Repository navigation
Conversation
castrojo
left a comment
There was a problem hiding this comment.
Reviewed: Introduces config/multimedia-closure.toml, tools/multimedia_closure.py, and packages/ffmpegthumbnailer to inventory and verify Bluefin's multimedia requirements, resolving part of #24. Clean trial merge against main and passes all tests and factory validation. Awaiting second independent review to satisfy merge policy.
6831290 to
a890272
Compare
There was a problem hiding this comment.
intent-alignment
- The "reporting only" publish step actually gates the publish (medium). The PR body says "The publish-time report step therefore reports and does not gate", and the step's own comment repeats it (
.github/workflows/rebuild-rpms.yml:738-743) — but the step has nocontinue-on-error: trueand no|| true, so any nonzero exit fromtools/multimedia_closure.pyfails the publish job before the consumer-transaction validation and image push. The--checkhalf is PR-gated in validate.yml, but the--repodatahalf only ever runs here:published_nevraraisesClosureErrorwhen no*primary.xml.gzmatches (tools/multimedia_closure.py:published_nevra), a failure mode no PR check can catch. The repo's precedent for advisory steps is explicitcontinue-on-error: true(e.g. the preflight job in.github/workflows/build-stage.yml).
correctness
published_nevrais coupled to createrepo_c's gzip default (low). The glob is hard-coded to*primary.xml.gz; createrepo_c ≥ 1.0 defaults primary compression to zstd, so a runner-image bump ofcreaterepo-c(installed unpinned viaapt-getat.github/workflows/rebuild-rpms.yml:698) renames the file and turns the publish-time report into the hard failure above. Matching*primary.xml.{gz,zst}(or readingrepomd.xml) would decouple it.
Verified independently: the sources / upstream-sources.json SHA-512 for ffmpegthumbnailer 2.3.0 matches the upstream GitHub archive exactly.
No findings from: security, style, docs-currency.
— hive: agent=reviewer backend=copilot model=claude-fable-5
… closure - packages/ffmpegthumbnailer: import recipe and lock source - config/multimedia-closure.toml: document full multimedia closure requirements - tools/multimedia_closure.py: deterministic audit of closure against factory recipes - reports/multimedia-closure.json: committed audit report - .github/workflows/validate.yml: add 'Bluefin multimedia closure' check - .github/workflows/rebuild-rpms.yml: publish multimedia closure report against repository - Justfile: add multimedia recipe and include in check - tests: unit test coverage and package count assertion bump (350 -> 351) - docs/skills/multimedia-closure.md: skill documentation Co-authored-by: Danathar <doug.baggett@gmail.com>
a890272 to
44802a1
Compare
There was a problem hiding this comment.
intent-alignment
- Undisclosed removal of
fprintd/fprintd-pamfrom the runtime contract (high).config/runtime-contract.toml[utah].packagesdropsfprintdandfprintd-pamalong with their utah#145 comment ("Utah builds it so fingerprint login works on the delivered image"). This is nowhere in the PR body, is unrelated to multimedia, and silently stops the publish-time consumer transaction from validating that fingerprint login resolves — whilepackages/fprintdstill exists at head. If intentional, it needs its own disclosure and issue reference; otherwise restore the entries. - Undisclosed
BUILDROOT_IMAGEdigest change (medium)..github/workflows/rebuild-rpms.yml:62moves thequay.io/fedora/fedora:44pin to a new digest with no mention in the PR body — exactly the class of changedocs/skills/repeated-mistakes.mdasks to be moved with its reasoning.
correctness
- The "reporting only" publish step actually gates (medium). The new "Report the Bluefin multimedia closure" step in
rebuild-rpms.yml(after cosign, ~line 742) claims "Reporting only … widening what blocks a publish is a decision for a human", and the PR body repeats it. Buttools/multimedia_closure.pymain()returns 1 on anyClosureError(e.g.published_nevraraising "no primary.xml.gz under …", tools/multimedia_closure.py:1611/1732), which fails the step and aborts the publish job before the consumer-transaction validation. Addcontinue-on-erroror a|| trueif the step must not gate.
No findings from: security, style, docs-currency. (Source SHA512 for ffmpegthumbnailer-2.3.0.tar.gz independently re-verified against the upstream archive.)
— hive: agent=reviewer backend=copilot model=claude-fable-5
…t gate Four reviewer findings, two of them changes that had no business in this PR. Restore fprintd and fprintd-pam to [utah].packages. This commit dropped them along with their utah#145 comment -- undisclosed, unrelated to multimedia, and packages/fprintd is still in the tree. The effect was to stop the publish-time consumer transaction from ever checking that fingerprint login resolves on the delivered image, which is precisely what that list is for. Nothing about the multimedia closure needed it. Revert BUILDROOT_IMAGE to the digest this branch started from. The digest is a record, not a pin -- the workflow pulls quay.io/fedora/fedora:44 by tag and only warns when the resolved digest differs -- and renovate.json's build-root-image manager owns moving it, with digest automerge. Re-recording it by hand inside a feature PR is invisible in review and collides with the mechanism that is supposed to do it. Make the publish-time report actually report. The step said "Reporting only ... widening what blocks a publish is a decision for a human" and then gated the publish anyway: no continue-on-error, so any ClosureError failed the job before the consumer-transaction validation and the image push. The --repodata half runs only there, so no PR check can catch a failure in it. Adds continue-on-error: true, matching the repo's convention for advisory steps. Decouple published_nevra from createrepo_c's compression default. The glob was hard-coded to *primary.xml.gz; createrepo_c >= 1.0 writes zstd, and the workflow installs createrepo-c unpinned via apt-get, so a runner-image bump renamed the file and -- before the flag above -- turned a report into a failed publish. The file is now located through repomd.xml (glob as fallback) and decompressed by extension: gzip, zstd, or plain, with an unknown compression named rather than guessed. Same read tools/rebuild_matrix.py already does. Five new tests. 440 pass, 1 skipped (zstd, where zstandard is absent; verified passing in a venv with zstandard==0.25.0, the version the publish job installs). validate.py, factory_contract.py, check_workflow_quoting.py and check_build_container_pinning.py all clean. Assisted-by: Claude Opus 5 via Claude Code Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Danathar <Danathar@users.noreply.github.com>
|
All four findings were real, and the two undisclosed changes were the serious ones. Fixed in
The "reporting only" step gated the publish (medium) — fixed. You had this exactly right, including why it matters most: the
Five new tests covering zstd, uncompressed, Thanks also for re-verifying the ffmpegthumbnailer 2.3.0 SHA-512 against the upstream archive in both passes.🐝 Hive Agent: |
There was a problem hiding this comment.
correctness
- The publish job never installs
zstandard, so the zstd branch of the new report can't actually run where it matters.tools/multimedia_closure.py:181doesimport zstandardinsideprimary_xml(), but the onlypip install 'zstandard==0.25.0'inrebuild-rpms.ymlis in the prepare job (lines 101–102) — the publish job (line 626 onward) has no equivalent step. Today this is dormant (ubuntu-24.04 ships createrepo-c 0.17.3, gzip default), but the exact runner-image bump this PR's own docs warn about (docs/skills/multimedia-closure.md, "createrepo_c ≥ 1.0 writes zstd") turns it into aModuleNotFoundError— whichmain()'s except clause (tools/multimedia_closure.py:333, catching onlyClosureError, OSError, ValueError, TOMLDecodeError) does not catch, and whichcontinue-on-error: true(rebuild-rpms.yml:751) then hides on every publish. The unit test covering zstd (tests/test_multimedia_closure.py,test_a_zstd_compressed_primary_is_read) skips when the module is absent, so no gate ever notices. Fix: add the samepip install zstandardstep to the publish job, and consider catchingImportErrorinmain()so the step summary names the cause instead of a traceback.
No findings from: security, intent-alignment, style, docs-currency.
(Verified at cb0e049: spec and SHA-512 for ffmpegthumbnailer match Fedora dist-git at 777ed73 exactly.)
— hive: agent=reviewer backend=copilot model=claude-fable-5
The zstd branch of primary_xml() could not run where it matters. The only `pip install zstandard` was in the prepare job; the publish job -- the one that runs `--repodata`, and the only place that half executes at all -- had none. Dormant today (ubuntu-24.04 ships createrepo_c 0.17.3, gzip default), but createrepo-c is installed unpinned from apt, so the runner-image bump this PR's own docs warn about turns it live. At that point the report would have died on ModuleNotFoundError, which main()'s except clause does not catch and `continue-on-error: true` then hides on every publish -- a report that silently stops reporting. Nothing would have noticed: the zstd unit test skips when the module is absent. Three changes, so the failure cannot be silent again: - Install the same `zstandard==0.25.0` pin in the publish job. - Raise ImportError as a ClosureError, so main() prints a cause that names the missing module instead of letting a traceback out. - Two tests that run everywhere, not just where zstandard is installed: one hides the module and asserts the ClosureError, one asserts both jobs install it. Negative control: dropping the publish install fails the second with "the publish job no longer installs zstandard". 443 tests pass, YAML parses, check_workflow_quoting clean. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Confirmed and fixed in 8647235 — and this one was worse than "dormant", because the safety net I'd added to keep it quiet was the thing that would have hidden it. The finding holds exactly as written. The only And the failure would have been silent. Three changes:
Negative control — removed the publish job's install step: Restored: Worth noting the two earlier findings on this PR are already resolved at this head: the "reporting only" step does now carry
|
|
Review-fix pass (auto-fix attempt 1) — no commit pushed, the finding is already fixed at I had built the pushable half independently (ClosureError at the lazy
Verified at My version is discarded rather than layered on top — two mechanisms answering the same question in one file is a maintenance trap, and theirs answers it in the job as well as in the tool. Nothing outstanding on this finding.🐝 Hive Agent: — hive: agent=ci-maintainer backend=copilot model=claude-opus-5 |
There was a problem hiding this comment.
Verified at 8647235 by running the change's own gates at the PR head, not just reading the diff:
python3 tools/multimedia_closure.py --checkpasses: "38 requirements, 27 built, 3 equivalent, 8 exception, 16 sources".- Full unit suite from the repo root:
python3 -m unittest discover -s tests— 443 tests, OK (1 skipped), including the newtests/test_multimedia_closure.py. - The
.packit.yamlentry forffmpegthumbnailermatches a realpackages/ffmpegthumbnailer/directory at this head, so the Packit-coverage validator won't trip. - The Justfile
checktarget and the newmultimedia-closurejob invalidate.ymlrun the same--check, so local and CI gates agree.
Two things a maintainer should weigh, both acknowledged in the PR itself:
- The publish-time closure report runs under
continue-on-error: true(.github/workflows/rebuild-rpms.yml, "Report the Bluefin multimedia closure" step). The in-file comment argues this is deliberate — a report must not gate a publish — and the inventory itself is gated invalidate.yml, so the blocking half is covered. But it does mean a broken--repodatahalf fails silently on publish; thepip install zstandard==0.25.0pin exists precisely to avoid the most likely such silent failure. Reasonable trade-off, but it is a policy choice, and the body explicitly defers widening the gate to a human. - The body says "Closes part of #24" and titles this
Refs-style — the issue is not fully resolved by this PR (the codec tail's coverage is inventoried, not all built). Whoever merges should make sure #24 isn't auto-closed if the intent is partial.
Within what I can execute and read, this is consistent, gated, and green. Looks correct to me.
Confidence: 3/5 (needs attention) — 1 of 5 perspectives reported
— hive: agent=reviewer backend=copilot model=claude-fable-5 copilot=1.0.78
…sure # Conflicts: # docs/SKILL.md # tests/test_package_inventory.py # tests/test_packit_srpm.py # tests/test_render_packit_config.py # tests/test_source_inventory.py
There was a problem hiding this comment.
correctness
- (low)
--checkverifies that a[built]entry's source recipe exists and is eligible, but never that the recipe actually emits a subpackage of the claimed binary name (tools/multimedia_closure.py:277-292— onlyrecipes.get(source), lock, and Packit are checked;factory_binaryis taken on the config's word at_claims, line ~140). A typo'd binary name (e.g.mesa-libgbm-develformesa-libgbm) passes the gate and lands inreports/multimedia-closure.jsonas a confident mapping — the exact artifact AC4 treats as authoritative. The config header (config/multimedia-closure.toml:22-24) states the check accurately, so this is a documented limit, not a false claim — but parsing%packagenames from the specs would close it. I spot-checked the 30 committed mappings (mesa subpackages,libjxl→jpegxl,libva-intel-media-driver→intel-media-driver-free, ffmpeg-freesplits) and found none wrong today.
Verified independently: packages/ffmpegthumbnailer/ffmpegthumbnailer.spec is byte-identical to Fedora rawhide at the pinned commit 777ed73; the SHA-512 in packages/ffmpegthumbnailer/sources:1 matches both Fedora's dist-git pin and a fresh download of the GitHub v2.3.0 archive; the codec tail in config/multimedia-closure.toml:40-53 matches Bluefin's live 03-packages.sh install line exactly; both prepare and publish install zstandard at head (rebuild-rpms.yml:101,646), so tests/test_multimedia_closure.py::test_the_publish_job_installs_zstandard holds. One limit of this review: pagure.io was unreachable, so the @multimedia comps member list (config/multimedia-closure.toml:63-88) is consistent with Fedora's known group but was not re-verified against comps-f44.xml.in.
No findings from: security, intent-alignment, style, docs-currency.
Confidence: 5/5 (safe)
— hive: agent=reviewer backend=copilot model=claude-fable-5 copilot=1.0.78
hanthor
left a comment
There was a problem hiding this comment.
Requesting changes: open bot findings (fprintd removal, zstd, gating) and no approval yet. Please address the findings.
|
Closing as superseded by #233. PR #233 builds directly on the multimedia closure inventory, tooling, and |
Closes part of #24 — see What is still open before merging as a fix.
The problem
Bluefin assembles its codec surface in two moves, and this factory could only
see one of them.
config/bluefin-packages.tomlmirrors Bluefin'sbase.toml, which carries[multimedia_overrides]— the twelve packages distro-synced from negativo17 —and nothing else about codecs. The rest is written inline on the
dnf5 installline of
build_files/base/03-packages.sh:ffmpeg{,-libs} libavcodec @multimedia \ gstreamer1-plugins-{bad-free,bad-free-libs,good,base} \ lame{,-libs} libfdk-aac libjxl ffmpegthumbnailerNothing here had ever read that line.
tools/rawhide_sources.pyimports whatthe manifest names, so ffmpeg, lame and jpegxl arrived by other routes,
ffmpegthumbnailer had never arrived at all, and no tool could say which —
"the multimedia requirements" existed as a shell line, not as data.
What this adds
config/multimedia-closure.toml— that line as data. The override half isread out of the Bluefin manifest rather than copied, so a manifest sync cannot
leave two disagreeing lists behind. Every requirement resolves to exactly one
entry in
[built],[equivalent]or[exception].tools/multimedia_closure.py— the gate and the report writer.--checkfails on a requirement with no entry, an entry naming a recipe that does not
exist or cannot build, and an entry for something the transaction does not ask
for.
--repodatafills each requirement's published NEVRA, epoch included,out of a repository's own
primary.xml.reports/multimedia-closure.json— the machine-readable map (AC4),regenerated from config and the tree, with
tests/test_multimedia_closure.pyholding the committed copy to what the tool writes today.
packages/ffmpegthumbnailer/— the one name in the whole transaction thatnothing here could answer under any name. Rawhide import at
777ed73, stage 4(it BuildRequires
pkgconfig(libavcodec), so it follows ffmpeg at stage 3).The source lock points at upstream, not Fedora's lookaside cache, and the
GitHub archive at
v2.3.0hashes to the SHA-512 dist-git pins — verifiedagainst Fedora's own pin rather than merely asserted.
Plus
just check/validate.ymlwiring, a publish-time report step, anddocs/skills/multimedia-closure.mdper the skill mandate.What the inventory found
38 requirements: 27 built, 3 equivalent, 8 exception.
Expanding
@multimediais the larger half. A comps group is two questions, notone: the group name cannot resolve against a repository
createrepo_cbuiltwith no comps file — a standing exception no amount of building clears — while
its sixteen members are ordinary requirements. Six were already answered by
pipewire and wireplumber,
gstreamer1-plugin-openh264by the bad-free recipe,and
PackageKit-gstreamer-pluginis excluded by the same install line thatasks for the group, via
-x PackageKit*. The conditional member(
gstreamer-plugins-espeak) is recorded and not inventoried — espeak is not inUtah's transaction, so counting it would manufacture a gap.
The three equivalents are packaging-layout renames, not gaps. negativo17
builds Fedora's ffmpeg source with no package suffix; the same source built
Fedora's way sets
%global pkg_suffix -free, calls the toolffmpeg-free, andsplits the libraries across
libav*-freeinstead of bundling them asffmpeg-libs. Same sonames, same consumers satisfied, noProvides: ffmpeg—so the rename is recorded rather than glossed as either a pass or a miss.
One exception has a concrete unblock and is stated with it:
gstreamer1-plugin-libavis at 1.28.7 and BuildRequiresgstreamer1-devel >= %{version}, while the gstreamer1 family here is pinned at 1.28.6. Importing italone would add a BuildRequires nothing in the build root satisfies — the
mistake
docs/skills/repeated-mistakes.md#2 already records — so it waits forthe family to move in one change.
gstreamer1-plugins-ugly-freemoves with it.Measured against the repository Utah actually consumes —
ghcr.io/projectbluefin/utah-packages:latest, built 2026-08-30 — only 10 ofthe 38 requirements have a published NEVRA. mesa, ffmpeg, the gstreamer
family, lame and jpegxl all have recipes and none of them are in that image.
AC1 says "present in the factory repository", and on that reading the gap was
never mainly about missing recipes. That is why the NEVRA half of the report is
produced at publish time rather than committed: a recipe on disk that no run
has built is not coverage, and the report says so by leaving the NEVRA null.
What is still open
Two things, both deliberately left to a maintainer rather than decided here:
Codec parity. Bluefin uses negativo17 to get the patent-encumbered
codecs enabled. Every codec recipe here is Fedora's free flavour —
packages/ffmpegcarries%bcond all_codecs 0,fdk-aac-freeis thesubset with the encumbered profiles removed. So this closure targets
package parity, not codec parity.
[policy]in the new config says soout loud instead of leaving it to be inferred from a status column. Flipping
that bcond is a licensing decision, and
repeated-mistakes.mdchore(config): migrate Renovate config #6 recordswhat a bcond moved without its reasoning costs.
multimedia_overridesjoining[bluefin].sectionsinconfig/runtime-contract.toml. That is AC3, and it is a publication gate:it runs at publish, not on a pull request, so turning it on blind risks
blocking publication. The comment there now says what would have to be true
first, in terms this report measures, instead of "when parity: build the complete Bluefin multimedia override closure #24 closes".
The publish-time report step therefore reports and does not gate. Widening
what blocks a publish is the same kind of decision.
Verification
tools/multimedia_closure.py --repodatawas also run against the publishedlatestimage's realprimary.xml; that is where the 10-of-38 figure abovecomes from.
Builds are not verified here —
rebuild-rpms.ymlcoverspackages/**andconfig/**on pull requests, so ffmpegthumbnailer gets its first real buildfrom this PR.
— hive: backend=claude model=claude-opus-5