CI

Workflows

Workflow	Trigger	Does
build-image.yml	push to main or stable, dispatch	Builds, signs, and pushes the image.
execute-release.yml	push to stable	Promotes the candidate digest. Does not rebuild.
promote-main-to-stable.yml	daily schedule, dispatch	Opens the squash promotion PR and runs the release gate on it.
sync-stable-to-main.yml	push to stable	Merges stable hotfixes back into main.
pr-validation.yml	pull request	The validate check: shellcheck and hadolint.
validate-brewfiles.yml	pull request	Brewfiles, without evaluating them.
validate-flatpaks.yml	pull request	Flatpak preinstall files against Flathub.
validate-justfiles.yml	pull request	just check.
validate-renovate.yml	pull request	Renovate config.
unit-tests.yml	push, pull request	The bats suite.
renovate.yml	schedule, config change	Runs Renovate.
clean.yml	schedule	Deletes images older than 90 days.

Most are thin callers of reusable workflows in projectbluefin/actions.

The release model

main publishes :stable-testing. stable never rebuilds: promotion is a squash PR from main to stable, and execute-release.yml copies the digest :testing resolves to. The README owns the release table and the promotion gate's current limits.

The factory reusable puts its release gate and its auto-merge enrollment behind one input, enqueue_promotion. A personal repository cannot enroll — there is no merge queue, and gh pr merge --auto refuses without a merge method — so enrollment is off, and promote-main-to-stable.yml runs the gate itself in its own gate job to keep the pre-merge check. That gate resolves :testing when it runs, so it attests the current candidate. The binding comes from execute-release.yml passing source_branch: main, which makes the reusable refuse to promote at all once main has moved past the promotion commit; a manual dispatch is exempt, because that path is deliberate recovery.

The same reusable builds the squash branch, and it stages deletions with git diff --diff-filter=D. Git reports a moved file as a rename, so a path main moved survives on its old path and the branch's tree stops matching main's — the state the guard above refuses, but only after the PR is merged. repair-promotion-branch in promote-main-to-stable.yml rebuilds the branch from main's tree when it has drifted, and the validate check fails a promotion PR whose tree does not match main. The sweep belongs to projectbluefin/actions; the one-line fix there is --no-renames.

Signing

Keyless OIDC via Cosign. There are no keys to generate or store; the workflow needs id-token: write and packages: write. Unsigned images fail the promotion gate. The README has the command to verify an image.

The promotion gate is the only enforcement point. Nothing checks the signature on an installed system, so 00-image-info.sh writes an unverified update transport (ostree-unverified-image:docker://…) and the README says so. ostree-image-signed: would send the client to /etc/containers/policy.json, which Common supplies with no scope for this namespace — it would verify against the "" catch-all, insecureAcceptAnything, and report success having checked nothing. Adding a scope does not rescue it while signing stays keyless: containers/image matches a Fulcio certificate on subjectEmail alone (mandatory, exact, with a standing FIXME for URI SANs in signature/fulcio_cert.go), and a GitHub Actions certificate names its workflow in a URI SAN with no email to match. Device-side verification is a key-based signing change first, a policy change second. tests/contract/image-signing_test.bats fails if either side moves alone.

The identity regexp the release workflows pass to the reusables is scoped with github.repository, not github.repository_owner. Matching the owner and then any repository accepts a signature minted by any repository in the org, and github.repository keeps the scope correct in a fork without hardcoding it.

Renovate

Self-hosted through projectbluefin/actions, running every six hours. It pins GitHub Actions to SHAs and updates image digests. The policy lives in .github/renovate.json: updates below a major automerge once checks pass; majors wait for a pull request — except quay.io/fedora-ostree-desktops/*, where cross-major bumps do not open a PR at all (the rule is enabled: false). A Fedora major rebase for that namespace is a deliberate manual step: bump the tag in Containerfile by hand when the next major is released and validated.

Renovate needs the RENOVATE_TOKEN secret and auto-merge enabled. Both are onboarding steps. The secret is optional: with it unset the workflow logs a skip and the run stays green, which is how upstream runs, where an org-wide app does the work instead. The check sits in its own token job because a job that calls a reusable workflow cannot hold steps, and jobs.<job_id>.if cannot read the secrets context — secrets in a job-level if is a parse error, not a skip.

Automerge deliberately covers GitHub Actions SHA bumps, which reverses a guard upstream kept. Those SHAs run in jobs holding packages: write, id-token: write, and secrets: inherit, and PR builds are disabled, so a bump merges with only shellcheck, hadolint, and the test suite having run. Putting the guard back is one rule — matchManagers: ["github-actions"] with automerge: false.

Making a change

1. Open a pull request against main.
2. Wait for validate and the image build.
3. Merge. main publishes :stable-testing.
4. Review and merge the promotion PR to publish :stable.

Apt mirror resilience

The unit-tests.yml Install bats step runs apt-get against the runner image's stock Ubuntu mirrors. A slow mirror can leave apt waiting past the 10-minute job timeout, cancelling a run whose diff cannot have caused the stall (issue #497, observed on PR #492 head 5769c24, job 112944421448). The step fronts both apt-get update and apt-get install with timeout 180, passes -o Acquire::http::Timeout=30 -o Acquire::https::Timeout=30 -o Acquire::Retries=2 to each so a stalled socket fails fast from either half, and --no-install-recommends to skip the bats-doc package the workflow never reads. The wall-clock cap is what catches a slow-but-alive mirror; the per-call timeouts only fire on inactivity.