NAT64 forms are denied as the Cilium agent holds the policy, since kind has neither; a command a session runs on a pod runs in bubblewrap and sees neither the pod's BEAM nor its OpenBao token; and a profile's own OpenBao role, Troupe.KMS.Policy.worker/2 bound to its namespace, reaches its own teams' keys, no other team's or person's, deletes none, and is refused to another profile's ServiceAccount. Each refusal is set beside a pod or a role for which the same thing succeeds, so it fails when its rule goes.
  • Not chosen. The CRD step alone on a pull request (a few minutes): it would catch #678 and nothing the suite catches, such as an admission policy that refuses the dev profile. The worker and plane apps in the list: their own suites run on every change to them, and the cluster suite on each would put twenty-five minutes on most pull requests.
  • What it costs. The job takes what it took in the nightly, about twenty-five minutes, plus the new tests; on a pull request it starts after the cheaper checks, so a change to the chart or the operator waits for it.

Governs: .github/workflows/cluster.yml, .github/workflows/ci.yml, .github/workflows/dev-check.yml