Repository navigation
ci: reproducibility covers every shipped element, shared CI scripts, doc drift - #403
Conversation
…isk setup reproducibility.yml's FINAL_ASSEMBLY missed the NVIDIA driver and Container Toolkit sysexts, the SBOM and the homelab templates, so their cached artifacts were reused by the second build and never compared. tests/unit/test_ci_workflows.py derives the shipped set from what oci/bluefin-server-image.bst stages plus the oci/* targets of `just validate` and fails when FINAL_ASSEMBLY misses one. The three copies of the runner disk setup (two in build.yml, one in reproducibility.yml) had drifted between `ln -sf` and `ln -sfn`; scripts/ci-runner-disk.sh replaces them with the idempotent `-sfn`. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The Justfile's fsdk_version and track-junctions.yml each re-parsed the junction ref with grep/sed. check-release-version.py gains --print-fsdk (the Justfile) and --fix (track-junctions: sync installer-version, then check). Also correct the Justfile comment: the bst2 image is pinned by a commit-named tag, not by SHA. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
.github/scripts/propose-pr.sh commits the given paths onto the tracker branch, force-pushes with a lease and creates or edits the PR. It pushes nothing when an open PR's branch already holds the same content, which stops track-junctions from force-pushing a fresh commit (and a full FSDK build) every day; track-binaries keeps not reopening a closed title via --skip-closed-title. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
One canonical figure in ci-tooling.md (build job on main: median ~50 min, p90 ~2 h over 56 runs; uncached kernel 1 h 43 min) that build.yml, kernel-cache.sh and image-build-needed.py link to. Also: the just install sites are six steps in four workflows, the old track-refs job carried freedesktop-sdk junction bumps (there is no gnome-build-meta junction), and the Core Process list is no longer mis-indented. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
tests/unit/test_ci_workflows.py also lands in #401 with other contracts (including the single just version), so the reproducibility coverage and the runner disk tests move to test_reproducibility_workflow.py and test_ci_runner_disk.py. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
There was a problem hiding this comment.
Reviewed at a72b225 (correctness, security, intent-alignment, style, docs-currency).
correctness
tests/unit/test_propose_pr.py:114(test_opens_a_pull_request_on_a_new_branch) asserts the author isgithub-actions[bot], butpropose-pr.sh:77-79sets identity withgit -c user.name/user.email, whichGIT_AUTHOR_NAME/GIT_COMMITTER_*env vars override. Theenvfixture (test_propose_pr.py:62-71) inheritsos.environ, so on any machine that exports those vars the test fails ('reviewer' != 'github-actions[bot]'here; passes with them unset, 13/13). Either strip/setGIT_AUTHOR_*/GIT_COMMITTER_*in the fixture (as thegit()helper at:38-40already does) or export them in the script so the bot identity holds regardless of the caller's environment. CI runners don't export these, so this is a dev-machine flake rather than a CI failure.
intent-alignment
- PR body: "
justis installed in six steps across four workflows, and a test checks that all pins are the same." No such test is in the diff — the onlyjust@1.58.0references added are workflow steps anddocs/skills/ci-tooling.md:56-59;tests/unit/at head has no test reading thejust@pin (checkedtest_release_workflow.py,test_ci_runner_disk.py,test_reproducibility_workflow.py). Either add it or drop the claim. (Body also names the new test filetests/unit/test_ci_workflows.py; it istests/unit/test_reproducibility_workflow.py.)
Verified: all 14 FINAL_ASSEMBLY entries are location: deps of elements/oci/bluefin-server-image.bst:60-104 and every just validate oci/* target (Justfile:73) is covered; bluefin-server-sbom.bst stamps creationInfo.created from fixed SOURCE_DATE_EPOCH (element description), so adding it to the reproducibility set is sound; --force-with-lease=refs/heads/B: with empty expect correctly requires the branch not to exist (propose-pr.sh:80).
No findings from: security, style, docs-currency.
Confidence: 3/5 (needs attention) — 2 perspectives requested changes
— hive: agent=reviewer backend=copilot model=claude-fable-5.1 copilot=1.0.88
Keeps #402's tracker changes (trackers environment, token minted after the tracking steps and narrowed, persist-credentials: false) and moves its one-off extraheader push into propose-pr.sh, which both trackers now call. Assisted-by: Claude Opus 5.5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add tests/unit/test_ci_tool_pins.py asserting every workflow installs the same just pin, as the ci-tooling doc describes. Strip inherited GIT_AUTHOR_*/GIT_COMMITTER_* from the propose-pr test env so the bot author assertion does not depend on the caller's environment. Hive-Fix-Attempt: 1/6 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: scanner <scanner@hive.kubestellar.io>
Review findings M4, L2, L3 and the doc-drift list from the GitHub Actions review. The kernel-cache work (M5) is in a separate PR.
M4: reproducibility rebuilds everything that ships
FINAL_ASSEMBLYnow also hasoci/nvidia-open-595-sysext.bst,oci/nvidia-container-toolkit-sysext.bst,oci/bluefin-server-sbom.bst(it ships asbluefin-server_<ver>.spdx.json) andhomelab/homelab-templates.bst(the.bu/.ignrelease assets). Before this, the second build reused their cached artifacts, so they were never compared.tests/unit/test_reproducibility_workflow.pybuilds the shipped set from two sources: every dependencyoci/bluefin-server-image.bststages at alocation, plus everyoci/*target ofjust validate. The test fails ifFINAL_ASSEMBLYmisses any of them. Checked: on the old workflow it names exactly the four missing elements.cron: '0 9 * * 1', so the first run is due Mon 2026-10-05 09:00 UTC. Other scheduled workflows in this repo do fire.L2: runner disk setup
scripts/ci-runner-disk.shreplaces the three inline copies (two in build.yml, one in reproducibility.yml). It usesln -sfn, so running it twice no longer nests a symlink inside/mnt/buildstream.sudo. Another test makes sure no workflow inlinesgraphroot/ln -sfagain, and that every job that builds calls the script.L3: one FSDK version parser and one tracker PR script
check-release-version.pygains two options:--print-fsdk, which the Justfile'sfsdk_versionnow uses.--fix, which track-junctions now uses: it syncsinstaller-versionand then runs the same check..github/scripts/propose-pr.shis the create-or-update-PR step for both trackers. It pushes nothing when:--skip-closed-title.--force-with-leaseon the branch state it saw. A failedghorls-remotestops the run before any push.gh.Doc drift
ci-tooling.md#build-time-and-cachesnow holds the one build-time figure: thebuildjob on main has a median of ~50 min and a p90 of ~2 h (56 runs, 2026-09-24..10-04), and the kernel takes 1 h 43 min when uncached. build.yml, kernel-cache.sh and image-build-needed.py link to it instead of repeating their own (conflicting) numbers.justis installed in six steps across four workflows, andtests/unit/test_ci_tool_pins.pychecks that all pins are the same.Gates
just validate, pytest (1103 passed; also in a detached worktree withoutfiles/boot-keys), bats, docs-checks, shellcheck, actionlint (only the existing SC2129), and zizmor (same 7 medium + 1 high as main, no new findings).Found while checking: both tracker workflows currently fail at "Get mergeraptor token" with
Not Found - get-a-repository-installation. The mergeraptor App does not look installed on this repo. That is a settings matter and is not changed here.