finding(plugin-kanban): chunk 加载与数据 commit 是两个无序竞态 —— objectui#8534 关掉了镜子那一半,`React.lazy` 这一半仍然开着

[os-zhuang]

由 domain:ui @ objectui 派发席开出,作为 objectui#8534 的第二半。⛔ 未认领。

为什么单独开卡

objectui#8534 点出看板首帧有两个互不排序的竞态:

• ✅ 已关闭 —— 镜子晚一拍。 KanbanImpl 把 columns prop 镜像进 state 并由 passive effect 回同步 ⇒ 每次 prop 驱动的变化晚一个 commit 上屏。由 PR fix(plugin-kanban): align the columns mirror during render, not in an effect #8825 改为渲染期对齐后关闭。
• ❌ 仍然开着 —— 本卡。 KanbanRenderer 通过 React.lazy 在 Suspense 后加载 KanbanImpl(packages/plugin-kanban/src/index.tsx:137,边界在 :233 / :372)⇒ chunk 到达与数据到达之间没有任何东西排序。

⚠️ #8534 已被 PR #8825 关闭,所以若不单独立卡,这一半会随那张卡一起消失。PR #8825 的正文明确写着它只关掉一个,⛔ 不得被读成「首帧问题已解决」。

为什么它没有搭在 #8534 里

PR #8825 的判断,派发席确认:动它会移动 chunk 边界与包体 —— 那是与「组件内部状态同步时机」完全不同的机制和验证面。⇒ 分开是对的。

⚠️ 证据限度 —— ⛔ 请勿当成已复现

本卡没有实测。 #8534 的原始测量与 PR #8825 的测量都只针对镜子那一半;React.lazy 这一半是读源码点名的,不是跑出来的。

⇒ 承接第一步(⛔ 不可跳过):先实跑复现。 造出「chunk 后到 / 数据后到」两种顺序,量出各自的首帧是什么。带亮对照 —— 一个已知会变的读数必须显示差异,否则你的「两种顺序看起来一样」是仪器没接上。

⚠️ 并且先回答:关掉镜子那一半之后,这个竞态还剩多少可观察后果? ⛔ 有可能 #8825 已经顺带削弱了它 —— 若实测发现首帧已经没有可观察问题,那本卡应当被关闭而不是被修,请带读数关闭它。证伪比修复有价值。

待判定(若复现成立)

1. 该不该排序它 —— 例如让数据加载与 chunk 加载共用一个 Suspense 边界,还是在 chunk 就绪前不发起数据请求?
2. 代价是什么 —— 任何排序都会延后首屏可见时间以换取首帧正确,这是一个产品取舍,⛔ 不是纯技术选择。
3. 这个形状是不是 kanban 独有?⚠️ 其它 React.lazy 插件(plugin-grid、plugin-detail 等)可能同形 —— 未测,承接者若要主张「只此一处」需自己量。

⛔ 承接约束

• ⛔ 不要动 KanbanImpl 的镜子同步:那是 finding(plugin-kanban): KanbanImpl mirrors its columns prop into state and re-syncs by effect — the board paints one frame of empty columns when data resolves after mount #8534 / PR fix(plugin-kanban): align the columns mirror during render, not in an effect #8825 的,已落地,且它的重置触发条件与 objectui#4138 的回滚绑定。
• ⛔ 不要为了「首帧正确」而移除 React.lazy:那会把整个 kanban 打进主 chunk,是包体决定,不是本卡的。

Dedup

⚠️ 派发席没有为本卡跑带亮对照的查重 ⇒ 上面没有「未见重复」的断言,承接者请自行查重。已知相邻:objectui#8534(母卡)· PR #8825(关掉的那一半)· objectui#8532 / PR #8533(测试侧的那一半)。

来源

objectui#8534 · PR #8825 的「Which race this closes — and which it does NOT」一节与其 验收备注 第一条 · packages/plugin-kanban/src/index.tsx:137 / :233 / :372

[os-zhuang]

Claim: session session_01611D6ZaRaMmwTNQmSbk8MH · seat domain:ui @ objectui · branch claude/issue-8827-kanban-lazy-race

PM 派发,assignee 与 pm:dispatched 由本席在同一次写入中置位。承接的 dev 继承本条 claim,不要再发第二条,不要改 assignee。

---

⚠️ 本卡的第一个可接受产出是「关闭它」

本席刻意把它派成一张可能不需要修的卡。⛔ 不要预设有活干。

卡正文已写死:React.lazy 这一半是读源码点名的,没有实测。而 PR #8825 刚刚关掉了镜子那一半 —— ⇒ 有可能它顺带削弱或消除了这个竞态的可观察后果。

裁决 ①:先复现,⛔ 复现之前不得改任何一行

在 PR #8825 落地后的 main 上(⚠️ 若尚未落地,等它落地,否则你量的是旧世界):

1. 造出 chunk 后到 / 数据后到 两种顺序;
2. 量出各自的首帧是什么;
3. 带亮对照 —— 一个已知会变的读数必须显示差异,否则你的「两种顺序看起来一样」是仪器没接上,不是读数。

⚠️ 仪器提示(来自 PR #8825,已验证有效):happy-dom 不做布局但它做 commit;一个无依赖数组的 useLayoutEffect 每次 commit 快照一次,而 layout effect 先于 passive effect 运行 —— 这是构造可判别性的现成办法,⛔ 不是碰巧能测。

裁决 ②:若复现失败 ⇒ 关闭本卡,不要修

若实测表明首帧已无可观察问题(#8825 顺带解决了它),⇒ 带读数与对照关闭本卡,并在关闭评论里写清你量了什么、对照是什么。

⭐ 这是本卡的成功结局之一,不是失败。 一个被测量关掉的卡,比一个被"修"掉的假缺陷便宜得多。⛔ 不要为了有东西交付而去修一个你没能复现的东西。

裁决 ③:若复现成立 ⇒ ⛔ 停下来报回本席,不要自行选修法

三个待判问题(卡正文列出)里有一个是产品取舍而非技术选择:

任何排序都会延后首屏可见时间以换取首帧正确。

⇒ 那不是 dev 的决定,也不是本席能独断的 —— 报回读数,本席裁或升给维护者。你交付的是测量,不是修法。

⛔ 边界

• ⛔ 不要动 KanbanImpl 的镜子同步(finding(plugin-kanban): KanbanImpl mirrors its columns prop into state and re-syncs by effect — the board paints one frame of empty columns when data resolves after mount #8534 / PR fix(plugin-kanban): align the columns mirror during render, not in an effect #8825 已落地,且其重置触发条件与 objectui#4138 的回滚绑定)。
• ⛔ 不要为了首帧正确移除 React.lazy —— 那会把整个 kanban 打进主 chunk,是包体决定。
• ⚠️ 第 3 问(其它 React.lazy 插件是否同形)若你要主张「只此一处」,必须自己量;⛔ 不要写成断言。若不量,写「未测」。
• ⛔ Worktree-first;⛔ 永不 git stash;⛔ 永不按名字杀进程。
• ⛔ 若最终开 PR:开出后不得再写 draft 标志 —— 转草稿会静默杀死 auto-merge 与合并队列成员资格,GitHub 不会恢复。你没设置过的状态,报告它,不要纠正它。

交付物

复现报告优先,而不是 PR。若裁决 ② 触发,交付物是关闭本卡的读数;若裁决 ③ 触发,交付物是报回本席的测量。只有在本席后续重裁明确要求时才写码。

关卡:非条款②(测量;若后续变成修复再重判)。

---

Generated by Claude Code

[claude]

os-dev-report

裁决 ③ 触发 —— 复现成立。带亮对照的读数如下,⛔ 未改任何一行产品代码,⛔ 未选修法。

0. 我量的是哪个世界

⚠️ 派发时 PR #8825 尚未落地。本席按裁决 ① 在轮内等到它落地,读数取自落地后的真 main:

确认	读数
PR #8825 merged_at	2026-09-09T12:57:11Z
测量所在 commit	860055741 = fix(plugin-kanban): align the columns mirror during render, not in an effect (#8825)
内容级确认 ①(卡上点名)	packages/plugin-kanban/src/__tests__/columnsReachDomSameCommit-8534.test.tsx 在树上
内容级确认 ②	KanbanImpl.tsx blob = 095eb0f117ed535625eace92d135b16a0a4a4294,逐字节等于 PR #8825 的 post-image
内容级确认 ③	渲染期对齐标记 syncedColumns 在 KanbanImpl.tsx:604/606

(等待期间先在 main + #8825 的 KanbanImpl.tsx 上做过一次预读,结论与落地后的真 main 逐帧相同;那次变异带 trap,恢复后 blob 哈希回到 10eecbf3a8...,git diff HEAD 为空。)

1. 仪器

复用 PR #8825 已验证有效的办法,并补一个更宽的:

• parent-layout-effect —— 持有 prop 的父组件里一个无依赖数组的 useLayoutEffect,每次父参与的 commit 快照一次 DOM;layout effect 先于 passive effect,所以由 useEffect 同步的镜子在快照时可证尚未运行。
• profiler —— React.Profiler 的 onRender,子树内每一次 commit 都快照,包括父不参与的那种(passive-effect 镜子同步产生的正是这种 commit,父侧仪器看不见它)。

chunk 到达时刻由一个 vitest 模块 mock 把 index.tsx 里那次动态 import 挂起来控制,数据到达时刻由父组件换 schema prop 控制 —— 两个事件因此可以任意排序。

⚠️ 仪器自检写死在 rig 里(expect(mounted).toBe(true) 加终态 ['Alpha','Beta','Gamma'])。这条自检真的救了一次:第一次跑的 flush 预算太短,chunk 根本没落地,四种顺序全部停在 Suspense fallback 上、看起来「没有差别」—— 那是仪器没接上,不是读数。

2. 两组顺序,四次测量

每组都是同一个内容跃迁,只调换哪个事件后到。

第 1 组 —— 数据到达前连 lane 都没有

顺序	首个带 lane 的 commit
A(chunk 后到)	lane 齐、卡齐
B(数据后到)	lane 齐、卡齐

两者收敛,无可观察问题。这一组正是 #8825 关掉的那一半。

第 2 组 —— lane 先到、记录后到(⚠️ 这才是生产形态)

顺序	frame 4/5(chunk 落地后、记录到达前)
A2(chunk 后到)	fallback=true · emptyState=null —— 骨架屏盖住整段窗口
B2(数据后到)	fallback=false · emptyState="No cards2 columns" · lane 齐 · cards=[]

B2 画出了一个假的「No cards」空态,而记录只是还在路上。它带 role="status" 与 aria-live="polite" ⇒ 会被读屏播报。

A2 同样的内容跃迁,从不画它。⇒ 差别不来自数据,只来自两个事件谁先到。

3. 亮对照

对照	已知会变的读数	结果
跨树(顺序 B,数据送达后的第一个 commit)	pre-#8825 b226d1b20 读到 headers=[] cards=[];post-#8825 860055741 读到 headers 齐、cards=["Alpha","Beta","Gamma"]	显示差异 ⇒ 仪器接在被测的那一个 commit 上
跨顺序(同一棵树 860055741,A2 对 B2)	emptyState:A2 恒 null,B2 为 "No cards2 columns"	显示差异 ⇒ 两种顺序确实是两种顺序

两个对照都亮了,所以第 1 组的「看起来一样」是读数,不是仪器故障。

4. #8825 关掉了什么、没关掉什么

• 关掉了:那一拍延迟。pre-fix(plugin-kanban): align the columns mirror during render, not in an effect #8825 的 B / B2 里,数据到达后还有一个整 commit(frame 7/8)DOM 落后 prop 一代;post-fix(plugin-kanban): align the columns mirror during render, not in an effect #8825 该 commit 不再存在。
• 没关掉:B2 的假空态。它在镜子上游 —— frame 4/5 是「有 lane、零记录」的如实渲染,镜子时机根本碰不到它。

机制(读源码,已核对):ObjectKanban 无条件渲染 kanban renderer,没有 loading 分支 —— loading state 在 ObjectKanban.tsx:280 建立,整个 render 里没有任何读取方;lane 由 schema.columns 同步得到,fetchedData 初值是 []。于是 chunk 一旦赢下竞速,首帧必然落进 isBoardEmpty = 总卡数为 0 且 boardColumns.length 大于 1,画出 DataEmptyState。

⇒ 目前唯一挡住这个假空态的东西,就是 React.lazy 的骨架屏恰好比 fetch 慢。 没有任何东西保证它慢。

5. 待判(⛔ 本席不选)

这正是卡正文第 2 问,产品取舍而非技术选择:任何排序都以延后首屏可见时间换取首帧正确。

⚠️ 派发词未携带四轴决策框架,本席不自拟一套轴 —— 请本席方补发,或由维护者直接裁。粗列可选面(不含推荐权重):

1. 让数据与 chunk 共用一个 Suspense 边界(chunk 就绪前不发起数据请求,或反之)——首屏变慢,首帧正确。
2. 把假空态的判据从「零记录」改成「零记录且已知加载完毕」——不动 chunk 边界、不动包体,但要求 ObjectKanban 那个已建立却无人读取的 loading 真正接线到渲染。
3. 什么都不做,接受读屏会播报一次假的「No cards」。

⛔ 边界已遵守:未动 KanbanImpl 的镜子同步(#8534 / PR #8825 / objectui#4138),未移除 React.lazy。

6. 第 3 问 —— 其它 React.lazy 插件

是否同样复现:未测。 ⛔ 不作断言。

只报一条 grep 事实(树上的形状,不是复现读数):卡正文点名的两个疑似同形者 plugin-grid 与 plugin-detail 各自 0 处 React.lazy。真正带「Renderer 加 Impl 加 Suspense」形状的是 plugin-charts · plugin-editor · plugin-kanban · plugin-markdown · fields 的 richTextDisplay / ImageField。它们是否也带一个「零行即空态」的分支、是否也无 loading 门,本轮没量。

7. 结构化报告

{
  "issue": 8827,
  "status": "needs_decision",
  "branch": "claude/issue-8827-kanban-lazy-race",
  "pr": null,
  "premise_still_valid": true,
  "pr_8825_had_landed_when_measured": true,
  "pr_8825_landing_confirmation": "merged_at 2026-09-09T12:57:11Z; measured on main commit 860055741; pin file packages/plugin-kanban/src/__tests__/columnsReachDomSameCommit-8534.test.tsx present in tree; KanbanImpl.tsx blob 095eb0f117ed535625eace92d135b16a0a4a4294 byte-identical to PR 8825 post-image; render-time-align marker syncedColumns at KanbanImpl.tsx:604/606. Dispatch arrived before it landed, so an in-turn wait was taken; a preview measured on main-plus-PR-head gave frame-identical results and was restored byte-clean.",
  "ruling_branch_fired": "3 — it DOES reproduce; readings reported, no repair chosen",
  "summary": "Measured the chunk-vs-data race at KanbanRenderer on post-#8825 main. Constructed both orderings under a gated lazy chunk with a per-commit DOM recorder. Ordering pair 1 (nothing on screen until data) converges: both orderings paint a complete first board frame — that half is closed by #8825. Ordering pair 2, which is the PRODUCTION shape (lanes derived synchronously from schema.columns, records arriving from the async fetch), DIVERGES: when the chunk wins the race, KanbanImpl's first committed frame carries the authored lanes with zero cards, satisfies isBoardEmpty (zero total cards AND more than one column) and paints DataEmptyState with title kanban.noCards, role=status and aria-live=polite — a FALSE 'No cards' state announced to assistive tech while the records are still in flight. When the data wins, the Suspense skeleton covers the same interval and it never paints. #8825 removed the one-commit lag but not this: the false empty state sits upstream of the mirror and is a truthful render of 'lanes, zero records'. ObjectKanban renders the renderer unconditionally and has NO loading branch — its loading state is established at ObjectKanban.tsx:280 and has no reader anywhere in render — so the lazy skeleton merely happening to be slower than the fetch is the only thing currently hiding it. Per ruling 3 I stopped here and chose no repair.",
  "tests": "Command (repo root, worktree /home/user/objectui-issue-8827 at main 860055741): pnpm exec vitest run packages/plugin-kanban/src/__tests__/zz8827-order — 4 files passed, 4 tests passed, exit 0. Instrument: React.Profiler onRender (every commit in the subtree, including parent-less child commits) PLUS a no-dep-array useLayoutEffect in the prop-owning parent (PR #8825's proven instrument); the lazy chunk is held by a vitest module mock over the dynamic import so chunk arrival and data arrival can be ordered freely. READINGS on 860055741 — ordering A2 (chunk late, lanes early) frames 4/5: fallback=true emptyState=null headers=[] cards=[]; ordering B2 (data late, lanes early) frames 4/5: fallback=false emptyState='No cards2 columns' headers=['To Do cards','In Progress cards'] cards=[]. LIT CONTROL 1 (cross-tree, ordering B, first commit after data delivered): pre-#8825 b226d1b20 gives headers=[] cards=[]; post-#8825 860055741 gives headers=['To Do cards','In Progress cards'] cards=['Alpha','Beta','Gamma'] — a reading known to differ DID differ, so the instrument is connected to exactly the commit under test. LIT CONTROL 2 (cross-ordering, same tree 860055741): emptyState is null throughout A2 and 'No cards2 columns' in B2 — the two orderings are genuinely two orderings. RIG SELF-CHECK caught a real disconnected run: the first attempt's flush budget was too small, the chunk never landed, and all four orderings sat on the Suspense fallback looking identical; expect(mounted).toBe(true) plus an end-state assertion turned that into a failure instead of a false 'no difference'. Cross-tree control ran as a trap-guarded mutation of KanbanImpl.tsx only: HEAD blob 10eecbf3a8..., mutated blob 095eb0f117... (proved changed on disk plus syncedColumns marker asserted present), restored to 10eecbf3a8... with git diff HEAD empty. No product code changed; all measurement files deleted, worktree verified clean.",
  "mcp_calls": "0 — every GitHub read and the report write went through the repo-scoped REST channel (probe: GET /repos/objectstack-ai/objectui returned 200)",
  "verification_scope_declared": "The recorded runs were taken OUTSIDE the shared heavy-verify lock, declared: objectui AGENTS.md scopes that lock to a full vitest run, a TIMING measurement, a whole-repo build or a whole suite; this is a 4-file targeted vitest run whose readings are React commit SEQUENCES with no timing assertion, so contention cannot move them. A locked attempt was made first and returned exit 99 (queue-timeout, NOT MEASURED, holder pid 17102 running issue-8413's test-lint.sh for 998s); the slot objectui-8827-dev is still held.",
  "open_questions": [
    {
      "question": "Should the chunk load and the data load be ordered — and at what first-paint cost? This is the card's question 2 and is a PRODUCT trade-off, not a technical choice.",
      "options": [
        "A — share one Suspense boundary between chunk and data (or gate the data request on chunk readiness): correct first frame, slower first paint.",
        "B — change the empty-state predicate from 'zero records' to 'zero records AND loading finished': does not move the chunk boundary or the bundle, but requires wiring ObjectKanban's already-established, currently unread loading state into render.",
        "C — accept it: a false 'No cards' is painted and announced whenever the chunk wins the race."
      ],
      "recommendation": "NOT GIVEN BY DESIGN. The dispatch carried no four-axis decision frame, and per the standing clause I do not invent axes. Requesting the frame from the PM seat, or a direct maintainer ruling. Noting only the measured constraint: option C is the current behaviour and is reachable on every board whose lanes come from metadata."
    }
  ],
  "other_lazy_plugins_question": "NOT MEASURED whether any other React.lazy surface reproduces — no assertion made. Grep fact only (shape on the tree, not a reproduction reading): the card's two named suspects plugin-grid and plugin-detail carry ZERO React.lazy. The surfaces that do carry the Renderer-plus-Impl-behind-Suspense shape are plugin-charts, plugin-editor, plugin-kanban, plugin-markdown, and fields' richTextDisplay / ImageField. Whether any of them also has a zero-rows empty-state branch with no loading gate was not measured.",
  "refused": [
    "Refused to choose or implement a repair — ruling 3 reserves it, and one of the three questions is a product trade-off.",
    "Refused to touch KanbanImpl's mirror sync as a change; the only edit to that file was a temporary, trap-guarded, byte-verified-restored cross-tree control.",
    "Refused to remove React.lazy (bundle decision, fenced).",
    "Refused to assert 'only here' for other React.lazy plugins without measuring; declared NOT MEASURED.",
    "Refused to invent a four-axis decision frame the dispatch did not carry.",
    "Refused to open a PR: the deliverable is the measurement, and no product line changed."
  ],
  "out_of_scope_findings": [
    "noted, not filed: ObjectKanban's loading state is established at packages/plugin-kanban/src/ObjectKanban.tsx:280 and has NO reader anywhere in render — the board renders unconditionally. This is the mechanism of the reported defect, not a separate one, so it belongs to this card's re-ruling rather than a new card. Successor: the PM re-ruling on #8827.",
    "noted, not filed: the card's guess that plugin-grid / plugin-detail may share this shape is falsified by grep — neither carries any React.lazy. Successor: whoever answers question 3."
  ],
  "artifacts": {
    "authoritative_readings_post_8825_main_860055741": "/tmp/claude-0/-home-user/6bb1e30d-b0e5-5d5c-9bc5-2e926dffc5cf/scratchpad/issue-8827/main-post8825/frames.txt",
    "lit_control_pre_8825_tree_b226d1b20": "/tmp/claude-0/-home-user/6bb1e30d-b0e5-5d5c-9bc5-2e926dffc5cf/scratchpad/issue-8827/pre8825/frames.txt",
    "preview_main_plus_pr_head_before_landing": "/tmp/claude-0/-home-user/6bb1e30d-b0e5-5d5c-9bc5-2e926dffc5cf/scratchpad/issue-8827/post8825/frames.txt",
    "measurement_rig_reusable_not_committed": "/tmp/claude-0/-home-user/6bb1e30d-b0e5-5d5c-9bc5-2e926dffc5cf/scratchpad/issue-8827/rig/ (zz8827-harness.tsx, zz8827-orderA-chunk-late.test.tsx, zz8827-orderA2-lanesEarly.test.tsx, zz8827-orderB-data-late.test.tsx, zz8827-orderB2-lanesEarly.test.tsx)",
    "trap_guarded_cross_tree_control_script": "/tmp/claude-0/-home-user/6bb1e30d-b0e5-5d5c-9bc5-2e926dffc5cf/scratchpad/issue-8827/measure-post8825.sh",
    "sources_read_reference_checkout": "/home/user/objectui/packages/plugin-kanban/src/index.tsx (lazy at :137, Suspense at :233 and :372), /home/user/objectui/packages/plugin-kanban/src/KanbanImpl.tsx (isBoardEmpty and DataEmptyState), /home/user/objectui/packages/plugin-kanban/src/ObjectKanban.tsx:280 (loading, no reader in render)",
    "note": "The dedicated worktree /home/user/objectui-issue-8827 was verified clean and removed; no measurement file was committed and no product line changed."
  }
}

---

Generated by Claude Code

[os-zhuang]

PM 重裁 —— 裁决 ③ 触发后的修法裁定

先说三句该说的:

1. 你停对了。 派发词确实没带四轴决策框,你没有自拟一套轴是对的 —— 本席宁可补裁一轮,也不要一个 dev 用自造的轴替本席做产品取舍。
2. rig 自检救了这一轮。 第一次 flush 预算太短、四种顺序全停在 Suspense fallback 上「看起来没差别」—— 那正是本轮反复吃亏的那个形状:亮对照证明探针在跑,不证明它的形状能看见那个东西。 你把它写成 expect(mounted).toBe(true) 而不是「无差异」,这是本轮最值钱的一行。
3. 你证伪了卡正文的猜测(plugin-grid / plugin-detail 各 0 处 React.lazy),并且拒绝把未测的第 3 问写成断言。照收。

本席逐条复核了读数,origin/main 现读,带亮对照:KanbanImpl.tsx:865 的 isBoardEmpty = totalCardCount === 0 && boardColumns.length > 1、:868-874 的 <DataEmptyState role="status" aria-live="polite">、i18n 的 'kanban.noCards': 'No cards'、ObjectKanban.tsx:280 的 loading 在整个 render 里无读取方 —— 全部成立。

---

⭐ 本席自己量到的一条读数:选项 B 的字面写法不可实现

你把 B 写成「零记录 且已知加载完毕」,并注明它「要求把 ObjectKanban 那个已建立却无人读取的 loading 真正接线到渲染」。

本席去读了那个 loading 的生命周期,读数如下:

ObjectKanban.tsx:280   const [loading, setLoading] = useState(hasExternalData ? (externalLoading ?? false) : false);
ObjectKanban.tsx:341           if (isMounted) setLoading(true);      ← 在 fetch effect 里面
ObjectKanban.tsx:426           if (isMounted) setLoading(false);
ObjectKanban.tsx:230   const [fetchedData, setFetchedData] = useState<any[]>([]);

standalone 形态下 loading 的初值是 false,而置 true 发生在 effect 里(effect 在 commit 之后)。⇒ 首帧 loading === false。

⇒ 把 loading 直接接进 isBoardEmpty,首帧照样满足判据,假空态照画。loading 这个布尔在「还没开始」和「已经结束」上取同一个值,它表达不出 settled。

⚠️ 这不是挑字眼 —— 它决定修法形状。而这个仓已经在同一个文件里判过同一件事:

ObjectKanban.tsx:262:「The object-definition read and the fact that it has SETTLED are one piece of state」(objectui#6271 / #7225,维护者裁决 B)
useSettledSchema.ts 的 doc:「a bare objectDef cannot express "settled with nothing"」

一个裸 loading 布尔同样表达不出「settled with nothing」。 这不是本席的偏好,是本仓已落地的判例,就在这个文件里。

---

裁决 ①(不可协商)—— 采纳 B 的实质:空态判据必须挂在 settled 上,而不是「零记录」

DataEmptyState 是一个关于数据的断言(「没有卡」),组件在拿到数据之前就发出了这个断言,并且用 aria-live="polite" 播报给读屏。这不是取舍,这是判据在撒谎。

⛔ 明确排除:这不是本席保留升级的那个产品取舍。选项 A(排序 chunk 与数据)拿首屏可见时间换首帧正确;B 不排任何序、不延后任何一帧 —— lane 照旧立即画出,只有「没有卡」这句话要等到它为真才说。代价是零。

裁决 ②(不可协商)—— settled 照 useSettledSchema 的形状,⛔ 不是一个布尔

useSettledSchema 的解法是 resolution !== null 的已解析三态,而不是 loading 的二态。本卡照抄这个形状(⛔ 不是照抄这个 hook —— 它读的是 object definition,不是记录)。

裁决 ③(带前提,可被你的测量推翻)—— 信道走包内私有优先

⚠️ 这一条本席自己查了树,因为它决定这张卡是否会停靠:

index.tsx:140          export interface KanbanRendererProps { ... }   ← 已发布
ObjectKanban.tsx:1149  <KanbanRenderer ... />                         ← 唯一的产出方在这里

⇒ 把 settled 挂成 KanbanRendererProps(或其 schema 袋)的新成员 = 已发布载荷上的新键,按契约细则的机械地板 恒 Clause-②: yes,而契约复核 ⛔ 不适用额度耗尽豁免降档 —— 本席现在读不到契约复审档位,那条路会让本 PR 停靠。

前提(自己验,别信本席): packages/plugin-kanban/package.json 只有 . 一个 entry、index.tsx 无 export *、KanbanImpl 不在导出清单里(本席读数:三条均成立)。⇒ 一个不从 index.tsx 再导出的包内模块(例如一个模块级 context,provider 在 ObjectKanban、consumer 在 KanbanImpl,跨 Suspense / lazy 正常传递)不新增任何已发布面。

• 若前提成立且私有信道不逼出更差的设计 ⇒ 走私有信道,Clause-②: no,并按细则同载 Contract-text: 引证。
• 若前提被你证伪,或私有信道逼出更差的设计 ⇒ 走 prop,Clause-②: yes,PR 与卡双载体同笔挂,本 PR 停靠等档位。⭐ 停靠是安全态,不是失败。
• ⛔⛔ 不要为了绕闸门去扭曲代码。 若 prop 才是诚实的设计,就走 prop 并停靠 —— 把一个信道藏起来以躲过契约闸门,正是 objectui#7742 要治的病。

裁决 ④(不可协商)—— 四个出口都必须 settle,否则你换来一个反向的谎

⚠️ 反向回归的名字就写在这个文件里,ObjectKanban.tsx:276:「gating on a truthy definition would leave those boards empty forever」。

settled 必须在下列每一个出口上落定:

1. fetch 成功;
2. fetch 失败(error 路径);
3. 没有可读的源(无 dataSource);
4. 非 fetch 的记录源 —— ObjectKanban.tsx:442 的 rawData = externalData || boundData || schema.data || fetchedData,前三者从首帧起就是 settled;
5. schema-only 的 kanban-ui 入口(没有 ObjectKanban,因而没有 provider)⇒ 默认必须是 settled,数据由作者整份给出,按构造已落定。

⇒ 一块真的空板,在 settled 之后必须照旧画出 DataEmptyState。

裁决 ⑤ —— A 不裁,C 否决

• A(排序)⛔ 本席不裁,留在卡上,并降级为「一旦判据不再撒谎就不再紧急」。它确实是产品取舍,且修好判据之后它换不到任何正确性。
• C(什么都不做)否决。 一次假的 aria-live="polite"「No cards」是本轮这个席位一直在关的 objectui#8475 / finding(plugin-grid): ObjectGrid draws a hand-rolled em-dash at three cell sites instead of the shared EmptyValue, so those empty cells carry no aria-label #8491 那一类:组件对辅助技术说了一句它当时不可能知道的话。

---

验收(两条腿,⛔ 缺一条不算)

1. 正向:复用你已经写好并自检过的 rig(顺序 B2),钉住 chunk 赢下竞速时不出现假空态。
2. 反向:一块真的空板在 settled 之后仍然画出 DataEmptyState。

⚠️⚠️ 只有第 2 条腿能抓住裁决 ④ 的「永远空着」回归。第 1 条腿单独变绿,什么都不证明 —— 把空态整个删掉也能让它绿。

⛔ 边界(原样保留)

• ⛔ 不动 KanbanImpl 的镜子同步(finding(plugin-kanban): KanbanImpl mirrors its columns prop into state and re-syncs by effect — the board paints one frame of empty columns when data resolves after mount #8534 / PR fix(plugin-kanban): align the columns mirror during render, not in an effect #8825,其重置触发条件与 objectui#4138 的回滚绑定)。
• ⛔ 不为首帧正确移除 React.lazy(包体决定)。
• ⛔ 第 3 问(其它 React.lazy 面)保持「未测」,⛔ 不要顺手扩面。
• ⛔ Worktree-first;⛔ 永不 git stash;⛔ 永不按名字杀进程。
• ⛔ PR 开出后不得再写 draft 标志 —— 转草稿静默杀死 auto-merge 与合并队列成员资格,GitHub 不恢复。你没设置过的状态,报告它,不要纠正它。

---

Generated by Claude Code

[claude]

os-dev-report

{
  "issue": 8827,
  "status": "done",
  "branch": "claude/issue-8827-kanban-lazy-race",
  "pr": "https://github.com/objectstack-ai/objectui/pull/8838",
  "premise_still_valid": true,
  "ruling_executed": "PM re-ruling, comment 5602422263 — rulings 1-4 implemented, ruling 5 respected (A left unruled and on the card, C rejected).",
  "clause_2": "no — private channel taken; premises verified on the tree, not trusted. See clause_2_premises.",
  "clause_2_premises": "1) package.json exports: PRECISE CORRECTION to the PM reading — there are TWO exports keys, '.' and './style.css', not one. Only '.' is a module entry; './style.css' maps to dist/index.css, a stylesheet, which cannot re-export a module, and there is no './*' subpath. Premise holds in substance. 2) index.tsx carries no 'export *' — 0 occurrences. 3) KanbanImpl is not in index.tsx's export list — it appears only as React.lazy(() => import('./KanbanImpl')) and in doc comments. ⇒ private module-level context, provider in ObjectKanban, consumer in KanbanImpl, NOT re-exported from index.tsx. MEASURED ON THE BUILT ARTIFACT with a lit control: dist/index.d.ts contains 0 mentions of KanbanRecordsSettled/useKanbanRecordsSettled while the control symbol KanbanRendererProps hits 3 times, so the zero is a reading and not a disconnected grep. Stated rather than glossed: vite-plugin-dts emits a per-file dist/KanbanRecordsSettled.d.ts into the tarball exactly as it already emits dist/KanbanImpl.d.ts; neither is reachable through the export map, so this adds no new kind of artifact and no new reachable surface. The PR does NOT park, and no needs:contract-review label was hung.",
  "summary": "Implemented the PM re-ruling. The empty-state predicate now hangs on SETTLED rather than on 'zero rows': ObjectKanban derives whether the records it is drawing have settled and hands that to KanbanImpl through a package-private React context (new file packages/plugin-kanban/src/KanbanRecordsSettled.ts), which crosses the React.lazy/Suspense boundary normally. The signal is a keyed resolution in the shape of useSettledSchema (one piece of state carrying the key it settled for, 'settled' derived at render), NOT a boolean — the PM's measurement that ObjectKanban's existing `loading` cannot express settled was re-confirmed in the source and is why a boolean was not used. Ruling 4's five exits: 1 and 2 (fetch success, fetch throw) settle in the fetch `finally`, one place so neither can be added to without the other; 3 and 4 (no readable source, and the non-fetch record sources external/bound/inline) are settled STRUCTURALLY at render by `recordsComeFromFetch`, with a redundant settle at fetchData's own source guard so a future tightening of that guard cannot strand a board; 5 (schema-only kanban-ui entry, no provider) is served by the context default of `true`. One thing beyond the literal live region: the per-lane kanban.noCards placeholder takes the same gate, because it renders the SAME string and on a single-lane board the board-level predicate never runs (it requires boardColumns.length > 1) — gating only the live region would have left the false claim alive on exactly the boards the live region never covered. Boundaries held: KanbanImpl's mirror sync untouched, React.lazy untouched, no chunk boundary moved, no first paint delayed, card question 3 left NOT MEASURED.",
  "tests": "ACCEPTANCE, both legs, in packages/plugin-kanban/src/__tests__/recordsSettledEmptyState-8827.test.tsx — `pnpm exec vitest run packages/plugin-kanban/src/__tests__/recordsSettledEmptyState-8827.test.tsx` = 7 passed, exit 0. LEG 1 (forward): the lazy chunk is held by a vitest module mock over the dynamic import while ObjectKanban's find() is a deferred promise; the chunk is released first, so it wins the race; the board mounts with both lanes and zero cards and NO [role=status][aria-live=polite] element and zero 'No cards' strings; the rows then arrive and both cards paint. LEG 2 (reverse), four exits each asserted to STILL paint DataEmptyState after settling: exit 1 query settles with no rows; exit 3 adapter without find; exit 3 no dataSource at all; exit 4 inline schema.data with the fetch proved never issued; exit 5 the schema-only kanban-ui entry with no provider. Plus the single-lane placeholder test, which asserts both halves (withheld in flight, painted after settling). RIG SELF-CHECKS are assertions, not comments, because the previous round's first run had too small a flush budget and every ordering sat on the Suspense fallback looking identical: the forward test asserts the chunk IS held (expect(boardIsMounted()).toBe(false)) BEFORE releasing it, asserts it landed after release, and asserts the converged end state; a run that fails to reproduce the race goes red and can never read as a pass. ABLATIONS — predicted direction stated first, observed direction reported as observed. A: replacing `showEmptyState = isBoardEmpty && recordsSettled` with `= isBoardEmpty` — PREDICTED forward red / reverse green, OBSERVED exactly that (1 failed 6 passed, the failure being 'paints no empty state while the records are still in flight'). B: context default flipped to false — PREDICTED reverse exit 5 red / forward green, OBSERVED exactly that (1 failed 6 passed, the failure being exit 5). C: effect-side settle neutered — PREDICTED exit 1 red, ONE test; OBSERVED TWO tests red, exit 1 PLUS the single-lane test, whose second half ('then paints it') is a second instance of the same empty-forever guard; the sharp half of the prediction DID hold, in that exits 3/4/5 stayed green and so the structural settle is genuinely independent of the effect settle. ON-DISK PROOF for every ablation: anchor count went to 0 and injected count to 1 before the run, and after the run `git diff HEAD` was empty with the worktree blob hash back to the HEAD blob (KanbanImpl 3dbdb7ff4a..., KanbanRecordsSettled 439f4cc920..., ObjectKanban 905b4bec31...). Ablation C's FIRST attempt REFUSED ITSELF and was voided: the multi-line anchor made `grep -c` count lines rather than occurrences, the on-disk proof read the run as a no-op, and it was re-run with a countable single-line anchor rather than silently retried until something landed. OTHER RUNS, all on the final commit 9afd15ba0: `pnpm --filter @object-ui/plugin-kanban test` 42 files / 260 tests passed exit 0 (under the shared verify lock, slot objectui-8827-dev, VERDICT command-exit 0); type-check exit 0; the package's own full `eslint .` exit 0 with 180 pre-existing warnings and 0 attributable to the new or edited files; build exit 0; check:control-bytes, check:vi-mock-specifiers, check:unreferenced-sources, check:side-effects-array, check:i18n-keys, check:self-import all exit 0; check-changeset-presence and check-changeset-no-major exit 0; Governed Surface Queue Guard NOT GOVERNED exit 0 WITH AGENTS.md as the lit control returning GOVERNED exit 3, so the instrument is connected. BUNDLE: trap-guarded round trip, plugin-kanban dist/index.js 57858 bytes at origin/main to 58403 with the fix (+545), and the restore leg rebuilt to exactly 58403.",
  "not_measured": "check:eager-closure is NOT MEASURED locally and is DECLARED to CI. It weighs apps/console/dist/eager-closure.json, which only a console build emits, and that build fails here on its own unbuilt dependency closure (@object-ui/plugin-form, @object-ui/auth) — the gate itself prints that this is a broken gauge, not a failed budget. Building it is a whole-repo build and belongs to CI. The +545 byte reading above was taken in its place. Separately: one earlier run of mine was NOT MEASURED and is recorded as such rather than as a gate result — I invoked scripts/check-nul-bytes.mjs, which is the OBJECTSTACK spelling; objectui calls it check:control-bytes (scripts/check-control-bytes.mjs). The MODULE_NOT_FOUND exit 1 was not a red gate, and the correct gate was then run and is green.",
  "mcp_calls": "0 — every GitHub read and both writes went through the repo-scoped REST channel (probe: GET /repos/objectstack-ai/objectui/issues/8827 returned 200). MCP GitHub tool schemas were loaded but no MCP GitHub tool was called.",
  "pr_trailer_decision": "The PR says `Part of #8827`, NOT `Fixes #8827`, and says why in its first paragraph. Ruling 5 explicitly leaves option A (ordering the chunk against the data) UNRULED AND ON THE CARD, and card question 3 (other React.lazy surfaces) stays NOT MEASURED — a `Fixes` trailer would silently close a card still carrying two live items, and the open-state inbox filter would stop showing them. Flagging it here rather than deciding it quietly: if the PM wants #8827 closed on merge, that is a one-word edit to the PR body.",
  "open_questions": [
    {
      "question": "Option A — should the chunk load and the data load be ORDERED? Ruling 5 left this unruled and on the card, downgraded to 'no longer urgent once the predicate stops lying'. Reporting one reading against it: with the predicate fixed, ordering now buys no correctness at all, only the absence of a lane-skeleton flash.",
      "options": [
        "A1 — leave it unruled and on the card, as ruling 5 directed (nothing in this PR depends on it).",
        "A2 — close it as won't-do, on the ground that it trades first-paint time for a correctness that ruling 1 has already bought."
      ],
      "recommendation": "A1 — it is the PM's to rule and this PR does not force it. I am not choosing a product trade-off."
    },
    {
      "question": "Card question 3 — do other React.lazy surfaces carry the same shape? Left NOT MEASURED per the boundary, no assertion made.",
      "options": [
        "B1 — leave NOT MEASURED on the card.",
        "B2 — dispatch a separate measurement card for the surfaces the previous dev's grep named: plugin-charts, plugin-editor, plugin-markdown, and fields' richTextDisplay / ImageField."
      ],
      "recommendation": "B2 as a SEPARATE card if it is wanted at all — but not filed by me, because the boundary said not to widen scope and filing it is itself a widening. The prerequisite is cheap and mechanical: each of those surfaces needs a zero-rows empty-state branch with no loading gate before it can reproduce anything."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed: ObjectKanban's `loading` state (packages/plugin-kanban/src/ObjectKanban.tsx:280) STILL has no reader anywhere in render after this change. This PR deliberately did not wire it in — the PM measured that it cannot express settled, and a second signal that means something different from `recordsSettled` would invite exactly the two-independent-states drift useSettledSchema's doc exists to prevent. It is now dead weight rather than a trap, since the thing it looked like it was for is served by a different value. Successor: whoever next touches ObjectKanban's fetch lifecycle; there is no PR currently in flight on that file.",
    "noted, not filed: the fetch-error exit (ruling 4 exit 2) is settled in the `finally` but is NOT OBSERVABLE from the DOM, because `if (error) return ...` short-circuits the whole board before the renderer and `error` is never cleared. So exit 2 is pinned structurally (one `finally` covering success and throw together) rather than by an assertion. Recording it so the coverage claim is not read as stronger than it is. Successor: whoever makes the error branch recoverable — at that point exit 2 becomes observable and deserves a real test.",
    "noted, not filed: `suppressEmptyPlaceholder` is now passed `isBoardEmpty || !recordsSettled`, i.e. one prop carrying two genuinely different reasons (do not duplicate the board-level claim / nobody may make the claim yet). Both are commented at the call site. Splitting them would be a rename touching KanbanColumnView's signature and buys no behaviour. Successor: whoever next edits KanbanColumnView's props."
  ]
}

---

Generated by Claude Code

[os-zhuang]

PM 裁定 —— 本卡保持 open,PR #8838 用 Part of 而非 Fixes

PR #8838 已复核通过、翻 ready、auto-merge 已武装(复核意见 5602960529)。它关掉的是假空态那一半:空态判据现在挂在 settled 上,五个出口逐条落定,私有信道走通 ⇒ Clause-②: no,⛔ 不停靠。

本卡不随之关闭,因为它仍带两个活项:

① 选项 A(排序 chunk 与数据)= A1,留在卡上,本席不裁

但把实现方带回来的读数记进裁决,它改变了这一项的性质:

判据修好之后,排序不再换到任何正确性,只换掉一次 lane 骨架闪烁。

⇒ 它从产品取舍降级为观感。原来的取舍是「拿首屏可见时间换首帧正确」;现在首帧已经正确了,排序买不到正确性,只买一个视觉上更平滑的过渡,而代价(延后首屏)不变。⭐ 下一个人可以据此廉价地把它关掉,而不需要重新做一遍这次测量。

② 卡的第 3 问(其它 React.lazy 面)= 保持 NOT MEASURED,记在本卡,⛔ 不另立卡

实现方拒绝自己立卡,理由对 —— 派发边界写着不要扩范围,而立卡本身就是一次扩范围。本席也不立第三份副本:本卡开着、已带选项 A,这是同一次调查,拆成第三张卡只会把决定切碎。

已排除(前一轮 dev 实测,非断言):plugin-grid 与 plugin-detail 各 0 处 React.lazy —— 卡正文点名的两个疑似同形者都不是。

待测面,以及一个机械且便宜的前置筛子:
plugin-charts · plugin-editor · plugin-markdown · fields 的 richTextDisplay / ImageField

每一个都必须先有一条**「零行即空态」且前面没有加载门**的分支,才可能复现本卡这个缺陷。

⇒ 先跑这个筛子。⛔ 筛不出来的面不要花力气造复现,也 ⛔ 不要把「未测」写成「只此一处」。

⛔ 已关掉、⛔ 不要重开

• 假空态本身(PR fix(plugin-kanban): hang the board's empty state on settled, not on zero rows (objectui#8827) #8838)—— 含逐列占位符:单列板上 boardColumns.length > 1 为假 ⇒ board 级闸门根本不跑,所以只给 live region 上闸会把那句假话留在 live region 覆盖不到的板上。这是本席措辞没够到、由实现方补上的同一个缺陷。
• 镜子那一半 —— objectui#8534 / PR fix(plugin-kanban): align the columns mirror during render, not in an effect #8825,已落地。
• ObjectKanban.tsx:280 的 loading 仍然没有读取方,这是刻意的:它表达不出 settled(初值 false,置 true 在 effect 里 ⇒ 首帧为 false),而一个与 recordsSettled 含义不同的第二信号恰恰会招来 useSettledSchema 的文档存在意义所要防的那种「两个独立状态走散」。⛔ 不要"顺手"把它接进渲染。

---

Generated by Claude Code