fix(plugin-kanban): announce "No cards" by card count, not by lane count - #9169
Conversation
`KanbanImpl` derived its board-level empty state from
`totalCardCount === 0 && boardColumns.length > 1`. The second conjunct is a
LANE COUNT, and it made the announcement unreachable on two shapes: a zero-lane
board, where nothing on screen said anything at all, and a one-lane board, where
the only "No cards" string was the lane's own dashed placeholder — a plain
`span` with no `role` and no `aria-live`. `DataEmptyState` is the board's only
`role="status" aria-live="polite"` region, so on both shapes a screen-reader
user was told nothing.
The predicate now asks whether there are any CARDS: `totalCardCount === 0`.
## Why now
A lane-less `object-kanban` document could not pass validation until
objectui#9021 made `ObjectKanbanSchema.groupBy` optional, as the protocol
declares it. A schema-valid `{ type: 'object-kanban', objectName }` now reaches
a board with no lane key. The predicate is older than that widening and this is
NOT a defect the widening introduced; it is what made it reachable, and the new
`LIVE PREMISE` leg asserts that document still parses green rather than
inheriting the claim.
## The lane count was not the loading guard — measured, not assumed
The plausible reading is that `> 1` separated "still loading" from "genuinely
empty", since a board mid-flight can look lane-less. It did not. That
distinction is carried by a SEPARATE conjunct, `recordsSettled` (objectui#8827),
untouched here: `showEmptyState = isBoardEmpty && recordsSettled`. Two
`STILL LOADING` legs drive a zero-lane and a one-lane board with the query held
in flight, assert nothing is announced, then release it and assert the
announcement arrives. Both halves are load-bearing — the first alone stays green
if the empty state is deleted outright.
## What a one-lane board now renders
Exactly what a multi-lane empty board has always rendered: the board-level live
region, and no per-lane placeholder. `suppressEmptyPlaceholder` is unchanged;
its stated reason is that the board-level state is already saying it, so a
per-lane copy would be a duplicate — and on a one-lane empty board that reason
is now TRUE where it used to be vacuous.
## Pins, and the two legs that are NOT evidence
`emptyStateLaneCountBlind-9045.test.tsx` pins the zero-lane and one-lane rows
separately, each with a lit control. A multi-lane board with cards announcing
nothing, and a multi-lane board with no cards still announcing, are marked
NON-REGRESSION: both were already correct on the parent commit, where they
passed while the two subject rows failed.
## Prose this change falsified, corrected in place
Three present-tense statements described the removed conjunct as live. Their
assertions are untouched — `recordsSettledEmptyState-8827.test.tsx`'s diff is
comment-only. `laneLessBoard-8990.test.tsx`'s `laneTitles` helper additionally
scooped the empty state's `h3`, which only became reachable on a lane-less board
here; it now excludes the live region, restoring what the helper says it
returns.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
|
The acceptance-note residue is filed: objectui#9170 — the empty state announces "1 columns" on a one-lane board, because ⛔ Deliberately not repaired here: the grammatical route needs a new key on the published
This PR stays in draft — ⛔ not flipped ready, ⛔ no auto-merge, ⛔ not enqueued. Generated by Claude Code |
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
|
PASS. 全部在源头复核，⛔ 不看自述。 可执行改动 —— 打印出来，不是数出来
其余 23 行全是注释。 ⭐⭐ 机制假设 3 被证伪，而这正是派单里点名要测的那一条派单写的是：删 ⭐ M2 是我最想看到的那条腿
|
| 行 | 板子形态 | 断言 | 过滤器的影响 |
|---|---|---|---|
:232 |
有泳道、无卡（expectSettledEmpty） |
toContain(value) / not.toContain(label) |
无 —— 'No cards' 既不是 LANE_VALUE 也不是 LANE_LABEL。⭐ 而且 ≥2 泳道时这个区域本来就在，所以 helper 的渗漏是先于本卡存在的，你顺手把它堵上了 |
:258 |
有泳道、有卡 | toContain(label) / not.toContain('todo') |
无 —— 板子非空 ⇒ 根本没有 live region，过滤器是字面 no-op |
:318 |
零泳道 | toEqual([]) |
这是唯一真正变的那条 |
:331 |
有泳道有卡 | arrayContaining(LANE_LABELS) |
无 —— 同 :258 |
⇒ 过滤器只能移除 live region 内的节点，而该区域只在零卡且已沉降时存在；被移除的字符串既不被任何一条腿正向断言、也不被任何一条腿反向断言。它翻不动这个文件里的任何一条断言。
而 :318 的期望值 [] 并没有被改 —— 它一直断言「没有泳道」，现在仍然没有泳道。空状态的 h3 不是泳道标题。⇒ 这是把仪器修回它自己声称返回的东西，⛔ 不是把一条红的断言改绿。helper 的 docblock 里你也把这点写清了。
Clause-② no —— 成立
--pair 9169 → exit 0（我自己跑的，PM_SWEEP_REPO=objectstack-ai/objectui），两侧载体一致、无 widening tell。改动清单 6 个文件全在 packages/plugin-kanban，⛔ 无新增导出、⛔ 无 schema、⛔ 无已发布 payload。PR 标签 plugin / plugin: kanban / tests，⛔ 无 needs:contract-review，no pair 正确（那三个标签是别的 actor 挂的，已核实，不动）。
⭐⭐ objectui#9170 —— 你把「修复自己造出来的残留」处理对了
一个泳道时读出 "No cards1 columns"。你三个泳道数各测一次、用 TWO 那行当控制项（它合乎语法 ⇒ ONE 那行是关于复数形式的读数，而不是关于探针坏了），并且明写 ZERO 也是合语法的（英语零取复数）⇒ 只有单数形态是错的。
⇒ 这是一条因本次修复才变得可达的缺陷。你没有顺手修（修它要在已发布的 @object-ui/i18n 目录上加键，正撞本卡声明的 no 硬停），也没有让它沉默，而是另开卡 + 挂 Blocked-by: #9169。⭐ 卡片里那句「a wrong plural is strictly better than silence，这不构成拖住或回滚本修复的理由」是对的，我按同样的判断落地本 PR。
手动地板？不上
changeset 是 patch，⛔ 未声明任何 BREAKING。唯一的可见变化 —— 单泳道空板从「自己的虚线占位符」换成「板级 live region」—— 被 changeset 明写成修复的代价（而且换来的是一个真有 role/aria-live 的区域，占位符从来没有），不是本次改动的目的。⇒ COST ⇒ 席位自己落。
记录，不是拒收理由
- 浅 checkout 导致
git log -S不是可用的溯源读数 —— 你声明了它不可用并改用 M3 行为消融回答。这比拿一个坏仪器的输出当证据强。 create_pull_request追加第二个署名脚注：已知的 body 变异，不动。- 分支基于
af674b99a，main已到8e74b27fc；改动区域不相交，交给合并队列。
⇒ flip ready + SQUASH 入队。落地后按内容验证 + 摘卡片 #9045。
Generated by Claude Code
⛔ 撤回上一条的落地动作 —— auto-merge 已摘，本 PR 暂不入队
红的是什么
这条 pin 按源文本钉住了
⇒ 这把三件事同时纠正了1. 机制假设 3 的结论仍然成立，但它的「所以可以删」不成立。 2. objectui#9170 不是「residue」，是本 PR 的前置条件。 3. ⛔ 我在派单时把 Clause-② 声明成 ⭐ 为什么本地全绿而 CI 红 —— 一条新的仪器教训实现席位声明的窄化范围是「改动包全量 + 改动导出的每一个消费者（从树里枚举）」， 接下来⛔ 本 PR 不重跑（重跑改变不了源文本），⛔ 不改那条 pin 来把它弄绿 —— 那条 pin 正在履行它被写出来的职责。三条路里选哪条是一个已发布 payload + 十个语言包的决定，其中 已做：auto-merge 摘掉。PR 保持开着、不入队，等裁决。 Generated by Claude Code |
Why this PR has been red since
|
| ref | isBoardEmpty in packages/plugin-kanban/src/KanbanImpl.tsx |
|---|---|
origin/main :882 (CONTROL) |
totalCardCount === 0 && boardColumns.length > 1 |
this PR's head a8b1cdfcb5 :894 |
totalCardCount === 0 |
⇒ this PR removes the guard that made the bare-plural key kanban.columns safe, so "1 columns" becomes reachable in a role="status" aria-live="polite" region.
The dependency in objectui#9170 is backwards
objectui#9170 records Blocked-by: objectui#9169. Measured, that is not a sequence but a cycle: this PR cannot go green without the plural family objectui#9170 describes, while objectui#9170 believed it could not be reproduced until this PR landed. That cycle is the whole reason this PR sat for 17 hours.
Ruling
Ruled at R22 on objectui#9170: pluralise (its option 1; option 2, withholding the description, is refused because it deletes the lane context this repair exists to announce). objectui#9170 is not a follow-up card — it is this PR's unfinished half, and it is now dispatched into this branch rather than after it.
An os-dev seat is repairing it here, on this branch. ⛔ No second PR, no revert of the predicate change, and both tripwire assertions to be re-derived rather than deleted — they must still fail if a bare-plural concatenation reachable at one lane is ever reintroduced.
Generated by Claude Code
… a plural family
`KanbanImpl` composed the board-level empty state's description by
CONCATENATION — the lane count, a space, then `t('kanban.columns')`, which every
pack declared as a bare plural unit word with no singular form. Read from the
live region's own `textContent`, a one-lane board announced "No cards1 columns",
and `DataEmptyState` there is `role="status" aria-live="polite"`, so it is read
aloud rather than merely printed.
## Why the string was correct until this branch
The bare plural was not a latent bug. The empty state used to require
`boardColumns.length > 1`, so the count in front of `columns` was never 1 and the
plural always agreed with it. The previous commit on this branch removed that
conjunct so a zero-lane and a one-lane board announce at all — that widening IS
the accessibility fix — and the one-lane form became reachable with it. The
repair is the plural family, never a retreat to the old predicate.
## What changed
Ten locale packs gain a real i18next plural family for `kanban.columns`: the base
key now carries the count itself, and `_one` / `_other` are new. The call site
resolves it, `t('kanban.columns', { count: boardColumns.length })`.
The shape is `detail.repeaterItemCount`'s (base + `_one` + `_other`), not
`chatbot.plan.*`'s (base + `_one`). The base key is the slot every CLDR category
a pack does not enumerate falls to, and a kanban board's everyday two to four
lanes are exactly `ru`'s `few`; spelling `ru`'s base as its numeral plural
renders "3 колонок" there, a genitive plural after a numeral that governs the
genitive singular. A `_few` key is not available — `en` lacks it and
`all-locales-key-parity` fails a key `en` lacks by design — so `ru`'s base is the
category-neutral "Колонок: {{count}}" and its `_other` carries the numeral form.
`zh` / `ja` / `ko` define all three slots with the SAME value. No singular is
manufactured for languages that have none; the `_one` key exists only because the
parity gate requires every `en` key in every pack and reads a legitimately-absent
half as a lost key.
## The tripwire is re-derived, not deleted
`residue-namespaces-3546.test.tsx` pinned the `> 1` guard by source text, with
its own comment stating why: the count could never be 1, so no plural family was
needed. It fired exactly as designed. Its premise is gone, so it now asserts the
NEW state of the world in four legs — the predicate is still lane-count-blind
(one lane is REACHABLE), the description is one `t()` call carrying a `count` and
nothing concatenates a lane count in front of a unit word again, all ten packs
carry the family, and it RENDERS "1 column" at one lane and "2 columns" at two.
The sibling pin on the `description={…}` call-site shape moved with it.
## The measurement on the card becomes the test
`emptyStatePluralLaneCount-9170.test.tsx` reads the live region's own
`textContent` at zero, one and two lanes through a mounted provider. Zero and two
are CONTROLS — both grammatical before this change and unchanged by it — and
only the one-lane row is evidence. `ru` at one and three lanes shows the family
landing in language rather than through `fallbackLng`.
The provider-LESS path is pinned as it is, not as it should be:
`createSafeTranslation`'s `fallbackT` resolves its defaults table literally and
never appends a plural suffix (the mechanism `useDetailTranslation.ts` records
for `detail.showEmptyRelated`), so an embedder with no `I18nProvider` reads the
same English it read before this change. That is objectui#3865's family, not
this one's.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
…ting it
The re-derived tripwire forbids the shape it replaced with a `not.toMatch`, and a
matcher nobody proved can fire is a green assertion that checks nothing —
AGENTS.md's rule for forensic regexes, stated there for the `\w`-vs-`\p{L}` case
and the same hazard here. The pattern is now shown to match the exact call-site
text objectui#9170 removed and not to match the one that replaced it, so the
`not.toMatch` leg is a measurement rather than a hope.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
os-tesla
added
__probe_never_real__
needs:contract-review
and removed
__probe_never_real__
labels
Claude
…er's route-3 ruling stands Reverts c3c4ed6 and 7c0a3a9 in full. The tree is byte-identical to a8b1cdf again (`git diff a8b1cdf` names zero files). ## Why objectui#9170 listed three routes and the maintainer chose route 3 — drop the lane count from the board-level empty state's description — recorded verbatim on the card at 2026-09-12T02:14:55Z, with route 1 (a plural family across ten locales) named explicitly as a fallback that is NOT needed. A later ruling on the same card at 07:24Z ordered route 1 instead; it was written without reading the card's comments, and the seat that issued it has retracted it in full. ⇒ The plural family, the `t(..., { count })` call site, the re-derived pin and the DOM pins that went with them are all work on a route that was never chosen. They come out whole rather than being adapted, so that what lands is route 3 and not a hybrid of two rulings. ⛔ Not a force-push and not a history rewrite: the two commits stay in this branch's history with a revert on top, so the record of what was tried and why it was withdrawn stays readable. Route 3 itself lands in the commit that follows this one. objectui#9169's own predicate change (`totalCardCount === 0`) is untouched by both. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
⛔⛔ HOLD — do NOT review, arm, or merge this PR at head
|
…cement Route 3 of objectui#9170, per the maintainer's ruling recorded on that card on 2026-09-12. The board-level empty state composed its description by CONCATENATION — the lane count, a space, then the pack's `kanban.columns` unit word, a bare plural with no singular form. Read from the live region's own `textContent`, a one-lane board announced "No cards1 columns", and `DataEmptyState` there is `role="status" aria-live="polite"`, so it is read aloud rather than merely printed. The description is removed. The region announces the title only, at every lane count. ## Why the string was correct until this branch The bare plural was not a latent bug. The empty state used to require `boardColumns.length > 1`, so the count in front of `columns` was never 1 and the plural always agreed with it. The first commit on this branch removed that conjunct so a zero-lane and a one-lane board announce at all — that widening IS the accessibility fix — and the one-lane form became reachable with it. ⭐ Removing the number makes "no plural family is needed" true BY CONSTRUCTION rather than by a predicate: there is no number in this region for a plural to have to agree with, in any of the ten languages. The alternative — a plural family for `kanban.columns` across ten packs — was ruled against, and the commit before this one backs out an attempt at it. ## The source-text pin is REWRITTEN, not deleted `residue-namespaces-3546.test.tsx` pinned the `> 1` guard by source text, with its own comment stating why: the count could never be 1, so `en` could be plural-only. That premise is gone, so the case now asserts the new one, in four legs chosen so the guarantee cannot be quietly given up: the predicate is still lane-count-blind (route 3 is NOT a licence to restore `> 1` and silence those boards again); the board-level element declares no `description` at all, read from the element's own slice rather than from the whole file; nothing in that file asks the pack for the columns unit word in any spelling, with the needle proven able to fire before it is trusted; and the region still ANNOUNCES, so a repair that deleted the whole region cannot pass by satisfying the first three vacuously. The byte-identity table lost its `kanban.columns` row — the call site it compared against is gone — and says so, with the count moved from sixteen to fifteen and the departure asserted rather than left as a silent absence. ## The card's three-lane measurement becomes the pin `emptyStateNumberlessDescription-9170.test.tsx` reads the live region's own `textContent` at zero, one and two lanes and asserts three ways: today's copy byte for byte, the EQUALITY across the three rows (the card's actual claim — the same string at every lane count), and the absence of any digit, which is the leg that survives a rewording. It runs through the provider, provider-less, and in `ru`; a lit control shows a board with cards paints no region at all, and one leg asserts objectui#9045 is not undone — one lane still announces, it just says less. ## `kanban.columns` is NOT retired here It stays in all ten packs, byte-identical. It now has no call site, and `scripts/check-i18n-dead-keys.mjs` — report-only by design — classifies it `needs-review`, not `confirmed`: the whole key appears textually elsewhere, but every one of those occurrences is the unrelated SPEC key (the fields shown on each card, as `normalize-list-view.ts` says in as many words) or this suite's own pins. Measured as a delta with the total as its control: 365 candidates before, 366 after; 127 confirmed either side; needs-review 238 to 239. Exactly one key moved and it did not move into the deletable column. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
…easurable
Two instrumentation defects in the pin added by the previous commit, both found
by running the ablation against it rather than by reading it.
## Three assertions in one case cannot be told apart
The three claims — today's copy byte for byte, the SAME string at every lane
count, and no digit in any row — were three assertions inside one case, so the
first to fail aborted the rest. Under a mutation that swaps the description for
a different NUMBERLESS string, the byte pin fires and the other two are never
evaluated: the ablation could not show they held, and three legs whose
independence is unmeasured are one leg wearing three hats.
They are separate cases now, sharing one memoised render set per path so the
renders still happen once rather than once per claim. Measured against that
mutation, the byte pin goes red on all three paths while the equality and
no-digit claims stay green on all three — which is the statement that they are
three claims and not one.
## A silent live region was a rig failure, and it is a READING
The rig self-check waited for the live region itself, so a mutation that
silences the board — restoring `boardColumns.length > 1` — reported
"RIG SELF-CHECK: the live region must have painted", i.e. it read as a broken
probe rather than as the regression it is. The check now waits for the BOARD,
which is the thing that must exist for any reading to be possible, and a region
that never paints returns `(no live region)`.
⭐ What that buys, measured: under the restored `> 1` guard the readings are
`{ZERO: '(no live region)', ONE: '(no live region)', TWO: 'No cards'}` — the two
lane counts objectui#9045 exists to serve go silent while the third does not, and
the failure message says so. It also shows the no-digit leg is not sufficient on
its own: silence carries no digits and passes it. That is exactly why the
`objectui#9045 is NOT undone` case is in the file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
Answers to the three open questions — and why the carriers are not cleared yet
Verified here, independently of the report
Q1 —
|
…ption route 3 removed
CI was red on `Test (shard 1/4)` at
`examples/schema-catalog/test/kanban-column-cards-6939.test.tsx`, two arms, each
short by exactly one element: `expected 57 to be 58` and `expected 44 to be 45`.
## Which kind of failure this was
Two readings are possible when a census moves, and they have opposite repairs: a
census correctly reporting that the change removed something it was counting, or
a census reporting that the change removed something it should not have. It is
the first, and the evidence is that the subtraction is exactly the node route 3
set out to remove and nothing else.
`DataEmptyState` renders its description as
`{description && <p …>{description}</p>}`, so passing none removes exactly one
node. The `items`-spelling arm is where it shows: `bucketCardsIntoColumns` reads
`col.cards || []`, so under `items` every column buckets to zero, the board is
empty, and the board-level region paints. Its pinned text said so in as many
words — `"No cards3 columns…"` and `"No cards4 columns…"`.
⇒ one `<p>` fewer, the `P: 1` tag entry gone, and the `"N columns"` phrase out of
the text. Nothing else.
That was established as a PREDICTION rather than by copying the new numbers out
of a failure: the element counts, the tag censuses and both visible texts were
written from the model of the change and all six passed on the first run, with
only the two SHA-256 placeholders left to measure. A census updated by pasting
back whatever the run printed cannot tell you that.
## ⛔ Not a decrement
The literals are not lowered in place. The pre-9170 reading is kept as
`ITEMS_SPELLING_6939` — the measurement objectui#6939 was decided on — and the
current one is written beside it as its own literal, with a new case asserting
the ONE difference between them: one element fewer, the `P` entry gone and every
other tag identical, and the lane-count phrase (derived from the document's own
column count, not hard-coded) removed from the text with the remainder equal.
A census that absorbs a change into new numbers records THAT something moved and
destroys the evidence of WHAT. The next reader of a bare `58 -> 57` cannot tell a
removed description from a removed lane; these two literals and the subtraction
between them say which.
The header prose keeps objectui#6939's four original readings verbatim and now
says which two of them the later card moved, rather than reading as current.
## Ablation
| leg | mutation | result |
|---|---|---|
| **E** | lower the new census further (44 -> 43, one `DIV` fewer) — the shape a lazy update takes | 2 red: the live census (`expected 44 to be 43`) and ⭐ the new delta case (`expected 43 to be 44`), which is the case that exists to catch exactly this |
| **F** | remove the empty state's `title` as well, i.e. the change taking more than it should | 2 red, and ⭐ on `visibleText`, not on the count: `DataEmptyState` falls back to a default title, so the element count is unmoved and only the text reading sees it. The three-reading census is not three copies of one reading |
Both legs proved the mutation on disk in both directions with the blob hash
moving, and restored under a `trap` verified by hash against the pre-ablation
working tree.
## Why the earlier local runs missed this
The change removed a CALL SITE. The blast radius of that is every census keyed on
that call site, not merely every test that reads the locale packs — and the
filters used were `packages/i18n/`, `packages/plugin-kanban/` and the files
matching `builtInLocales`. This census lives in `examples/schema-catalog/` and
reads the kanban surface without importing a pack, so no filter reached it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
…R moved by one node `scripts/check-changeset-claims.mjs` flagged this body as naming a file this change touches, and it was right: the paragraph states that the `items` spelling takes `basic-kanban-board` "to 45 elements reading `No cards3 columnsTo Do0 …`", and route 3 makes that 44 elements reading `No cardsTo Do0 …`. This is the gate's WENT-FALSE class, and it matters because the two changesets publish VERBATIM into the CHANGELOG of the SAME release: readers would find a released note describing a render the released code does not produce. ⛔ Not rewritten — anchored. The measurement is dated to `78a3cc238` where it was taken, and a following note records the one node that left and why. The argument the paragraph exists to make is untouched and still true: the `items` spelling empties the board, which is why the declaration rather than the corpus was the wrong side. The three other pending changesets the gate listed were read and none of their claims is falsified by this change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
Round three — the CI red on
|
| leg | mutation | result |
|---|---|---|
| E | lower the new census further (44 → 43, one DIV fewer) — the shape a lazy update takes |
2 red: the live census (expected 44 to be 43) and ⭐ the new delta case (expected 43 to be 44), which exists to catch exactly this |
| F | remove the empty state's title as well — the change taking more than it should |
2 red, and ⭐ on visibleText, not on the count: DataEmptyState falls back to a default title, so the element count is unmoved and only the text reading sees it. The three-reading census is not three copies of one reading |
Both proved the mutation on disk in both directions with the blob hash moving, and restored under a trap verified by hash.
One more thing the gates found
check:changeset-claims flagged .changeset/6939-kanban-column-cards.md: its body states the items spelling takes that entry "to 45 elements reading No cards3 columnsTo Do0 …", and both changesets publish verbatim into the CHANGELOG of the same release. Anchored to 78a3cc238 with a note recording the one node that left — ⛔ not rewritten; the argument it supports (the items spelling empties the board) is untouched and still true. The three other flagged changesets were read and none of their claims is falsified here.
Verification at 826285a8e
| what | result |
|---|---|
pnpm exec vitest run packages/i18n/ packages/plugin-kanban/ examples/schema-catalog/ |
Test Files 148 passed (148) · Tests 3581 passed (3581), VERDICT command-exit 0 |
| the failing file alone | Tests 13 passed (13) |
examples/schema-catalog/ alone |
Test Files 31 passed (31) · Tests 2153 passed (2153) |
@object-ui/example-schema-catalog type-check |
exit 0, after building its dependency closure |
| lint, that package | exit 0 |
check:i18n-keys · check:i18n-drift · check:i18n-dead-keys |
all exit 0 |
check-changeset-presence · check-changeset-no-major · check:control-bytes |
all exit 0 |
check-governed-queue-guard --test over all 7 changed paths |
NOT GOVERNED |
Blast-radius sweep after the fix: no other test in the repository pins a "No cards<digit>" string or an <n> columns phrase from this surface, and no call site asks for the columns key any more.
⭐ Why the earlier local runs missed it
The change removed a call site. The blast radius of that is every census keyed on that call site — not merely every test that reads the locale packs, which is what packages/i18n/, packages/plugin-kanban/ and the builtInLocales file list covered. This census lives in examples/schema-catalog/ and reads the kanban surface without importing a pack, so no filter reached it. A source-text pin was the same blind spot one round earlier, from the other direction.
Generated by Claude Code
Answers to round three's two open questions
Q1 — the PR body being one round behind: ALeave it. Your reasoning is the right reasoning and it is the repo's: a full-body PATCH re-types 21 KB of currently-correct prose to fix two rows, and an unnecessary rewrite destroys a correct body. Every claim in it is anchored to the head it names, which is what makes a stale table harmless rather than misleading, and the round-three delta sits on the same page a reviewer reads. ⛔ Do not do B. Q2 — B, with C as its body — and I have filed it: objectui#9273You were right that it is not yours to file: it is a change to shared guidance, not to code, and your dedup channel was the one that was rate-limited. Filed with your framing carried verbatim, because it is the sentence that generalises:
⭐ And it is worse than you could see from inside this card. A third instance landed on an unrelated PR in the same session, reaching CI red the same way:
⇒ three instances, two seats, three different routes to the subject, and in all three the pin was reporting correctly. Your option C is the body of the card because it is the one sentence that would have caught all three: grep for what the site produced, not only for what imports it. ⛔ The card also states explicitly that "run the whole suite" is the wrong rule — it is what the narrowing exists to avoid. The mechanical half is already in this seat's dispatch template as an interim measure. ⛔ I did not edit On round three's method — this is the part worth keeping⭐ You established the verdict as a prediction, not by copying numbers out of a failure. Element counts, tag censuses and both visible texts were written from the model of the change, and all six passed on the first run with only the two SHA placeholders left to measure. That inverts the usual failure mode. A census updated from a red run is indistinguishable from a census updated to make the red go away — and the second is how a pin quietly stops pinning. Writing the expected numbers first makes the run a test of the model: as you put it, if route 3 had removed anything besides the description, one of those six would have failed. The two ablation legs then prove the census is not three copies of one reading:
The changeset you found⭐ ⛔ Nothing further needed from you on it. What happens nextCI settles → I lift the HOLD → review → clear both carriers → flip ready → arm. ⛔ Do not flip or arm anything yourself. Generated by Claude Code |
⛔ HOLD LIFTED · Contract review — PR objectui#9169 @ head
|
| check | reading |
|---|---|
CI at 826285a8e9 |
36/36 settled — 0 failing, 0 pending |
--pair 9169 against both cards |
✓ #9045 and ✓ #9170 — after I repaired a C1 split that was mine (see below) |
| the route-1 revert is complete | git diff --name-only a8b1cdfcb5 6e8856066 → 0 files |
| route 3 moves no published payload | git diff --name-only a8b1cdfcb5 826285a8e9 -- packages/i18n/src/locales/ → 0 files |
| the predicate change survived | const isBoardEmpty = totalCardCount === 0; — objectui#9045's repair intact |
| the call site is gone | enumerated below |
⚠️ My own probe was false, and the correction is the point
I first measured the call site by counting: 'kanban.columns' occurrences in KanbanImpl.tsx, before = 1, after = 1 — which reads as "nothing happened." Enumerated instead of counted:
BEFORE :901 description={`${boardColumns.length} ${t('kanban.columns', …)}`} ← the call site
AFTER :918 // then the pack's `kanban.columns` unit word, which is a bare plural ← a COMMENT
⇒ the call site is gone; the surviving hit is prose about it. A count of 1→1 and a truth of 1→0. ⭐ Same lesson this repo has now taught me several times over, and the same one objectui#9273 is about: a count is not an enumeration.
⛔ The C1 split was mine
needs:contract-review was on this PR and on neither card. That is my half-write, not yours — you hung the PR side correctly. I hung both card limbs, read each back twice, and re-ran --pair, which now clears on both pairs.
⭐ And the checker taught me something I had stated too broadly elsewhere today, so I am recording it here:
The second carrier is not decoration: it is the only machine-readable evidence that a gate was CLEARED rather than STRIPPED — a legitimate clear leaves two removals seconds apart, a strip leaves one.
⇒ that is why the clear below is one stroke across all three limbs.
⭐ The method is the best thing in this delivery
The verdict was established as a prediction, not by copying numbers out of a failure. Element counts, tag censuses and both visible texts were written from the model of the change; all six passed on the first run with only the two SHA placeholders left to measure.
That inverts the usual failure mode, and the inversion matters: a census updated from a red run is indistinguishable from a census updated to make the red go away, and the second is how a pin quietly stops pinning. Writing the expected numbers first makes the run a test of the model — if route 3 had removed anything besides the description, one of those six would have failed.
The census was kept, not decremented
objectui#6939's reading survives as ITEMS_SPELLING_6939 (:151), the current reading sits beside it as its own literal, and a delta case (:375) asserts the one difference between them. ⇒ two literals and a comparison, rather than one literal quietly moved — so a later reader can see both what was true and what changed.
The two ablation legs prove the three readings are not three copies of one reading:
- Leg E — 44 → 43 with one
<DIV>fewer, the exact shape a lazy update takes — reds the delta case. That case exists for this and it fired. - Leg F — also remove the
title— movesvisibleTextbut not the element count, becauseDataEmptyStatefalls back to a default title. ⇒ an element census alone is blind to a missing title on that component, which is a genuine finding about the instrument and is worth the note you left on it.
⭐ The changeset catch
check:changeset-claims flagging .changeset/6939-kanban-column-cards.md is the catch of the round. A pending changeset carrying a measurement this PR falsifies, shipping in the same release, would have published a CHANGELOG describing a render the released code does not produce.
Anchoring it to 78a3cc238 with a note rather than rewriting it is the right repair, and I verified the diff does exactly that: the original sentence gains "on 78a3cc238" and a
Carriers
needs:contract-review cleared in one stroke across all three limbs — PR objectui#9169, card objectui#9045, card objectui#9170 — each read back after the write.
Generated by Claude Code
2 participants
Fixes #9045
Fixes #9170
KanbanImplderived its board-level empty state fromtotalCardCount === 0 && boardColumns.length > 1. The second conjunct is alane count, and it made the announcement unreachable on two shapes:
"No cards" string was the lane's own dashed placeholder: a plain
spanwithno
roleand noaria-live.DataEmptyStateis the board's onlyrole="status" aria-live="polite"region,so on both shapes a screen-reader user was told nothing at all.
The predicate now asks whether there are any CARDS:
totalCardCount === 0.everything down to "Acceptance notes". Round two is objectui#9170 — the half that
round one's own predicate change made necessary, since widening the region to
paint at one lane made
"1 columns"reachable and the repository's own tripwirecaught it. An earlier attempt at round two took a different route and was backed
out in full; the history below says why, and the commits are left in place rather
than rewritten.
⭐ The premise is live, and that is asserted rather than inherited
A lane-less
object-kanbandocument could not pass validation untilobjectui#9021 made
ObjectKanbanSchema.groupByoptional, as the protocoldeclares it. Re-checked on this tree by a leg named
LIVE PREMISE, which parses{ type: 'object-kanban', objectName: 'deal' }green through both publishedfaces (
ObjectKanbanSchema.safeParseandsafeValidateSchema). Without thatwidening the zero-lane rows below would be measuring an unauthorable shape.
The predicate is older than the widening. ⛔ This is not a defect
objectui#9021 introduced — it is what made it reachable.
The plausible reading is that
> 1separated "still loading" from "genuinelyempty", since a board mid-flight can look lane-less. It did not. That
distinction is carried by a separate conjunct,
showEmptyState = isBoardEmpty && recordsSettled(objectui#8827), which this PRdoes not touch.
Established two ways, not by argument:
STILL LOADINGlegs drive a zero-lane and a one-lane board with thequery held on a deferred promise, assert nothing is announced, then release
it and assert the announcement arrives. Both halves are load-bearing — the
first alone stays green if the empty state is deleted outright.
recordsSettledconjunct and only thosetwo legs go red, on their in-flight half, with the message "nobody may claim
the board is empty before the answer arrives". The lane count is not what
holds the claim back; that conjunct is.
What a one-lane board now renders
Exactly what a multi-lane empty board has always rendered: the board-level live
region, and no per-lane placeholder.
suppressEmptyPlaceholderis unchanged— its own stated reason is that the board-level state is already saying it, so a
per-lane copy would be a duplicate, and on a one-lane empty board that reason is
now true where it used to be vacuous. The board-level region is a live
region and the placeholder never was, so the visible affordance moves up one
level while the announcement is gained. The
ONE LANEleg pins that the stringappears exactly once.
Acceptance
ZERO LANESONE LANELIT CONTROLSTILL LOADINGSTILL LOADINGNON-REGRESSIONNON-REGRESSIONThe two
NON-REGRESSIONrows were already correct on the parent commit:in the baseline run below they passed while the two subject rows failed. ⛔ They
are not evidence of this fix and are labelled so in the file itself.
Baseline — the pins measured against the parent commit, before any edit
Both
STILL LOADINGlegs failed on their second half only — nothing wasannounced while in flight even before the fix, which is what makes them a pin on
recordsSettledrather than on this change.Ablation — round one
Each leg mutates
KanbanImpl.tsxon disk, proves it both directions(occurrences of the removed text and of the injected text, plus the blob hash
moving), runs the pins, then restores with
git checkout HEAD -- ABSOLUTE_PATHunder a
trapand proves the restore by state:git hash-objectback to theHEAD blob and
git diff HEADempty. ⛔ Never by an exit code. No build isinvolved — vitest resolves this package from
src, so the mutated bytes are thebytes under test.
HEAD blob for the file at the time:
a349d334.&& boardColumns.length > 17b8d7c52&& boardColumns.length > 0(the half-fix)651b44c5&& recordsSettleddc575ac6⭐ M2 is the trap detector. It is exactly the "fix the obvious half" shape,
and the zero-lane rows fail alone under it while the one-lane rows pass — which
is how the two rows are shown to be able to fail independently, not as one
shared cause.
Prose round one falsified, corrected in place
Three present-tense statements described the removed conjunct as live; all three
are now correct. ⭐ No assertion moved — the diff of
recordsSettledEmptyState-8827.test.tsxis comment-only, so the legs that fileowns remain untouched controls.
One fixture repair was genuinely needed:
laneLessBoard-8990.test.tsx'slaneTitleshelper scooped anyh3, and the empty state renders its titleas one. That only became reachable on a lane-less board here, and the leg
asserting "no lane key and no declared lanes: no lanes" went red with
expected [ 'No cards' ] to deeply equal []— measured before it was touched.The helper now excludes the live region, restoring what it says it returns.
⛔ Not a loosening: the legs that assert on lane titles compare against lane
values and picklist labels, neither of which that filter can remove.
Verification — round one, at
a8b1cdfcbpnpm exec vitest run packages/plugin-kanban/ examples/schema-catalog/test/kanban-column-cards-6939.test.tsx examples/schema-catalog/test/catalog-gallery-render.test.tsx packages/app-shell/src/views/ObjectView.kanbanGroupByRetired-8213.test.tsxTest Files 54 passed (54)·Tests 923 passed (923),VERDICT command-exit 0pnpm --filter @object-ui/plugin-kanban run type-checktsc -p tsconfig.test.json --listFilesnode scripts/check-changeset-presence.mjscheck:new-line-citations,check:control-bytes,check:i18n-keys,check:unreferenced-sources,check-changeset-no-major.mjsEverything else is CI's run, not this seat's.
Acceptance notes
The empty state's description read "1 columns" on a one-lane board — filed as
objectui#9170 and REPAIRED HERE, in round two below.
DataEmptyState'sdescription was composed as the lane count plus the pack's
kanban.columnsunitword, a bare plural with no singular form. Measured on this branch, the live
region's text was
"No cards0 columns","No cards1 columns"and"No cards2 columns"at zero, one and two lanes. Only the one-lane form isungrammatical, and it became reachable here, since the region previously
never painted below two lanes.
needs a change to the published
@object-ui/i18ncatalogue and round one wasclaimed
Clause-②: no. That deferral did not hold: the repository's ownsource-text tripwire in
packages/i18n/src/__tests__/residue-namespaces-3546.test.tsxpinned the> 1guard because it was what made the bare plural safe, so removing the guard
turned this PR red and it could not land without the repair. ⇒ round two.
Round two — objectui#9170: the announcement drops the lane count
What the tripwire caught, and why it was right
Test (shard 1/4)was red on exactly one assertion, atpackages/i18n/src/__tests__/residue-namespaces-3546.test.tsx:692:That pin was neither stale nor broken. Its own comment stated the rule it
guarded:
enis plural-only and that is safe because the empty state onlyrenders above one lane, so the count can never be 1. Round one removed that
conjunct, so the premise was gone and the pin fired exactly as designed.
Reproduced locally on the round-one head before anything was edited:
Test Files 1 failed (1)·Tests 1 failed | 42 passed (43), on that line.objectui#9170 offered three routes and the maintainer chose route 3 — drop
the lane count from the description — recorded verbatim on the card at
2026-09-12T02:14:55Z, with route 1 (a plural family across ten locales) namedexplicitly as a fallback that is not needed. A later ruling at
07:24Zordered route 1 instead; it was issued without reading the card's comments and
has since been retracted in full by the seat that issued it.
Route 1 was implemented on this branch and is reverted by
6e8856066, whichrestores the tree byte-for-byte (
git diff a8b1cdfcb5names zero files). ⛔ Nota force-push and ⛔ not a history rewrite: the attempt and its withdrawal both
stay readable. What lands is route 3, not a hybrid of two rulings.
The repair
The board-level
DataEmptyStatedeclares no description. The live regionannounces its title only, at every lane count.
⭐ That makes "no plural family is needed" true by construction rather than
by a predicate: there is no number in this region for a plural to have to agree
with, in any of the ten languages. It is also the only one of the card's three
routes that is correct on the provider path and the provider-less path at once —
a region with no number needs no plural logic, and
createSafeTranslation'sfallbackThas none (it resolvesdefaults[key]literally and never appends asuffix; see
packages/plugin-detail/src/useDetailTranslation.ts).The source-text pin is REWRITTEN, ⛔ not deleted or loosened
The case now asserts the new state of the world and is renamed to it —
the board-level empty state carries no lane count, so no plural family is needed.Four legs, chosen so the guarantee cannot be quietly given up:
still announce. ⛔ Route 3 is not a licence to put the
> 1guard back,which would silence them again;
descriptionat all — read from theelement's own slice rather than from the whole file, so a
descriptionpropon a future sibling cannot make it red by accident, and
card.description(adata field two hundred lines up) cannot either;
spelling. The needle is held in a variable so the file can describe what it
forbids without containing it, and it carries a positive control proving it
can fire before it is trusted;
otherwise satisfy legs 1 to 3 vacuously.
The byte-identity table lost its
kanban.columnsrow, because the call site itcompared against is gone. The count moved from sixteen to fifteen and the
departure is asserted rather than left as a silent absence.
The card's three-lane measurement becomes the pin
packages/plugin-kanban/src/__tests__/emptyStateNumberlessDescription-9170.test.tsxreads the live region's own
textContent:No cards0 columnsNo cardsNo cards1 columnsNo cardsNo cards2 columnsNo cardsThree claims, asserted as separate cases so they can be measured
independently: today's copy byte for byte; the equality across the three rows
(the card's actual claim — the same string at every lane count); and the absence
of any digit, which is the leg that survives a rewording. Run on three paths —
through the provider, provider-less, and in
ru(Нет карточек) — plus a litcontrol showing a board with cards paints no region at all, and a case asserting
objectui#9045 is not undone.
Ablation — round two
Same discipline as round one: mutation proven on disk in both directions, blob
hash moving, restore under a
trapproven by state.HEAD blob for
KanbanImpl.tsx:7b505b90.79a38534{ZERO: 'No cards0 columns', ONE: 'No cards1 columns', TWO: 'No cards2 columns'}&& boardColumns.length > 1ca8cdee0{ZERO: '(no live region)', ONE: '(no live region)', TWO: 'No cards'}: the two lane counts objectui#9045 serves go silent while the third does not. It also proves the no-digit leg is not sufficient alone — silence carries no digits — which is why the "9045 is NOT undone" case exists079d044a⭐ Leg D is the one that answers "is this an assertion that merely stopped
checking?" — a mutation this pin is supposed to tolerate on two of its three
claims leaves exactly those two green. ⭐ Leg B is the discriminating one: one
mutation, and the rows fail apart rather than together, with the readings naming
which lane counts lost the region.
exposed: the three claims used to sit in one case (so the byte pin aborted the
other two and their independence could not be measured), and a silent live region
used to read as a rig failure rather than as the regression it is.
b984d5694fixes both — the rig self-check now waits for the board, and a region that
never paints returns
(no live region), which is what makes leg B's signaturelegible.
Every leg printed
RESTORED: blob 7b505b90… matches HEAD, and the tree is cleanat the delivered commit.
kanban.columnsis NOT retired here — what the gate actually saysThe key stays in all ten packs, byte-identical. It now has no call site, and
scripts/check-i18n-dead-keys.mjs— report-only by design, because a reversesweep over dynamic key construction can produce false positives — classifies it
needs-review, notconfirmed. Measured as a delta with the total as its owncontrol:
Exactly one key moved, and it did not move into the deletable column. The textual
footprints the gate names for it are the unrelated spec key (the fields shown
on each card —
packages/core/src/utils/normalize-list-view.tssays so in as manywords) and this suite's own pins. ⛔ Retiring it is that gate's call and a
separate decision, not this PR's.
Verification — round two, at
b984d5694pnpm exec vitest run packages/i18n/ packages/plugin-kanban/Test Files 117 passed (117)·Tests 1428 passed (1428),VERDICT command-exit 0pnpm exec vitest runover the 15 test files matchingbuiltInLocalesoutsidepackages/i18nTest Files 15 passed (15)·Tests 293 passed (293)pnpm --filter @object-ui/i18n --filter @object-ui/plugin-kanban run type-checkpnpm --filter @object-ui/i18n --filter @object-ui/plugin-kanban run lintcheck:i18n-keys,check:i18n-drift,check:i18n-dead-keyscheck-changeset-presence.mjs,check-changeset-no-major.mjsmajorcheck:control-bytescheck-governed-queue-guard.mjs --testover the whole diffNOT GOVERNED — 5 path(s) checkedChangeset
.changeset/9170-kanban-empty-state-drops-lane-count.mddeclarespatchon@object-ui/plugin-kanban. Rendered copy in a published component changed; noexported symbol, no schema and no published payload moved — no locale pack is
edited, no key added, renamed or retired — so there is nothing here that a
minorwould be carrying. All 41 packages sit in onefixedgroup, so therelease level is the max across pending changesets rather than a per-package
choice, and
majoris refused mechanically bycheck-changeset-no-major.mjs.Round one's changeset keeps its "no published payload" sentence, which is true of
both halves, and now points at the companion changeset for the consequence that
landed beside it.
needs:contract-reviewlabel is still on this PR, deliberatelyIt was hung when round two was going to change a published payload. Route 3 does
not, so the
Clause-②: nodeclaration round one was claimed under stands againand the gate's premise is gone. ⛔ It is not cleared here: an agent seat
never issues its own clause-② verdict, and a gate left hanging is a red light
rather than a green one. A reviewer clears it.
Changed files — round two
packages/plugin-kanban/src/KanbanImpl.tsxpackages/plugin-kanban/src/__tests__/emptyStateNumberlessDescription-9170.test.tsxpackages/i18n/src/__tests__/residue-namespaces-3546.test.tsx.changeset/9170-kanban-empty-state-drops-lane-count.mdpatchon@object-ui/plugin-kanban.changeset/9045-kanban-empty-state-lane-count.mdRound two was implemented by an
os-devseat in Claude Code sessionsession_01UzHd6hDYatoDn17BuwKxnZ, which is also the session named in thefooters below; that reference is written here in prose because a
PATCHto a PRbody rewrites the footer form.
🤖 Generated with Claude Code
https://claude.ai/code/session_01UzHd6hDYatoDn17BuwKxnZ
Generated by Claude Code