This repository was archived by the owner on Oct 24, 2026. It is now read-only.
Repository navigation
feat: replay main work onto testing - #726
Merged
Merged
Conversation
* feat(factory): build the GNOME stack against CentOS Stream 10
bluefin-lts is built on CentOS Stream 10, which ships GNOME 47. The GNOME 50
stack has come from a third-party COPR project. That is not a durable answer
for a release line: its contents, its rebuild policy and its availability are
all outside this repository's control.
factory/ builds the stack here instead, from recipes we control, and publishes
it as a signed OCI repository image the Containerfile copies in. 77 recipes
are seeded from the working c10s-gnome-50 repository, so the EL10 adaptations
already exist and the version bump to GNOME 51 is the remaining work rather
than the port.
Design decisions:
- Packit is the SRPM gate, not the builder. Copr has no CentOS Stream 10
chroot -- EL10 builds there happen in epel-10 or alma-kitten+epel-10, a
different ABI from the c10s base this image ships on. .packit.yaml therefore
carries no `jobs:` section; binaries are built by factory-build.yml against
the digest-pinned c10s root.
- Build order is solved from real BuildRequires parsed by rpmspec in the build
root, never hand-assigned. A hand-assigned stage hides a new edge until a
build fails downstream, and a stage no longer needed still serializes work.
- No build downloads anything. config/upstream-sources.json pins the resolved
upstream URL and SHA-512 for 75 of 77 recipes; the source is verified before
the build and again inside it.
- Publication is incremental and gated on the consumer transaction: a failed
package keeps its previous build, and the tag does not move unless what the
image installs resolves against the candidate.
Three things found while building this, worth not rediscovering:
- `rpmbuild` with no spec argument globs the working directory, finds nothing
in the build output directory, and exits 0 having written nothing. A green
build producing no RPMs. Every invocation now names the spec explicitly,
with an artifact guard behind it.
- CRB ships disabled in c10s, and meson, ninja-build, wayland-devel and gcc++
live only there, so BuildRequires fail with "No matching package to install:
meson" -- which reads as a missing package rather than a disabled
repository. The requirement is declared in config/buildroot.yaml and
asserted by validate.py.
- The imported docbook-style-xsl spec carried an upstream typo, `%2F{%version}`
with malformed braces, producing a URL that 404s. The digest check surfaced
it rather than letting it build from something unexpected.
Not done, stated rather than implied:
- The recipes are the GNOME 50 set. The tooling is version-agnostic; moving to
51 is a version bump across the stack.
- amd64 only. The image publishes arm64 as well, so its manifest cannot
consume factory packages yet.
- Builds are not hermetic. They run in the pinned container, which gives the
right ABI but is not a build root in the sense mock means -- no build user,
no isolation, no reset between packages -- and mock-core-configs ships no
EL10 chroot. There is no build cache yet.
- Nothing has been published to GHCR, so image-versions.yaml carries an empty
gnome_packages digest. The image build detects that and falls back to the
base image's GNOME rather than failing.
Verified: `just factory-check` green; 53 unit tests; actionlint clean across
all workflows; `just check` passes; and a real RPM built end to end against the
pinned root (wayland-protocols-devel-1.47-2.el10.noarch.rpm).
Not verified: `just unit-tests` and `pre-commit run --all-files` were not run,
because bats and pre-commit are not installed in this environment. No image
build or VM test was run.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): fetch full history where the SRPM gate diffs against the PR base
The pull-request path of the Packit SRPM gate ran `git diff <base.sha> HEAD`
against the default depth-1 checkout, which contains only the single commit
that was checked out. The base commit is not in that object database, so the
step died on `fatal: bad object 2170146...` -- naming a SHA and saying nothing
about the shallow clone being the cause, which is what actually had to change.
Two changes:
- The checkout for that job requests `fetch-depth: 0`, so the base commit is
present.
- A `git cat-file -e` guard names the real cause if the base is missing for
some other reason, instead of emitting git's "bad object" again.
Reproduced locally against a depth-1 clone to confirm both the failure and the
fix, rather than assuming.
`pkg-cadence.yml` was checked for the same shape and is unaffected: it diffs the
working tree with no commit argument, which a shallow clone serves correctly.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): drop the PyYAML dependency so the tools run on a bare runner
`factory/tools/validate.py` and `buildroot.py` imported `yaml` to read the
build root config. GitHub Actions runners install no Python packages --
`actions/setup-python` provides a bare interpreter -- so the validation gate
died on `ModuleNotFoundError: No module named 'yaml'` before running a single
check.
This also contradicted a convention the repository already states: the
`read-packages` helper is stdlib-only "no new dependencies", and CI callers
rely on that. The dependency passed every local gate because the authoring
machine happened to have PyYAML installed, which is the failure mode worth
being explicit about: a tool that only works on the machine that wrote it.
Changes:
- `config/buildroot.yaml` becomes `config/buildroot.json`, read with the
standard library `json`. The explanatory comments move into `_comment` keys
and the architecture document, since JSON has no comment syntax.
- Both tools drop the `yaml` import.
- New `validate.py --check dependencies` rejects any non-stdlib import in
`factory/tools/`, using `sys.stdlib_module_names` so the list cannot drift.
This is the guard against the same mistake recurring: a ModuleNotFoundError
names a module and says nothing about the fact that CI never installed it.
Verified by reinjecting `import yaml` and confirming the check fails, then
removing it again.
- `rebuild_plan.py` reads the build root through `buildroot_config()` rather
than opening the path itself. It had already been updated to the new path,
but a literal path here fails silently: the build root stops contributing to
the input digest, and moving it stops selecting anything to rebuild.
Verified the digest moves when the pinned digest changes and returns when it
is restored.
All fifteen tools were then imported under `python3 -S -E`, which suppresses
site-packages, to simulate the runner directly rather than reasoning about it.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): resolve container mount paths in the Packit SRPM gate
`srpm_gate.py` passed the caller's `--output` directory straight into
`docker -v`. Both workflows invoke it with a relative path
(`--output work/srpm/<package>.src.rpm`), and docker reads a relative host
path as a *volume name* rather than a directory:
create work/srpm: "work/srpm" includes invalid characters for a local
volume name, only "[a-zA-Z0-9][a-zA-Z0-9_.-]" are allowed
That message names the path and says nothing about the cause, so it reads like
a bad package name rather than a relative path.
A missing host path is the more dangerous variant of the same mistake: docker
silently creates it as an empty named volume, and the build then runs against
nothing. Both are now avoided by resolving the paths and checking the output
directory exists before invoking docker.
The docker invocation is extracted into `command()` so the paths are testable
without a container engine, and `factory/tests/test_srpm_gate.py` pins the
behaviour: every `-v` host path must be absolute even when the caller passes a
relative one, and a missing output directory is refused rather than becoming a
volume. The test was confirmed to fail when the fix is reverted.
Checked `extract_buildrequires.py` and `render_source_locks.py` for the same
shape; both derive their mount paths from `factory_root()`, which resolves, so
neither could be given a relative path.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(factory): stop asserting on the checkout directory name
`test_repository_mount_is_the_repository_not_the_factory` asserted that the
repository mount ended with `bluefin-lts-1`. That is this working copy's
directory name; CI checks the repository out as `bluefin-lts`, so the test
failed on a name and passed only where it happened to be written.
The test now compares against the root derived from the test file's own
location, and additionally asserts the factory directory itself is *not* what
is mounted -- which is the thing the test was actually for, since
packit_source0.py resolves the spec against the repository root and mounting
only factory/ would leave .packit.yaml and build_scripts out of reach.
Verified by copying factory/ into a directory named `bluefin-lts` and running
the suite there.
Two further fixes while here, both about the test not lying about what it
covers:
- The recipe used to build a command was hardcoded to gnome-shell, coupling
the test to the recipe inventory. It now uses whichever recipe sorts first
and skips when the inventory is empty.
- `test_end_to_end_through_the_cli` let publish_gate's stderr escape into the
test output, printing "did not build: gtk4" mid-run. In a CI log that reads
as a real failure. It is now captured and asserted on instead.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): lock every source a recipe downloads, not just Source0
The Packit SRPM gate failed on glycin with:
error: Bad file: .../glycin-2.0.8-vendor.tar.xz: No such file or directory
Two separate problems were behind that, and only one of them was the error
being reported.
**The lock carried one URL per recipe.** glycin has a second Source --
libjxl 0.11.1, a private copy to avoid the EPEL libjxl-0.10 conflict. Three
more recipes do the same: malcontent carries libgsystemservice, docbook-style-xsl
carries its docs, source-highlight carries its signature. `render_source_locks`
took the first remote Source and ignored the rest, so those recipes passed
validation and then failed in `rpmbuild -bs` naming an archive nobody had
staged. The lock now records every source, and the pipeline fetches and
verifies all of them.
Two details that follow from it:
- A URL fragment (`...tar.gz#/libjxl-0.11.1`) selects a subdirectory, and rpm
names the file after the URL path with the fragment stripped. Staging it as
`libjxl-0.11.1.tar.gz` would have left rpmbuild looking for a file that does
not exist, so the filename is derived that way.
- `.sig` and `.asc` are legitimate sources -- the spec verifies against them --
so they are locked and verified like any other rather than rejected as "not
an archive".
Re-rendering now reuses digests for URLs that have not changed, so
`just factory-relock` fetches only what moved instead of every archive in the
factory. This run fetched four.
**Three recipes cannot be built at all.** glycin and gnome-user-share need a
`cargo vendor` tarball generated by hand; malcontent bundles gvdb and
tinycdb. Those sources are neither downloadable nor committed. They are now
marked `blocked` in the lock with the reason, excluded from the build list, and
refused by the SRPM gate -- so the failure names the reason instead of a
missing archive. `validate.py --check generated` reports them without failing,
because a recipe correctly identified as unbuildable is a correct state of the
configuration, and a permanently red gate is one people stop reading.
Committing a generated vendor tarball would mean committing a binary blob whose
contents depend on the crates.io index at generation time. That is not a
provenance statement, so this is recorded as a decision for a human rather than
worked around.
Verified: docbook-style-xsl stages and verifies both of its sources; the build
list is 74 with the three blocked recipes excluded and none leaked; 72 tests
pass, including new coverage for the multi-source lock shape, the fragment
handling, and duplicate filenames.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): stop pre-commit rewriting imported recipes; retry flaky sources
Three problems from one CI run, only the first of which was the one that
stopped it.
**pre-commit was rewriting the imported recipes.** `end-of-file-fixer` and
`trailing-whitespace` ran over `factory/packages/` and modified 68 files, 51
of them patches. That is not cosmetic. Those files are upstream content,
imported byte-for-byte and applied with `patch -p1`, and:
- a patch that removes a line carrying trailing whitespace stops matching once
that whitespace is stripped, so the patch no longer applies.
`factory/packages/fontconfig/fontconfig-fix-crash.patch` line 26 removes a
line ending in `-- ` and is exactly this case;
- AGENTS.md forbids modifying installed upstream/vendor content unless the
task is about that content;
- it churns the input digest, which decides whether a package is selected for
a rebuild, so re-importing the same spec would select the factory for no
reason.
Both hooks now exclude `^factory/packages/`. Verified by installing pre-commit
and running it: `trailing-whitespace` passes, and `git status factory/packages/`
is empty afterwards.
**fontconfig failed to download.** The gate reported
`no source verified ... HTTP Error 500`. The URL was fine -- gitlab.freedesktop.org
served the same bytes moments later, digest matching the lock exactly -- but
that host answers 500 on a fraction of `/-/archive/` requests, which makes a
healthy source a random red build.
`_fetch` now retries with exponential backoff, four attempts. Only transient
answers are retried: 5xx and 429, never 4xx, because a 404 means the URL is
wrong and hammering a host that is certain to say no helps nobody. Retries are
announced on stderr rather than hidden -- a source that needed three attempts
is a signal about its host, and hiding it makes the gate look more reliable
than it is.
Verified against a simulated host failing twice with 500 then succeeding
(3 attempts, 2 retry messages) and a 404 (1 attempt, raised immediately).
**actionlint's shellcheck rejected a workflow.** The comment `this stack's
build roots` sits inside a single-quoted `bash -exc '...'` argument, so the
apostrophe terminated the string and the rest of the script would have been
parsed as shell. All six workflows were swept for the same mistake; this was
the only one.
Also adds final newlines to 35 files the EOF hook corrected in this run --
these are the repository's own files, not imported recipes.
Verified: `pre-commit run --all-files` passes, `just factory-check` green with
72 unit tests, `actionlint` clean, and `bats tests/unit/` green at 176 tests
including the 7 new factory-repo assertions. Both were previously unavailable
in this environment and have now been run.
Assisted-by: Claude
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): repair CentOS build and incremental publication
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): resolve RPM capabilities and preserve input witnesses
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(factory): exercise Containerfile wiring and shared SRPM changes in CI
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): pin the SRPM tool and gate tags on signed provenance
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* ci(factory): keep smoke runs isolated by commit
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): name Containerfiles explicitly for Docker builds
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): load CentOS DNF variables and plan the selected build set
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): verify sources against the mounted factory root
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* feat(factory): target GNOME 51 and verify the complete CentOS stack
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* ci(factory): exercise full GNOME builds on pull requests
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): align GNOME 51 archive names and glycin dependency edges
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): solve the GNOME stack against exact upstream minima
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* ci(factory): retain per-wave evidence and admit deeper GNOME graphs
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(ci): restrict wave six to strictly earlier dependencies
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(factory): require GNOME 51 in the consumer transaction
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(factory): verify reproducible and checksum-locked Rust bundles
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): discover bootstrap specs and port GNOME 51 build options
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): port GNOME 51 PAM packaging and missing prerequisites
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): require the GNOME 51 Meson and Rust toolchains
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): map new capabilities and retry stale DNF metadata
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): resolve Pango prerequisites and verify installed GNOME 51
Disable Meson source fallback downloads and install the resolved consumer transaction before checking core RPM and shell versions.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): port desktop packaging and Rust target detection
Apply the Utah Rust 1.98 vendor-target fix, package new GOA files, correct versioned prerequisites, and unblock the source-locked parental controls build.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): stage the malcontent archive under its RPM filename
Add the C compiler required by the selected notification library.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): expose missing desktop library capabilities
Include the sound, color, mock hardware and Glycin GTK libraries in the calculated dependency closure.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): honor RPM source aliases and current Autoar options
Stage URL fragment aliases under the names rpmbuild actually expects and reject mismatched locks before CI. Remove obsolete Autoar GTK options.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* feat(factory): build Orca 51 with offline MathCAT support
Vendor the explicitly selected nested Cargo.lock, verify its bundle digest, build a native RPM, and test the installed Python binding.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): match the GNOME 51 JPEG XL binding minimum
Upgrade the source-locked private libjxl bundle to 0.11.2 as required by jpegxl-sys, retaining the existing system-library separation.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): capture Rust license summaries before packaging
Redirect the expanded cargo license commands as a group and require their output to be nonempty.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): satisfy GNOME 51 native build and test requirements
Add GTK libdrm headers and the Localsearch fixed-width type include. Detect active SELinux in umockdev tests and allow only open_tree beyond the pinned default seccomp profile, preserving all tests and capabilities.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): require the GLib API used by malcontent
Require GLib 2.84 headers so the pinned CentOS build upgrades the installed base. Keep the umockdev syscall profile with its recipe so incremental witnesses include profile changes.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): fetch the pinned AppStream source over HTTPS
Preserve the AppStream version and SHA-512 checksum while avoiding failures from the plain HTTP endpoint.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): test packaged glycin loaders in their sandbox
Match the manual Meson directory, stage a separate packaged test install under the read-only /usr bind, and permit only the required Bubblewrap namespace syscalls beyond the pinned default profile. Drop the removed GTK symbolic-SVG tool from its file list.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): resolve private libraries in staged Glycin tests
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): remove retired GTK demo file entries
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* test(factory): gate GNOME builds on dependency-tree preflight
Check exact CentOS and candidate capability witnesses before wave zero. Reject missing or incompatible dependencies and invalid plans, with native RPM version regression coverage and a metadata-only local command.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): reject truncated dependencies and probe Glycin tests early
Capture RPM queries with checked writes and verify declared dependency coverage. Qualify workspace loader test IDs, probe Bubblewrap before stack compilation, and retain full integration-test failure logs.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): allow Glycin namespace mounts with a confined AppArmor profile
Retain the pinned Moby default policy except mount and pivot-root permissions needed inside Bubblewrap's private user namespace. Load and select the recipe profile in preflight and both builders, using container-engine feature detection and unchanged capabilities.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): match Glycin fixtures to enabled loaders
Exclude only fixtures for disabled optional HEIF or JPEG XL loaders, keeping every compiled loader's sandbox tests enabled. Verify the RPM macro expansion for all four loader combinations.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): declare GNOME 51 portal and account prerequisites
Adapt the locked Utah portal and AccountsService recipes for CentOS, align declared GNOME source API floors, and retain a checksum-identical Malcontent staging fallback.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): verify the required SELinux runtime policy
Include the existing locked policy recipe in full-stack verification because the compatibility RPM requires a newer policy than the CentOS baseline.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): package SELinux module stores for CentOS
Set the module store relative to the rpmbuild policy root and replace the older exact-version CentOS extra RPMs whose modules are bundled in this policy.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): complete GNOME 51 runtime package metadata
Ship Mutter’s experimental schema, preserve the replaced GdkPixbuf modules capability required by CentOS GTK 3, and exclude the bundled Localsearch ZIP library from external requirements.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* docs(factory): clarify bootc policy store handling
Distinguish the pinned image-mode runtime store from the RPM module store used during package staging.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): package GNOME Shell calendar desktop entry
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* fix(factory): resolve Initial Setup runtime build prerequisites
Check all immutable Podman repository digests while retaining exact pin verification. Order Initial Setup after Shell 51 to avoid the CentOS Control Center AccountsService ABI conflict.
Assisted-by: GPT-6 via Codex
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---------
Co-authored-by: Robin Groves <robin@obsidianos.xyz>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* refactor!: move to thallium-suite/uranium and rebrand to Uranium The repository now lives at thallium-suite/uranium and is no longer a GitHub network fork, so the published image paths change with it. Shared CI and packaging continue to be consumed from projectbluefin. Registry and identity: - Publish ghcr.io/thallium-suite/uranium and ghcr.io/thallium-suite/uranium-nvidia. factory publishes ghcr.io/thallium-suite/uranium-packages; the buildroot mirror prefix follows suit. - Rework the cosign identity regexps to accept thallium-suite/uranium. The projectbluefin/bluefin-lts and projectbluefin/actions alternatives are kept for one migration window so already-signed testing artifacts still verify; drop them after the first post-migration promotion. Workflow wiring: - Rename the build workflows to "Build Uranium", "Build Uranium Nvidia" and "Build Uranium Nvidia (aarch64)", and update every workflow_run filter that referenced them by display name. The sibling lookup in build-nvidia-manifest.yml matches on filename and is unaffected. - Move CODEOWNERS to @thallium-suite/maintainers. - Drop docs_url to the Bluefin docs site; release notes are published as GitHub release bodies, so link there. An empty value is not safe: render_notes.py interpolates docs_url unconditionally. - Correct SECURITY.md, which documented an lts branch that does not exist. Branding: - os-release NAME/PRETTY_NAME to Uranium, CPE to universal-blue:uranium, DEFAULT_HOSTNAME to uranium, and rename the countme units and the dconf profile. Internal-only identifiers keep their upstream names and are deliberately unchanged: the gschema override, the /system_files/bluefin payload copied from the common image, and the *bluefin*.js match, all of which are owned by projectbluefin. Removals: - Count me reporting. There is no Uranium endpoint to report to, and the only available badge endpoint reported Bluefin's install count. - track-common.yml and lab-check.yml. Renovate owns dependency bumps, including image-versions.yaml digests via a customManagers entry declared locally, because the shared preset resolves its inherited config through {{parentOrg}} and this repository has no parent org. A PR opened with the default GITHUB_TOKEN would never satisfy the required validate check, so Renovate's own app token is the reason this is a swap rather than a conversion. - The stale .github/renovate.json5. Renovate never read it, but CI validated it, so fixes to it passed review and did nothing. - The offline Bluefin PDF from projectbluefin/documentation. The org Discord webhook was deleted at the repository level; it fired on check_run and workflow_run, so it notified on every run rather than only factory work. BREAKING CHANGE: images are published to a new registry path. Existing installs keep updating from their current reference and are not moved automatically; a manual `bootc switch` is required. There is no automatic migration service, so the README now states that plainly instead of claiming one exists. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chore: drop CODEOWNERS The file gated nothing. The only active ruleset has require_code_owner_review: false, and the other one is in evaluate mode, so neither could block a merge. Its only live effect was GitHub requesting review from a team that does not exist. Enabling it would have been worse than leaving it inert. The repository has a single write collaborator, so a code-owner requirement would demand approval from yourself, which GitHub does not allow, and the rulesets have no bypass actor. That combination makes the repository unmergeable. The file also asserted "Sensitive paths - require maintainer review" over .github/workflows/, Justfile and build_scripts/ on a repository that cosign-signs its releases, which is false assurance rather than a control. Release skill corrected along with it: it attributed the --admin bypass to CODEOWNERS. The actual cause is the merge queue blocking direct merges, and the required context is validate, which merge-queue-validate.yml posts on merge_group, not the Lint & syntax check. Branch protection that still applies: PR only, squash-only, non-fast-forward, merge queue, one required status check, 0 approvals. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chore: remove brand emoji and vestigial brand inputs bonedigger.yml passed brand_name and brand_emoji, but upstream marks both as "Retained for backward compatibility; ignored by the slim workflow" and supplies defaults, so neither reached any job. Dropped the whole with: block rather than just emptying it, since carrying inputs that nothing reads invites the next reader to assume they matter. Also dropped the decorative list emoji from the README community links. Left the check and cross marks in the SECURITY.md support table: those encode a boolean rather than decorate it, and plain prose is harder to scan. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * chore(sig): narrow cosign identity to this repository The identity regexps accepted the pre-migration identity alongside the new one, on the assumption that already-published testing artifacts would need to be promoted once. That assumption does not hold: this registry path has never existed, so testing content is built fresh and signed by this repo. A reusable workflow invoked through `uses:` also takes the caller's OIDC identity rather than its own, so the projectbluefin/actions alternative was only ever defensive. Both workflow inputs are now the same single-identity regexp, which rejects the old identities and any other repository. Verified against adversarial variants: uranium-evil, uraniumx, a case mismatch and a ../ traversal are all rejected. Kept the remaining bluefin-lts references, which are load-bearing: README names the previous image in the migration table, and 40-services.sh masks the pre-rebrand countme timer so upgraded machines do not carry an orphaned unit. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix(factory): repin packit container after quay garbage-collected the manifest The SRPM gate failed to pull its container: Unable to find image 'quay.io/packit/packit@sha256:b4667f39...' locally manifest unknown: manifest unknown quay.io deletes a manifest once the tag that pointed at it moves on, and packit republishes :latest often, so an untagged digest pin goes stale on its own. Nothing about the factory or the recipe changed; only the registry side of the reference expired. Repointed PACKIT_IMAGE at the current :latest digest. Verified the reference resolves, carries both amd64 and arm64 legs so either runner architecture works, and that the previously failing invocation now succeeds: python3 factory/tools/srpm_gate.py gnome-session --engine podman -> gnome-session-51.0-1.fc43.src.rpm, "verified 1 staged sources" The failure was image resolution, so it was package-independent and affected every recipe the gate iterates, not only gnome-session. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * feat: spiced up readme * style: strip trailing whitespace from README Introduced by da7c882 and caught by the trailing-whitespace pre-commit hook in CI. One line, in the Community section. Assisted-by: space-bunny-free Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Removed image from README and updated formatting.
… match the validated combination on the testing branch (#723)
…gning identity wait-nvidia-e2e polled 30 x 30s while the job timeout was also 15 minutes, so an exhausted budget killed the job before the script could exit 1. Every scheduled promotion ended as "exceeded the maximum execution time" with no cause, promote was skipped, and testing never reached main. The budget is now 120 x 60s with a 135-minute timeout, and a known-bad run fails fast instead of consuming the whole budget. sign-and-publish@v1 defaults certificate-identity-regexp to the projectbluefin org, so cosign signed the multi-arch manifest and then rejected its own signature. Pass the thallium-suite identity, matching execute-release.yml and promote-testing-to-main.yml. Also guards the SELinux module install and restores nvidia override coverage: ublue-os-nvidia-addons 0.15 dropped ublue-nvctk-cdi.service, and `systemctl enable` on the missing unit aborted the whole image build under `set -e`. Assisted-by: Space Bunny Free via OpenCode
image-info-set resolved FLAVOR with parameter expansion before 'set -u' took effect, so invoking it with neither a FLAVOR env var nor a positional argument expanded the positional parameter to an empty string instead of erroring. The script then wrote image-flavor: "" into image-info.json and exited 0, silently mislabelling the image. Guard the flavor before the rewrite and fail closed with a clear message when neither FLAVOR nor the positional argument is supplied. Also restores tests/unit/image_info_set_test.bats (added in #554, lost from main by a later promotion commit) and flips the final test from asserting the blanking behaviour to asserting a non-zero exit. Fixes #555 Signed-off-by: mrbobbytables <mrbobbytables@users.noreply.github.com> Co-authored-by: mrbobbytables <mrbobbytables@users.noreply.github.com> Co-authored-by: Jorge O. Castro <jorge.castro@gmail.com>
Replaying the main-targeted PRs onto testing surfaced four mismatches that only appear when the two histories meet: - Containerfile declared ARG GNOME_VERSION and the Justfile passed it, but nothing consumed it, and the recipe signature no longer bound $gnome_version at all. The GNOME version is owned by the factory and pinned by digest in image-versions.yaml (factory/config/gnome-stack.json says 51.0), so the vestigial arg goes rather than gaining a consumer. - packages_post_test.bats asserted the offline Bluefin documentation PDF. The rebrand drops it and there is no Uranium release carrying a replacement, so the three cases go with it. - nvidia_test.bats asserted an inline kargs tee; #651 moved the kargs into system_files_overrides/nvidia. Assert the overlay instead, and cover its contents directly. - image-info-set lacked the empty-flavor guard from #583. Replayed it; the restored image_info_set_test.bats covers it. Assisted-by: Space Bunny Free via OpenCode
pkg-cadence.yml still queried ghcr.io/projectbluefin/bluefin-lts:stable because it is a testing-side workflow the rebrand never touched. Point it at thallium-suite/uranium. docs/architecture.md still documented the countme timer, its helper and its opt-out file after the rebrand removed all three, and nothing documented the masks in 40-services.sh that stand in for them. Assisted-by: Space Bunny Free via OpenCode
* fix: restore terminal welcome banner
Only set the inherited UWELCOME_SHOWN deduplication marker for interactive shells so desktop login does not suppress Ptyxis greetings. Verify the inherited binaries remain available and cover the adjustment with Bats.
Assisted-by: GPT-5
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: TB516 <TB516@users.noreply.github.com>
* fix: silence deliberate SC2016 in 25-welcome.sh sed/grep literals
The single-quoted grep -F patterns and sed expression intentionally
contain $( ) and ${ } as literal text to match in uwelcome.sh.
Plain shellcheck in the validate workflow fails on info-level SC2016.
Scope a disable/enable around the three commands.
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
* fix: per-command SC2016 directives (0.10 lacks enable=, trailing comment broke scope)
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
---------
Signed-off-by: TB516 <TB516@users.noreply.github.com>
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
Co-authored-by: TB516 <TB516@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
#561 added x86_64 zfs variant coverage on main only. Without it the promotion would revert the file on the next testing -> main squash. Deliberately not ported: - build_scripts/overrides/aarch64/hwe/.gitkeep is an orphan; build.sh never dispatches it and build_dispatch_test.bats fails on it. - .github/workflows/merge-queue-validate.yml is redundant here; pr-testsuite.yml publishes the required `validate` context that the ruleset demands. Assisted-by: Space Bunny Free via OpenCode
#721 deleted .github/CODEOWNERS without putting anything in its place, though its message says ownership moved to @thallium-suite/maintainers. The old file pointed at @projectbluefin/maintainers, a team in another org, so it could not have resolved here either. The team now exists on thallium-suite. The sensitive-path list gains factory/ and .packit.yaml, which AGENTS.md already gates behind a pull request. The enforcement note records why nothing requires a code-owner review yet: the team has one member, GitHub will not accept a code-owner review from the PR author, so requiring one would make every PR unmergeable. Add MD-2016 or TheOddCell before switching the testing ruleset from "evaluate" to "active". Assisted-by: Space Bunny Free via OpenCode
end-of-file-fixer rewrites the file in CI, which fails the validate check. My earlier local pre-commit run reported success while the file was still missing its final newline; the hook output was swallowed. Assisted-by: Space Bunny Free via OpenCode
releases.pagure.org stopped serving libosinfo; the SRPM gate failed with "Source0: no source verified / HTTP Error 404". Upstream moved release assets to the GitLab project, which serves the same tarball as a release download asset, next to its .asc signature. The tarball at the new URL hashes to the existing lock (13647b6d...e6572), so this is a relocation of the identical artifact: the SHA-512 is unchanged, no patch regeneration is needed, and provenance is preserved. Verified by running the gate locally, which now builds libosinfo-1.12.0-1.fc43 with "verified 1 staged sources". Assisted-by: Space Bunny Free via OpenCode
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
PRs #709, #721, #722, #723 and #725 were opened against
maininstead oftesting.pr-validation.yml— thecheck-base-branchguard that rejectsmain-targeted PRs — exists only on
testing, so nothing stopped them. Theconsequences compounded:
testingnever received the Uranium rebrand, so it still built and publishedbluefin-lts/bluefin-lts-nvidia. Green CI ontestingproved the oldimage builds, not the one being released.
build-nvidia-manifest.ymlexpectsuranium-nvidiadigests, so thearchscheck failed andpost-testing-e2e-nvidia.ymlwas never dispatched.wait-nvidia-e2epolled 30x30s under a 15-minute job timeout, so it alwaysdied on job timeout rather than exiting 1. Every scheduled promotion has been
cancelled since 2026-10-05; the last real promotion PR was ci(promote): bluefin-lts testing → main 2026-09-05 #570 on 2026-09-05.
This replays the main-only work onto
testing, where the guard lives, so thesanctioned
testing -> mainpromotion can carry it again.What is in here
image-info-setempty-flavor guardwait-nvidia-e2e: 120x60s budget under a 135-minute timeout, plus fail-fastwhen a run already exists and did not pass
sign-and-publish@v1calls now passcertificate-identity-regexp; theaction's default only matches the projectbluefin org, so cosign signed the
manifest and then rejected its own signature
ublue-nvctk-cdi.serviceenable and guard the SELinuxmodule install (ublue-os-nvidia-addons 0.15 dropped both)
Conflicts resolved by hand
~30 files. Notable choices:
testing'shead_branch == 'testing'guard,--ref testing, and theactions: write+ dispatch fix from ci(nvidia): forward-port the NVIDIA e2e gate to testing and let it dispatch #685 over main's copywait-nvidia-e2e/check-release-windowjobs: they are whatthe 04:00 UTC schedule actually runs, since schedules fire from the default
branch
lab-check.yml,
.github/renovate.json5), and deleted the orphanedsystem_files/usr/libexec/bluefin-lts-countmehelper left behindFixes the replay forced
ARG GNOME_VERSIONwas dead and$gnome_versionwas no longer bound by therecipe signature. The factory owns the GNOME version now
(
factory/config/gnome-stack.jsonsays 51.0), so the vestigial arg is gone.packages_post_test.batscases for the Bluefin documentationPDF; the rebrand drops it and no Uranium release carries a replacement.
pkg-cadence.ymlstill queriedghcr.io/projectbluefin/bluefin-lts:stable.docs/architecture.mdstill documented the removed countme timer and itsopt-out file.
Deliberately not ported
build_scripts/overrides/aarch64/hwe/.gitkeep— an orphan;build.shneverdispatches it and
build_dispatch_test.batsfails on it..github/workflows/merge-queue-validate.yml— redundant;pr-testsuite.ymlpublishes the
validatecontext the ruleset requires.Checks
just check,just lint,just unit-tests(238 pass),just factory-check,actionlint,pre-commit run --all-files— all green locally.Not verified
No image build has run on this branch. The first
testingpush will rebuilduraniumanduranium-nvidiafor the first time on this content, and theNVIDIA aarch64 build has to succeed before
post-testing-e2e-nvidia.ymlcandispatch. Expect that first promotion to take a full build cycle.
Follow-ups
CODEOWNERSwas deleted by refactor: move to thallium-suite/uranium and rebrand to Uranium #721 without a replacement, though the commitmessage says it moved to
@thallium-suite/maintainers. That team404s, sorestoring it needs the team created first.
origin/scanner/fix-nvctk-cdi-retired(issue [scanner] build: nvidia image builds fail — ublue-nvctk-cdi.service retired upstream in ublue-os-nvidia-addons 0.15 #622) is superseded by the NVIDIAchange here.
mainstill holdsmerge-queue-validate.ymland the orphanhwedir; theywill be dropped by this promotion.