Five claims in the skill's Core Behavioral Spine that misstate the contract they condense

[ptr727]

The ## The Core Behavioral Spine section of .agents/skills/workflow-ci-contract/SKILL.md is the condensed statement of the contract an agent reads before touching a workflow. Five of its claims either cannot be implemented as written or drop the qualifier that carries the rule. Found by a whole-unit canonical pass on PR #1203's fix branch, and deferred out of it, which corrected only the one sentence a reviewer had raised.

The aggregator described here is a dead gate

one required aggregator gates the merge, treating skipped smoke as pass and blocking on failure or cancelled

The behavior cannot be implemented as stated, because if: always() is the mechanism that makes it possible and nothing in this SKILL.md mentions it anywhere. D1.5 requires it explicitly, "run under if: always() so a failed or skipped dependency cannot skip the gate itself".

Without it the aggregator inherits success() and skips whenever smoke skips, so it does not treat skipped smoke as pass, it stops being a gate at all, and GitHub reads a skipped required check as satisfied. That is precisely the failure D1.5's Prevents clause names, a paths-filter error letting a target-changing PR merge unbuilt. A reader following this unit literally builds a gate that looks right and is dead in its main case.

The expect_release_assets rule inverts for NuGet

A repo with no file target passes expect_release_assets: false at the caller.

"File target" is undefined in the unit, and the obvious reading is wrong for NuGet. D4.3 spends a sentence on exactly this case: "A NuGet target is not among them, since its leaf uploads a release-asset-* carrying the package, so a NuGet-only caller keeps the default true." The no-file-target shapes are Docker-only, PyPI-only and source-only.

A reader with only this unit reasons that a NuGet package goes to a registry rather than onto the release, sets false, and thereby relaxes fail_on_unmatched_files and skips the asset download on a repo that does produce an asset. The PyPI and NuGet asymmetry is not derivable from the unit.

A portable guarantee is stated as a literal

main a clean X.Y.Z, anything else a prerelease X.Y.Z-g<sha>

WORKFLOW.md states this portably and flags divergence on it: "Guarantees say 'default branch' portably. It is implemented as the literal main in several places (the validate gate, the prerelease expression, and version.json's publicReleaseRefSpec). These MUST all reference the repo's actual default branch. A divergence is a defect (section 5A)." D3.2 repeats it.

The unit's other uses of main, the bot push and the weekly Docker schedule, track D4.1, which is main-literal by contract. This one clause is the place the contract is deliberately portable, and the unit gives no cue that the literal is repository-specific. registry/repos.json carries no defaultBranch field, so nothing downstream re-derives it either.

The artifact rule drops its only exemption

Artifacts are an intra-run handoff: consume-then-delete at the point of consumption

D5.1 scopes the requirement to a cross-job transfer artifact and adds "An intermediate consumed only within the same run MAY instead rely on the retention-days: 1 backstop." The unit names that same property, "intra-run handoff", as the description of everything the rule covers, so the MAY-case disappears and the rule reads as unconditional.

In the same bullet, "gated to the half of the consumption whose failure would leave the artifact not yet redundant" is true but not actionable from the unit. Nothing in it lets a reader reach D5.2's if: ${{ !cancelled() && steps.<download-step-id>.outcome == 'success' }}, nor the reason it is required, that a delete carrying no status-check function inherits success() and skips on exactly the failed-push path it was written for.

The methodology reference is an index, not the procedure

Auditing, tracing, and probing a repo's workflows is references/test-methodology.md.

That file says otherwise three times. Its 5A section: "read this as an index into them rather than as the sweep itself" and "Apply only the ones the repo's types imply, and read them in WORKFLOW.md 5A rather than from this list". Its 5B section gives only "The load-bearing ones", S1, S5/S6, S7, S8, S9, S10 and S12/S13, and defers the expected end-state table to WORKFLOW.md 5B. Its Verdict section puts the per-type walkthroughs in WORKFLOW.md section 6.

An agent told the audit "is" this file runs a sweep missing the per-type addenda and the unlisted scenarios, and reports a verdict anyway.

Scope note

A sixth finding, that the adjacent sentence called references/d-guarantees.md "The full catalog, each guarantee with the failure mode it prevents", was corrected on the fix branch to "The condensed catalog is in references/d-guarantees.md", because a reviewer raised the stale framing and the file calls itself condensed. The underlying gap that made "full" wrong is unchanged and belongs to #1205: WORKFLOW.md section 4 carries 44 numbered items and that file carries 39, collapsing all five D9 items into one unlabeled line with no addressable ids, so D9.1 and D9.2, the two this skill's own frontmatter singles out as easy to break, cannot be cited from the file the skill routes to.

Changing this file means regenerating .github/skills/ and .claude-plugin/ with scripts/build_dist.py. Related to #1200, #1204, #1205, #1206, #1207.

[ptr727]

A second whole-unit pass over the same unit adds three findings to the five above, and re-confirmed the rest.

The spine restates the file it introduces as the place the catalog lives, and nothing detects divergence between the two copies. The tail of every upload-artifact gated on smoke being false, which is \!inputs.smoke` at the workflow layer and `inputs.smoke != 'true'` in a composite action, whose inputs are stringsis byte-identical tod-guarantees.mdD1.3.strips +buildmetadataandon smoke skips the check while the job still succeedsare identical to D2.2. Thenever artifact-ids:clause plus its canonical-for-single-target sentence duplicates D6.1.a package target's separate publish- job needs: the release-task call, so no build failure ships anything partial. A failed package push is outside that` duplicates D4.5 clause for clause. An edit to either copy leaves the other stale, and the paragraph's own framing tells a maintainer the reference is the single place to edit.

**The release tags the built commit's SHA (\GitCommitId`), never a branch name.** is imprecise against D4.2 and drops the one auditable name. D4.2's output is that "the release \target_commitish` is the built commit's SHA (NBGV's `GitCommitId`), never a branch name or a separately re-resolved ref". A release's tag name is the version, and it is target_commitish that carries the SHA, so "the release tags the SHA" reads as a claim about the tag name and leaves an auditor no field to grep for. The unit also drops D4.2's second prohibition, "never a separately re-resolved ref".

the caller's own job reaching the reusable validator, or the replacement it points its aggregator at, always runs has no antecedent for "it". The actor in D1.2 is "A repo whose validation it cannot express", and this unit introduces no repo, so a first-time reader cannot tell who does the re-pointing. The unit also drops D1.2's condition entirely, presenting a replacement validator as an ordinary alternative rather than as reserved for a repository the reusable validator cannot express, and drops "replaces the call (not deletes it)", which is the half that keeps the aggregator's needs: from dangling.

Checked and sound in both passes: the release-asset-<branch>-<target> name shape and the pattern:/merge-multiple: collection (D6.1, and build-release-task.yml:317), expect_release_assets: false for a no-file-target caller (D4.3), the publish triggers including the main-only weekly Docker schedule (D4.1), the two-directional release gate (D2.2), the !failure() && !cancelled() guard and the failed-package-push carve-out (D4.5), retention-days: 1 with best-effort cleanup and the no-blanket-delete rule (D5.3 to D5.5), and the no-op republish and server-side dedupe clauses (D4.4).

[ptr727]

Correction to this issue's scope note. It quotes the shipped pointer sentence as "The condensed catalog is in references/d-guarantees.md", which was an intermediate state on the fix branch and not what ships. A later whole-unit pass found that wording carried no antecedent, since the unit never uses the word catalog and cites no D-number, so a reader could not tell what was cataloged or what it was condensed from. The shipped sentence names the source:

A condensed catalog of WORKFLOW.md section 4 is in references/d-guarantees.md.

The substance of the scope note is unchanged. The reason "full" was wrong in the original still stands and still belongs to #1205.

[ptr727]

Scope update: one of the five is now fixed, so four remain.

The fifth, "The methodology reference is an index, not the procedure", was raised independently by Copilot on #1209 and fixed there in ef0b0f0, because the sentence beside it was already moving on that branch. SKILL.md:43 now reads:

A condensed guide to WORKFLOW.md section 5's audit, trace, and probe procedure is in references/test-methodology.md.

"Condensed guide" rather than "index" deliberately. A pass found that "index" takes the file's own word for its 5A part while understating 5B, which states the load-bearing scenarios and their expected outcomes rather than pointing at them.

Two corrections to this issue's own text while I am here. The Scope note quotes the adjacent fix as "The condensed catalog is in references/d-guarantees.md", already corrected in an earlier comment: the shipped wording is "A condensed catalog of WORKFLOW.md section 4 is in references/d-guarantees.md". And the title says "Five claims", which is now four.

The four that remain, all unchanged and all still worth settling together: the aggregator described without if: always(), which cannot be implemented as stated; expect_release_assets: false inverting for NuGet; the main literal where D3.2 is portable; and the artifact rule dropping D5.1's same-run exemption. The three added in the second comment above also stand.