Repository navigation
Description
Reviewers and the explorative testers of lanes #55 (PR #57) and #56 (PR #67) raised these follow-ups against the images job's render and its test runner. Each is a bullet; the first two are the ones with a security reading.
- The render image by mutable tag without
--entrypoint. Where:.github/workflows/ci.ymlimages job (RENDER_IMAGE: ubuntu:26.04) andscripts/verify_image_release.shrender(). The image is pulled by mutable tag, anddocker runpassesflux build ...as the command without--entrypoint; an image that sets an ENTRYPOINT would receive those words as arguments and could print a render of its choosing. Suggested fix: pin the image by digest, or pass--entrypoint /usr/local/bin/flux. The bootstrap job's unpinnedubuntu:26.04already exists on main, so whether this attacker is in scope is the trust model's call. - The render has no timeout of its own. Where:
scripts/verify_image_release.shrender(); a second Kustomization withspec.path: ./clusters/pc(the root's own path) makesflux build --recursiverun without end (scratch treef6-recursive-self-pathhung past 150 s, timeout exit 124, no output). In CI the job times out at 10 minutes, so the check goes red without a message naming the Kustomization. Suggested fix: a render timeout that becomes a red record naming the render. - Two owners decide which sources are PC-only. Where:
scripts/verify_image_release.sh(PC_ONLY_SOURCE),.github/workflows/ci.ymlcluster overlay patches, AGENTS.md cluster line. AGENTS.md's "a new PC-only source joins that list" names only the overlay; drift fails loud (an unmapped source makes the render red), but the list has two owners. Suggested fix: one owner for the list, or the AGENTS.md line naming both. - The runner does not assert the G2 log lines. Where:
tests/images/run.sh. The contract's "each substituteFrom Secret is named in the log as a gap" is not asserted, since the runner compares only exit status and red set; the G2 lines were seen printed for all three Secret cases, but nothing catches a regression. Suggested fix: compare the gap lines too, or rule it out. - The runner's header omits
RENDER_IMAGE. Where:tests/images/run.shheader. The runner exportsGITHUB_*andDEFAULT_BRANCHbut notRENDER_IMAGE; run as its header says (mise exec -- tests/images/run.sh) it fails every case with exit 2 "RENDER_IMAGE is not set"; only CI's job env supplies it. Suggested fix: nameRENDER_IMAGEin the header or default it in the runner. - Unknown-kind reds print numeric field paths. Where:
scripts/verify_image_release.shjudge_unknown_kinds(FIELD_PATH). They print.spec.template.spec.containers.0.imagewhile Pod-template reds print.containers[main].image. Cosmetic; both name object, path and image. Suggested fix: one path form, or rule it out. - A short ref in a custom kind is unjudged. Where: the unknown-kind candidate rule. A custom kind whose field holds a Docker Hub short ref (a Widget with
spec.imageName: nginx:1orlibrary/nginx:1) is neither red nor judged; within the contract (host-qualified only), but an unjudged way to name an image. Suggested fix: name it as a gap on Platform CI refuses every image reference that is neither an attested app release nor a listed vendor at a digest #49 or extend the candidate rule. - A
platform-secretsurl with another owner case is red. Where: the PC-only source url check.OvernightWorks/platform-secretsis red as "not the repository overnightworks/platform-secrets", although GitHub treats the owner case-insensitively; a false red in the safe direction, and the committed url is lowercase. Suggested fix: compare the owner case-insensitively, or rule it out. - The missing vendor list message. Where: the vendor list read in
scripts/verify_image_release.sh. With the file absent the red record says "yq cannot read the vendor list" (exit 1); it does not say the file is missing. Suggested fix: test for the file and say so. - A kustomize component under
.github/is judged. Where: the render and the contract's source-field text. The script renders and judges such a component (red when it adds an unlisted image), while Flux's artifact excludes.github/by default, so on the PC that layer would fail to build. The direction is safe (CI sees a superset), but the contract text does not name it. Suggested fix: name it in the contract text or in AGENTS.md, or rule it out. - The scratch draft PR showed two red causes. Where: the evidence for the verdict wiring. The scratch draft PR also turned the cluster job red (the root's inventory held a Deployment without a health check), so
quality / verdictwas red from two causes and that run alone does not show thatimagesby itself turns the verdict red; the wiring is confirmed byquality/config.tomlmandatory_jobsand theci.ymlneeds list. Suggested fix: one scratch draft PR with only an images red, or rule it out. - An app ref in a private or non-existent owner repository is red with identical text (from B1b, B1b: the job images checks each app ref's tag, attestation and signer commit #56). Where:
scripts/verify_image_release.shresolve_tag_commit. In CI both are red withgit ls-remote ... failed: fatal: could not read Username for https://github.com: terminal prompts disabled; it fails closed, but the message does not say the repository is private or absent. hopin is public, so nothing is affected today. Suggested fix: say so in the red record when a private app repository appears, or rule it out. - E4's other-workflow half was not driven (from B1b, B1b: the job images checks each app ref's tag, attestation and signer commit #56). Where: the attestation check in
scripts/verify_image_release.sh, which passes--signer-workflow overnightworks/quality/.github/workflows/image.yml. Whether gh's SAN pattern rejects a lookalike workflow file name (such asimage.yml.bak) or another workflow of quality needs a real attestation from such a workflow; none exists, so the case is judged only from the flags. Suggested fix: drive it once a lookalike or other quality workflow attests an image, or rule it out.
Done when each bullet is fixed or ruled out with its reason.
version = 1
now = "B1a landed with these follow-ups open; none changes what its job refuses today."
next = "Pull when the images job is next touched, or before a second PC-only source is added."
done_when = "Each bullet of this item is fixed in scripts/verify_image_release.sh, tests/images or .github/workflows/ci.yml, or ruled out with its reason in the item's journal."
scope = [".github/workflows/ci.yml", "AGENTS.md", "scripts/verify_image_release.sh", "tests/images"]
size = "S"
whole = "The bullets all concern the one images job and its runner, and a fix to one (the render image, its timeout, the PC-only list) changes the lines the others read."
Metadata
Metadata
Assignees
Labels
No labels
Activity
[-]B1a's follow-ups after its first landing[/-][+]B1's follow-ups after B1a and B1b landed[/+]FlexOr2 commentedon Nov 9, 2026
Lane #65, PR 247. Bullets ruled out, with the builder's reasons; the review and gate judge them:
tests/flux/reconcile_the_candidate_tree.shis a kustomize patch pair per source. It swaps the source's spec and the sibling Kustomization'ssourceRef, which a name from the script cannot derive. The script also runsmainwhen sourced. An unmapped source fails the render, so the two lists cannot drift silently. AGENTS.md now names both.platform-secretsURL: a case-insensitive compare would pass a URL that is red today. The false red is in the safe direction, and the committed URL is lowercase..github/kustomize component: the safe direction (red when it adds an unlisted image). That Flux's artifact excludes.github/is not verified here.quality/config.tomlmandatory_jobsand theneedslist.git ls-remoteprints the same text for both. The check fails closed.FlexOr2 commentedon Nov 9, 2026
Correction to the journal above: 6 bullets are fixed (the render image, the render timeout, the PC-only sentence in AGENTS.md, the G2 gap lines, the runner's RENDER_IMAGE, the missing vendor list) and 7 are ruled out (numeric field paths, short ref in a custom kind, owner case, the .github/ component, the scratch draft PR, private or absent app repository, E4's other half). The matrix review of PR 247 found every reason true; bullet 7's gap already stands on #49.
FlexOr2 commentedon Nov 9, 2026
Journal, from the final gate of PR 247:
sourceignore.ExcludeCIlists.github/, andflux buildreads the checkout, a superset.signerWorkflowSANPatternanchors both ends, so neither a lookalike nor another quality workflow matches.FlexOr2 commentedon Nov 9, 2026
Landed in #247 at 3afb204.