发布态消费者渲染插件组件没有样式路径:插件包不发 style.css,两张已发布表也不含插件的 21 个主题化 utility

发现来源

objectui#4865 的实施(分支 claude/issue-4865-skill-entry-residue)。该卡要求「改写 eval 3 前先从树上回答 plugin-package style.css 的实际问题」。查证结论足够改 eval,但同时量出一个 eval 不该替仓库现场发明答案的缺口,按「相邻缺陷立新单不扩围」单独记录,未在 #4865 的 PR 里修。

事实(实测,worktree 基线 `40d3a3318`)

1. 只有两个包发布样式表

遍历全部 40 个 packages/*/package.json 的 exports,声明 "./style.css": "./dist/index.css" 的只有 @object-ui/components 与 @object-ui/fields。没有任何 @object-ui/plugin-* 有这个导出。

2. 插件包根本不产出 CSS

包	`scripts.build`
`@object-ui/components`	`vite build && node scripts/build-css.mjs`
`@object-ui/fields`	`tsc && vite build && node scripts/build-css.mjs`
`@object-ui/plugin-grid`	`vite build`
`@object-ui/plugin-kanban`	`vite build`

scripts/build-css.mjs 正是把 src/index.css 编成 dist/index.css 的那一步,插件包没有。packages/plugin-grid/src 与 packages/plugin-kanban/src 下一个 .css 文件都没有(find ... -name '*.css' 零命中)。

3. 两张已发布表按构造就不含插件的类

packages/components/src/index.css:14 —— @source '../src/**/*.{ts,tsx}';,只扫自己 src。
packages/fields/src/index.css:43 —— @source './**/*.{ts,tsx}';,同样只扫自己。

所以插件独有的候选类不可能出现在任一张表里。逐字面 grep 量了一遍:21 个主题化 utility 被 packages/plugin-grid/src / packages/plugin-kanban/src 使用,而在 packages/components/src 与 packages/fields/src 里零出现——

text-muted-foreground/80  focus:border-input  focus:ring-destructive/30  border-foreground
divide-border/50  bg-card/60  hover:border-primary/40  text-primary/90  bg-card/20
text-foreground/85  bg-muted/10  bg-muted/15  ring-primary/60  ring-destructive/30
bg-muted/70  border-primary/60  text-muted-foreground/60  bg-foreground/70
bg-muted-foreground/30  shadow-primary/25  ring-primary/40

(另有约 140 个非主题化的插件独有类,同样不在两张表内 —— 那些理论上消费者自己扫 dist 能生成,主题化这 21 个不能,因为 @theme 块在未发布的源码里。)

4. 今天的路只覆盖工作区

工作区:apps/console/src/index.css:19-30 为每个插件包各写一行 @source '../../../packages/plugin-*/src/**/*.{ts,tsx}',并自己声明 @theme —— 成立。
发布态:packages/cli/src/commands/init.ts:537 生成的 scaffold 只有 @import 'tailwindcss' + @import '@object-ui/components/style.css',模板不含任何插件类型,所以没暴露问题。仓内唯一覆盖发布态插件的实现是 packages/cli/src/utils/app-generator.ts:335 的临时 app —— 它逐条重述自己的 @theme(注释原话「Token-for-token the set packages/components/src/index.css declares」)再 @source '../node_modules/@object-ui/*/dist/**/*.js'。那是 objectui run 的临时 app,不是任何指南教给手写消费者的形状。

影响

装了 @object-ui/plugin-grid / plugin-kanban 的发布态 Vite 应用,照 quick-start 导两张 style.css 之后渲染 grid / kanban,上列 21 个主题化 utility 没有任何来源:bg-muted/10、bg-card/60、text-muted-foreground/60 都是常规外观类,不是边角分支。#4858 修正后的三份指南在「不要扫 node_modules」这一句上是对的(对 components / fields 成立,#3884 实测过 dist 是超集),但那条论据的前半句对插件不成立 —— 两张表没有这些类,于是发布态插件消费者被留在没有答案的位置上。

可能的方向(未定,交维护者)

给每个 @object-ui/plugin-* 加 scripts/build-css.mjs + "./style.css" 导出,按 fields 的减法形状(@reference components 的 theme,再减掉 components 已发的规则),消费者多导一张表。
把插件类合并进 @object-ui/components/style.css(components 的 @source 扩到插件 src)—— 但那会让不装插件的消费者也背上体积,且方向上是包依赖倒挂。
只补文档:告诉发布态插件消费者走 app-generator 那条「自己声明 @theme + 扫 node_modules/@object-ui/*/dist」的路。

三者是公开契约层面的取舍(是否新增包导出、体积归属、是否把「自己声明 theme」重新变成推荐做法),不适合由实施方替维护者定。

关联:#4865(本单来源)、#4858 / PR #4866(三份指南的修正)、#4059(style.css 为何必须存在)、#3884(dist 是 node_modules glob 的超集,该结论只对 components 量过)。未认领。

Maintainer ruling (2026-08-17, recorded by PM session session_01NYgmGheCzM6NrHZN436Cxf). The maintainer accepted the decision-inbox report's recommendations in full, verbatim: 「其他接受你的所有建议」. For this card:

Ruled: Direction 1 — every @object-ui/plugin-* ships its own stylesheet.

Each plugin package gains the scripts/build-css.mjs step + a "./style.css": "./dist/index.css" export, built in the subtraction shape @object-ui/fields already uses (@reference the components theme, emit only what components' published sheet does not already carry) — so a published-state consumer imports one extra sheet per installed plugin and the 21 themed utilities (plus the ~140 plain ones) have a source.
⛔ Direction 2 (fold plugin classes into components' sheet) rejected — dependency inversion, and every non-plugin consumer pays the bytes. ⛔ Direction 3 (docs-only: teach self-declared @theme + scanning node_modules dist) rejected — it re-recommends exactly what the 三份 skill 指南教消费者 @source 扫 node_modules/@object-ui/*/src —— 四个包里三个从不发布 src,且三份文件完全不提 style.css #4858 guide correction just retired.
Follow-through: the quick-start/plugin guides gain the one-line import per plugin; the 三份 skill 指南教消费者 @source 扫 node_modules/@object-ui/*/src —— 四个包里三个从不发布 src,且三份文件完全不提 style.css #4858-corrected "don't scan node_modules" prose stays true (it holds once the sheets exist).
Scope today: plugin-grid + plugin-kanban (the two with measured classes); the build step is the pattern any future plugin inherits.

First-touch grading completes: label flipped needs-user-decision → pm:queue in the same stroke.

Generated by Claude Code

Claim: PM session session_011SfZeFWrhGLHmfq61xbz4q (domain:ui execution seat) — branch claude/issue-4929-plugin-published-stylesheets. pm:queue → pm:dispatched in the same label write.

⛔ First, a correction to this seat's own handling of this card

This card sat undispatched for an entire round on my note "has a maintainer ruling, but touches ~20 published packages — needs a size call; the skill routes L/XL to cloud, unavailable this session."

That premise was wrong, and the answer was on this card the whole time. The 2026-08-17 ruling ends with:

Scope today: plugin-grid + plugin-kanban (the two with measured classes); the build step is the pattern any future plugin inherits.

Two packages. I read the card's measurement — a survey of all 40 packages/*/package.json finding no plugin declares the export — and inferred the scope from it, without reading the ruling's scope sentence. ⭐ Third instance this round of the same error: a check was run against the wrong surface. I had even recorded "has a maintainer ruling (Direction 1)" in my own notes and still never read the line that sized it.

No size call was ever needed. Dispatching at M.

Pre-dispatch gate

1. Thread read. The 2026-08-17 maintainer ruling read in full, including the scope line above and both rejections.

2. Shadow check. Nothing in flight touches packages/plugin-grid/ or packages/plugin-kanban/ — #6269 is in packages/types/. #6393 (in the merge queue) is plugin-form. No hold; merge origin/main again before you push.

3. Precedent verified against the tree — the ruling names it and it is complete and copyable:

packages/fields/package.json      build: tsc && vite build && node scripts/build-css.mjs
                                  exports["./style.css"]: ./dist/index.css
packages/fields/scripts/build-css.mjs        exists
packages/components/scripts/build-css.mjs    exists
packages/fields/src/index.css     carries the @reference subtraction shape, with its
                                  rationale documented in-source (objectui#4059)

And the two targets today:

packages/plugin-grid/package.json     build: vite build     style export: (none)
packages/plugin-kanban/package.json   build: vite build     style export: (none)

Scope — Direction 1, as ruled

plugin-grid and plugin-kanban each gain the scripts/build-css.mjs step and a "./style.css": "./dist/index.css" export, built in the subtraction shape @object-ui/fields already uses: @reference the components theme so this compilation gets its tokens, its class-based dark variant and its tailwindcss-animate plugin without emitting a byte of them, and emit only what components' published sheet does not already carry.

⭐ Read packages/fields/src/index.css's header comment before you write anything. It explains why the narrow shape works, and that reasoning is the whole design — copying the mechanism without it is how the subtraction silently becomes a duplication.

Follow-through, also ruled: the quick-start / plugin guides gain the one-line import per plugin. The #4858-corrected "don't scan node_modules" prose stays true and must stay — it holds once the sheets exist. ⛔ Do not re-introduce any advice to scan node_modules or to hand-declare @theme.

⛔ Both rejected directions — do not drift into either

Direction 2 (fold plugin classes into components' sheet) — rejected: dependency inversion, and every non-plugin consumer pays the bytes.
Direction 3 (docs-only: teach self-declared @theme + scanning node_modules dist) — rejected: it re-recommends exactly what the 三份 skill 指南教消费者 @source 扫 node_modules/@object-ui/*/src —— 四个包里三个从不发布 src,且三份文件完全不提 style.css #4858 guide correction just retired.

⚠️ Direction 3 is the tempting one because it is cheap. It is ruled out by name.

Measure, don't inherit

The card's numbers are from 40d3a3318 and are ~4 days old. Re-derive them on the merged tree: the 21 themed utilities used by the two plugins and absent from both published sheets, and the ~140 plain ones. Report your counts next to the card's and say where they differ.

⭐ The 21 are the load-bearing set, and the card says why: the ~140 plain classes a consumer could in principle generate by scanning dist, but the themed ones cannot come from anywhere, because the @theme blocks live in unpublished source. If your re-measure finds the themed count is now zero, that is a premise fork — stop and report.

Tests

Pin that each new sheet actually contains the themed utilities that motivated it, and that it does not re-emit what @object-ui/components' sheet already carries — the subtraction is the point, and a sheet that duplicates the base passes a naive "does it have the class" test.
⛔ Ghost-assertion guard (mandatory): run each new assertion against unmodified origin/main, capture it failing, then passing. Both readings in the PR body.
⛔ Degenerate-control guard: assert a class that components' sheet does carry is absent from the plugin sheet. Without that control, "the sheet has classes" proves nothing about subtraction.
objectui run's temporary app (packages/cli/src/utils/app-generator.ts:335) currently restates its own @theme token-for-token. If the published sheets make that restatement removable, report it — do not remove it here. That is a separate card.

Ground rules

Worktree-first: git worktree add ../objectui-4929 -b claude/issue-4929-plugin-published-stylesheets main && cd ../objectui-4929 && pnpm install. ⛔ Never git stash — refs/stash lives in the common .git dir and is shared across every worktree of this repo; a pop silently restores another agent's work and reports success. ⛔ Never touch content/docs/releases/. Changeset: patch, one entry per package changed. Run the repo's fast checks before pushing — and build both packages to confirm the CSS step actually emits. Open the PR as draft with Fixes #4929 and report back — the PM lands it, ⛔ never self-merge.

Size: M. Model: opus.

Generated by Claude Code

os-dev-report

{
  "issue": 4929,
  "status": "done",
  "branch": "claude/issue-4929-plugin-published-stylesheets",
  "pr": "https://github.com/objectstack-ai/objectui/pull/6403",
  "premise_still_valid": true,
  "summary": "Direction 1 as ruled 2026-08-17: `@object-ui/plugin-grid` and `@object-ui/plugin-kanban` each gain `src/index.css` (the `@reference` narrow shape), a `scripts/build-css.mjs` step and a `\"./style.css\": \"./dist/index.css\"` export, built by subtracting every rule `@object-ui/components`' published sheet already ships — 16.30 kB and 11.41 kB rather than the 172/164 kB the unsubtracted compile weighs. The build step is shared (`scripts/build-plugin-stylesheet.mjs`, thin per-package wrappers) so the next plugin inherits the pattern instead of copying it; `packages/fields`' own copy is deliberately untouched and filed as #6405. Guides, both plugin pages and both READMEs gain the one-line import per plugin, and the #4858 'do not scan node_modules / do not hand-declare @theme' prose is untouched and now true for plugins too. Three CI failures on the first push (entry-guard, turbo-build-inputs, package-files-exist) were reproduced locally, fixed and re-verified in one push.",
  "tests": "MEASUREMENT (re-derived on the merged tree, compiled-diff method rather than the card's literal grep): 128 plugin-only classes absent from BOTH published sheets — 25 themed, 103 plain; grid 83 (7 themed), kanban 47 (19 themed). All 21 themed utilities the card listed at 40d3a3318 still hold, plus 4 the grep method could not see ([&>h3]:text-foreground/80, border-l-primary/40, border-primary/30, hover:text-primary). Themed count is 25, not zero: no premise fork. || BUILD (both sheets deleted first, so emission is observed not assumed): `turbo run build --filter=@object-ui/plugin-grid --filter=@object-ui/plugin-kanban` -> 15/15 tasks, 'plugin-grid:build: OK built dist/index.css (16.30 kB) - 128 rules kept (126 classes), 1356 rules + 75 at-rules already in @object-ui/components sheet', kanban '(11.41 kB) - 52 rules kept (50 classes)'; files back on disk at 16696 and 11684 bytes. || SUITE: new `scripts/__tests__/plugin-published-stylesheet.test.ts`, 14 assertions; it COMPILES rather than reading dist/ because CI runs tests on an unbuilt worktree. Union at pushed HEAD 820631805: `vitest run packages/plugin-grid packages/plugin-kanban scripts/__tests__` -> 'Test Files 188 passed (188) / Tests 3245 passed (3245)'. `pnpm lint` (47 tasks) 0 errors; lint:root, both packages' lint, type-check:scripts, both packages' type-check all exit 0 (tsc --listFiles confirms all four new files are in the scripts program). Gates green at that head: check:entry-guard (--self-test 63 cases + the real run, '48 scripts/ file(s) - no entry guard outside the baseline'), check-control-bytes, check-doc-fence-languages, check-doc-component-types, check-doc-links, check-phantom-dependencies, check-lint-coverage, check-type-check-coverage, check-changeset-presence ('2 source file(s) of 2 released package(s) changed, and this change declares 1 changeset(s)'), check-changeset-no-major, check-changeset-fixed, check-cross-repo-closer-outcome. || GHOST-ASSERTION GUARD, two layers. (a) The test file copied onto unmodified origin/main (faa863dce) in a throwaway worktree: FAIL, exit 1, \"Cannot find module '../../packages/plugin-grid/scripts/build-css.mjs'\" — on main there is no build step to assert about. (b) Because an import-level failure only proves the subject is absent, every assertion group was driven red individually by MUTATING THE FIX on the branch. Baseline first (14 passed, exit 0), then: drop @reference + empty the builder's MUST_SURVIVE -> 4 failed (card coverage, themed emitted, theme-token resolution, control); make the subtraction skip rules whose selector contains 'flex' -> 2 failed, the degenerate control in both packages, \"expected [ 'flex' ] to deeply equal []\"; remove the ./style.css export -> 1 failed (wiring); drop source(none) -> 1 failed (byte-identical-from-any-cwd); inject ':root{--color-ghost-4929}' -> 1 failed (utilities-only). Dropping @reference alone makes the BUILD throw first ('removed 3 utility(ies) that only this build can produce'). Each mutation was proved on disk by grep count + blob hash before the run; each restore by `git checkout HEAD -- <abs path>`, a hash back to the HEAD blob and a clean `git diff HEAD`; the driver carried a trap restoring every touched path. || DEGENERATE CONTROL: flex, text-sm, rounded-md, bg-background, sr-only are asserted PRESENT in the pre-subtraction compile and ABSENT from the published sheet, so the absence cannot be explained by the plugin never using the class. || TWO NEGATIVE RESULTS REPORTED RATHER THAN BURIED. (1) A sixth mutation — swapping the narrow entry for a plain `@import 'tailwindcss'` — left the suite GREEN. Not a ghost: the subtraction removes the duplicated preflight/theme too and the artifact comes out the same size (11.26 kB vs 11.41 kB). The narrow entry stays because it makes a :root theme block impossible to emit at all, and such a block loaded after the base sheet would override a consumer's own token overrides. My in-source comments had claimed the wide entry would ship ~170 kB of duplication; the measurement disproved it and commit 2 corrects the comments. (2) The FIRST run of that mutation matrix was discarded: every leg exited 1 through `--reporter=basic`, which vitest 4 does not have, so five legs read as five confirmations while measuring nothing. The matrix was re-run with a baseline row so a non-zero exit means an assertion. || NOT MEASURED LOCALLY, named rather than glossed: `check:doc-snippets` exits 2 'PRECONDITION NOT MET' (it needs most packages built; this worktree was not) and `check:readme-exports` fails on an unbuilt worktree for ~20 other packages ('type entry not on disk — run pnpm build first') with ZERO findings against the two packages touched here. CI runs both after a build.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #6404: objectui run's temporary app (packages/cli/src/utils/app-generator.ts) still restates the components @theme token-for-token and scans node_modules/@object-ui/*/dist; the new sheets make it only PARTLY removable, because the other @object-ui/plugin-* packages still publish no stylesheet — dropping the restatement today would move the #4929 defect rather than fix it. Reported per the dispatch order, not touched.",
    "filed as #6405: packages/fields/scripts/build-css.mjs now duplicates the logic in the shared scripts/build-plugin-stylesheet.mjs; collapsing it needs byte-for-byte verification of a published artifact, which is its own change.",
    "not filed, recorded in the PR body: plugin-kanban's sheet keeps a 2.7 kB @layer properties block because 74 of its 75 --tw-* declarations are byte-identical to components' and the 75th (--tw-scroll-snap-strictness, from snap-mandatory) is not, so rule-level subtraction cannot drop it. Idempotent; left as-is rather than moving the subtraction to declaration level."
  ]
}

Generated by Claude Code

ACCEPT — enqueued. PM session session_011SfZeFWrhGLHmfq61xbz4q, domain:ui execution seat. 24 ✅ / 3 skipped / 0 ❌ / 0 pending, clean. Reviewed against the tree, not the report.

⭐⭐ Two negative results reported rather than buried — the best thing in this run

1. A mutation that left the suite GREEN, disclosed as such. Swapping the narrow @reference entry for a plain @import 'tailwindcss' did not turn anything red. The dev reported it instead of quietly dropping the leg, explained why it is not a ghost (the subtraction removes the duplicated preflight/theme anyway, so the artifact lands at 11.26 kB vs 11.41 kB), and kept the narrow entry for a different, measured reason: it makes a :root theme block impossible to emit at all, and such a block loaded after the base sheet would override a consumer's own token overrides.

⭐ And it corrected its own source comments in a second commit: they had claimed the wide entry would ship ~170 kB of duplication. Its own measurement disproved its own claim, and it changed the claim rather than the measurement.

2. ⭐ The first mutation matrix was discarded as a blind instrument. Every leg exited 1 through --reporter=basic, which vitest 4 does not have — "so five legs read as five confirmations while measuring nothing." Re-run with a baseline row, so a non-zero exit means an assertion rather than a broken invocation.

Five false confirmations that all pointed the way the dev wanted. That is precisely the trap a ghost-assertion guard exists for, applied one level up — to the harness rather than the assertions — and caught by the dev on its own work.

The three CI failures, verified fixed the right way

gate	how it was fixed	verified
`check:entry-guard`	`isEntrypoint(import.meta.url)` in the per-package wrappers	✅ and ⛔ the SHRINK-ONLY baseline was never touched — `check:check-entry-guard` reports "48 scripts/ file(s) — no entry guard outside the baseline"
`turbo-build-inputs`	added the script to `turbo.json` `build` inputs	✅ — and ⭐ the dev also added `$TURBO_ROOT$/scripts/invoked-as.mjs`, the transitive dependency my message did not mention. A hash that covers the script but not what the script imports is still stale-able.
`package-files-exist`	the pinned exports list grew to four	✅ with the three-things treatment written into the test file, quoting the 2026-08-17 ruling verbatim

⭐ That pin comment does something better than what I asked for: it records which direction is dangerous. Growing is expected as further plugins adopt the step; "shrinking it, by contrast, means a package stopped producing a sheet it still promises: that is objectui#4059's defect returning." A pinned list that explains its own asymmetry will not be misread by the next person who has to move it.

⭐ The guard mechanism is also placed correctly: isEntrypoint sits in the wrappers, which are the entrypoints; the shared scripts/build-plugin-stylesheet.mjs has no guard because it is a library. My own check initially grepped the shared module and found nothing — the design is right, my probe was aimed one file over.

The work itself

Both packages gain src/index.css in the @reference narrow shape (with the same header-comment discipline fields uses), a scripts/build-css.mjs step and the "./style.css" export. 16.30 kB and 11.41 kB — against 172/164 kB for the unsubtracted compile, so the subtraction is doing real work. Build observed with both sheets deleted first, so emission is measured rather than assumed.

Re-measured on the merged tree by a compiled-diff method rather than the card's literal grep: 128 plugin-only classes absent from both published sheets, 25 themed — all 21 the card listed still hold, plus 4 the grep method could not see. No premise fork.

Degenerate control: flex, text-sm, rounded-md, bg-background, sr-only asserted present in the pre-subtraction compile and absent from the published sheet — so absence cannot be explained by the plugin never using the class.

Two unbuilt-worktree gate limitations (check:doc-snippets, check:readme-exports) named rather than glossed, with zero findings against the two touched packages. content/docs/releases/ clean; changeset present.

⚠️ Filed by this run: #6404 (objectui run's temp app is only partly de-duplicable today — the other plugins still publish no sheet, so removing the restatement now would move the defect rather than fix it) and #6405 (fields' build-css now duplicates the shared module). Both ungraded — triage's.

Generated by Claude Code

Navigation Menu

Repository navigation

发布态消费者渲染插件组件没有样式路径:插件包不发 style.css,两张已发布表也不含插件的 21 个主题化 utility #4929

发现来源

事实(实测,worktree 基线 `40d3a3318`)

1. 只有两个包发布样式表

2. 插件包根本不产出 CSS

3. 两张已发布表按构造就不含插件的类

4. 今天的路只覆盖工作区

影响

可能的方向(未定,交维护者)

⛔ First, a correction to this seat's own handling of this card

Pre-dispatch gate

Scope — Direction 1, as ruled

⛔ Both rejected directions — do not drift into either

Measure, don't inherit

Tests

Ground rules

⭐⭐ Two negative results reported rather than buried — the best thing in this run

The three CI failures, verified fixed the right way

The work itself

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

发布态消费者渲染插件组件没有样式路径:插件包不发 style.css,两张已发布表也不含插件的 21 个主题化 utility #4929

Description

发现来源

事实(实测,worktree 基线 40d3a3318)

1. 只有两个包发布样式表

2. 插件包根本不产出 CSS

3. 两张已发布表按构造就不含插件的类

4. 今天的路只覆盖工作区

影响

可能的方向(未定,交维护者)

Activity

os-zhuang commented on Aug 17, 2026

os-support-ai commented on Aug 25, 2026

⛔ First, a correction to this seat's own handling of this card

Pre-dispatch gate

Scope — Direction 1, as ruled

⛔ Both rejected directions — do not drift into either

Measure, don't inherit

Tests

Ground rules

claude commented on Aug 25, 2026

os-support-ai commented on Aug 25, 2026

⭐⭐ Two negative results reported rather than buried — the best thing in this run

The three CI failures, verified fixed the right way

The work itself

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

事实(实测,worktree 基线 `40d3a3318`)