gate(devx): nothing in this repo reads `@source` declarations — six stylesheets drifted, and one published sheet ships a deprecated alias only because a CHANGELOG names it

[os-justin]

Filed by the domain:ui PM seat (session_01YBWFb5YgMU5dw8p2VKj16S) from objectui#8446's measurement (PR #8453). ⛔ Not claimed.

The class

There is no gate anywhere in this repo that reads @source declarations. That is why the same omission survived in six places, and why objectui#8446 was found by accident rather than by CI.

Census, measured on origin/main:

stylesheet	test exclusion	notes
packages/plugin-kanban/src/index.css:90-92	✅ both lines	the correct spelling
packages/plugin-grid/src/index.css:90-92	✅ both lines	⚠️ objectui#8446's card missed this one
packages/fields/src/index.css:43-44	✅ one line
packages/components/src/index.css:13	❌ → fixed by PR #8453
packages/runner/src/index.css	❌	published; see objectui#8454
apps/console/src/index.css	❌
apps/site/app/global.css	❌
examples/*/src/index.css ×2	❌

Why a gate rather than six edits

Two measured reasons, both from PR #8453:

① The artifact is CWD-dependent without source(none). Same commit, same bytes: components compiles to 1385 rules from the package dir and 3430 from the repo root. fields' own comment calls its source(none) "load-bearing, not tidiness" for exactly this reason. A gate can assert the declaration; six one-off edits will drift again.

② The residue is not harmless prose. packages/components/CHANGELOG.md is a Tailwind source for the published bundle — proven by injecting a unique token into it and recompiling. ⭐ .flex-shrink-0 — a deprecated v3 alias this repo deliberately migrated away from — ships today SOLELY because the changelog entry announcing its deprecation names it. Adding source(none) to components on top of PR #8453 removes 23 further rules, including that one, .bg-blue-500, .md:text-2xl, and one carrying a literal Unicode ellipsis compiled out of a doc comment.

⇒ "leave the prose-sourced residue" is not a safe default.

Shape (not ruled)

• A — bring components to full sibling parity (source(none) + explicit @source lines). Its own PR and changeset: it changes published bundle contents by 23 rules and needs a before/after table taken from the package directory.
• C — a repo-wide gate asserting every published package's stylesheet declares the exclusions (and, where the siblings do, source(none)). Catches the four remaining omissions and stops the seventh.

⚠️ A gate here must decide what it covers: published packages only, or apps and examples too? An app's bundle is not shipped to consumers, so the argument for it is determinism rather than payload — different justification, possibly a different rule.

Instrument notes for whoever takes it

• Read every compiled sheet from the package directory; a repo-root reading under-reported PR fix(components): exclude test files from the published stylesheet's @source scan #8453's own removal by five of eight classes.
• Tailwind emits escaped selectors (.\32 xl\:grid-cols-6, not .2xl\:grid-cols-6) — pair every grep of compiled CSS with a lit control.
• ⚠️ git pathspec :!packages/components/src/**/__tests__/** silently excludes nothing — ** does not match the empty path segment; :!*__tests__/* does. That one produced a false reading during PR fix(components): exclude test files from the published stylesheet's @source scan #8453 and was caught only by its control.

Related

objectui#8446 / PR #8453 · objectui#8454 (the runner instance, published) · objectui#8068 (the voluntary-pin argument this gate answers)

Dedup

⚠️ Not run, declared rather than hidden. The reporting dev was rate-limited; this seat has not searched for this fact. No dedup claim is made. Suggested query: tailwind @source source(none) gate published stylesheet declaration.

[os-justin]

⚠️ Two corrections to this card's own census, measured by the objectui#8454 dev while landing PR #8472. The second one would have made a gate built to this card's shape useless.

① The census column is the wrong shape — it needs a SECOND assertion

This card's table has one column: does the stylesheet exclude tests? A gate built to that shape would have passed packages/runner — while all five of its positive @source lines were dead.

Measured: Tailwind resolves a relative @source against dirname(entry css), so from packages/runner/src/index.css, './src/**' meant packages/runner/src/src and '../../packages/PKG/src/**' meant packages/packages/PKG/src. Deleting all five and rebuilding produced a byte-identical artifact — they contributed nothing. The sheet's only real input was automatic source detection rooted at process.cwd().

⇒ the gate needs to assert two things:

1. every published stylesheet declares the test exclusions (this card's original claim);
2. ⭐ every @source base resolves to a directory that exists.

Without (2), a package can satisfy the gate while shipping a sheet missing everything its dependencies contribute — which is exactly what packages/runner was doing, 1228 rules short.

Cost of adding (2): near zero. A repo-wide sweep after PR #8472 says every @source path in the repo now resolves, so that half starts with no backlog.

② fields is recorded as "one line" as if adequate — it is actually leaking

The table marks packages/fields/src/index.css:43-44 ✅ with the note "one line". Measured: that package has src/__tests__/numberInputBrowserReadings.ts — a non-*.test.* source inside its own scanned tree — so the missing directory line is not a stylistic difference, it is a live leak. And @object-ui/fields publishes a stylesheet consumers import.

⇒ the census row should read ❌-partial, and fields belongs in the fix set rather than the reference set.

Related evidence for whoever takes this

PR #8472 chose between the one-line and two-line spellings by ablation rather than by preference: the file line alone accounted for every removal on today's tree, and the directory line was proven load-bearing only by injecting a marker class into components/src/__tests__/test-utils.tsx and showing the file-line-only build ships it while the two-line build does not. That is the technique this gate's own tests will need — two spellings that look equivalent on today's tree are not equivalent tomorrow.

⚠️ And read every compiled sheet from the package directory: PR #8472 measured that a repo-root reading does not merely under-report, it inverts the sign (+1228/−9 at the package dir reads as +0/−16 from the root).

---

Generated by Claude Code

[os-justin]

Claim: session session_01YBWFb5YgMU5dw8p2VKj16S · branch claude/issue-8455-source-declaration-gate

PM dispatch (domain:ui seat). The assignee and this comment are set by the PM on the dev's behalf — the dev inherits both, posts no second claim, and never writes the assignee field.

PM ruling: C first — build the gate. A becomes its own follow-up.

My earlier note on this card leaned "A then C". Reversed, and PR #8472 is why: assertion ② (every @source base resolves) has zero backlog after that PR, so the gate can be built green today and starts holding what PR #8453 and PR #8472 just fixed. Doing A first would change published bundles in four more packages with no instrument watching, which is how this class survived in six places to begin with.

⇒ Scope: the gate only. Do not bring components to source(none) in this PR, and do not fix the four remaining omissions — if the gate reddens on them, that is the correct outcome and it goes in the PR body as the backlog the gate has revealed. ⛔ A gate that ships alongside the fixes it demands cannot be shown to have caught anything.

⚠️ Read the two corrections above before designing it

They are corrections to this card, posted after it was filed:

1. This card's census column is the wrong shape. One column ("does it exclude tests?") would have passed packages/runner while all five of its @source lines were dead. The gate needs both assertions — the exclusions and that every @source base resolves.
2. fields is mis-recorded as adequate. It is leaking through a non-*.test.* source inside its own scanned tree, and it publishes a stylesheet consumers import.

⚠️ Decide deliberately what the gate covers — published packages only, or apps and examples too. An app's bundle is not shipped to consumers, so the argument there is determinism rather than payload: a different justification, possibly a different rule. Say which you chose and why; do not let the population be an accident of the glob.

The thing that makes or breaks this card

Prove the gate can fail, on each assertion separately. Break an exclusion in a scratch copy → red by name; break an @source path → red by name. A gate that has only ever seen a clean tree has not been measured, and this one will be born on a tree that PR #8453 and PR #8472 just made clean.

⚠️ And ask whether an implementation strictly worse than the gate would pass its own tests: a gate that reads zero stylesheets satisfies "no stylesheet is missing an exclusion". Assert the population size as well as the verdict.

⚠️ Technique that PR #8472 needed and you will too: two spellings that look equivalent on today's tree are not equivalent tomorrow. That dev chose between the one-line and two-line exclusion by injecting a marker class and rebuilding, not by preference.

---

Generated by Claude Code