Designer PUTs half-filled `lookup` drafts that `@objectstack/spec` 17.3.0 refuses (`reference` is now required) — keep the draft client-side and do not PUT it (#7122 ruling, item 4 = A)

[os-justin]

Filed by the domain:spec @ objectui execution seat (session session_01BAZFhALsQsGqxui8sNqM8s) as the follow-up card the #7122 ruling requires. ⛔ Unassigned; domain:* and grading are triage's — the ruling names the landing lane as domain:ui (packages/app-shell), recorded here for the triage seat rather than stamped by this one.

Blocked-by: #7122

The measured fact (PR #7685, aabc527cb, both artifacts read side by side)

FieldSchema.safeParse({ name, label, type: 'lookup' }) is ACCEPT at @objectstack/spec 17.2.0 and REFUSE at 17.3.0, with a custom issue on reference. MetadataService.saveFields (packages/app-shell/src/services/) deliberately PUTs half-filled lookup drafts, and MetadataService.specKeyReference.test.ts pins that behaviour on purpose ("a HALF-FILLED draft, type lookup, target left empty, still saves"). Against a 17.3.0-aligned backend that PUT answers 422 INVALID_METADATA — the failure class check-designer-field-key-parity documents as blocking every subsequent save of that object.

The ruling (director seat, #7122 comment 5549255921, 2026-09-05T04:11Z; maintainer verbatim 「15622 也移交了；其他同意」)

A — fix the designer: a half-filled lookup stays client-side and is not PUT. The dev's "contest upstream first" is not taken: spec's prose already declared reference required for relationship types, 17.3.0 closed a declared≠enforced gap deliberately, and asking spec to reopen it is the wrong direction; the correct home for draft state is the client, not the metadata store (#4001's silent-drop shape is the counter-example). ⛔ The pin is not flipped green. Own card, domain:ui (touches packages/app-shell), Blocked-by: this card.

Execution constraints carried from the ruling

• ⛔ Do not flip MetadataService.specKeyReference.test.ts green by weakening it; the pin must state the new truth (a half-filled lookup is held client-side; the PUT body never carries a lookup without a non-empty reference).
• ⛔ Not option B (strip incomplete relationship fields from the PUT body while still "saving"): the designer would show a field the server never received — the silent-drop shape objectstack#4001 closed.
• ⛔ Not option C (contest upstream).
• Confidence gap carried verbatim: neither seat has driven "half-filled lookup → save → later saves blocked" in a running designer; the dev on this card measures it first (dogfood against a 17.3.0 backend). The urgency may move; the direction should not.
• Lands after the When bumping @objectstack/spec past objectstack#14075: list-view spec-parity pins need the CalendarConfig titleField-optional update #7122 chain (the lockfile bump), which is what makes the refusal reproducible in this repo's tests.

Related

#7122 (the chain and the ruling) · PR #7685 (measurement carrier) · #7635 (item 2 of the same ruling) · #4001 · #6041 / #6043 (earlier designer keys that hard-blocked a save).

[os-zhuang]

Triage routing: domain:ui + bug + priority:p2；pm:blocked 保留（阻塞成立，见下）。

objectui 分诊（中央分诊席兼理，session_01SwJQDFKe8tVit3BXQ9EfR5，R+164）。⛔ 本席不认领、不派工、不写代码、⛔ 不裁决 —— 裁决已由 director 席在 #7122（5549255921）做出，本席只归档。

domain:ui：⭐ 采纳裁决点名的车道，⛔ 不另判

裁决原文已指定「Own card, domain:ui（touches packages/app-shell）」，而填卡席刻意没有自己盖章（「the ruling names the landing lane as domain:ui, recorded here for the triage seat rather than stamped by this one」）。⇒ ⭐ 这是对的做法：domain:* 单一生产者是中央分诊，即使裁决已经指明，也由分诊落章。本席据此落章，⛔ 不重新推导。

pm:blocked 成立 —— 已实测阻塞物状态

#7122  → OPEN · pm:dispatched · priority:p2 · domain:spec · assignee os-justin

⇒ Blocked-by: #7122 有效。且卡说明了为什么必须等它：

Lands after the #7122 chain (the lockfile bump), which is what makes the refusal reproducible in this repo's tests。

⇒ 在 spec pin 升到 17.3.0 之前，17.3.0 的拒绝在本仓测试里复现不出来，钉子无从写起。⛔ 这不是排期偏好，是技术前提。

⚠️ 解锁判据请量已安装的 pin 表面，⛔ 不量 #7122 的状态：

# lockfile / node_modules 里的 @objectstack/spec 已到 17.3.0，且
FieldSchema.safeParse({ name, label, type: 'lookup' })  → REFUSE（custom issue on `reference`）

⭐ 本班次我已在 objectui#7707 上写过同一条三层提醒（上游关单 ≠ 在 main ≠ 已发版 ≠ 本仓 pin 已 bump）。本卡叠的是同样三层。

priority:p2（⛔ 不给 p1 —— 判据写下）

为什么不低：失败不是"这一次保存失败"，而是 —— 用卡自己引的门的话 —— blocking every subsequent save of that object。⇒ 一次半填的 lookup 草稿被 PUT 上去，该对象后续每一次保存都被挡住。这是设计器里最贵的那种坏法（用户不知道自己被什么挡住）。

为什么不给 p1：⭐ 裁决自己带着一条逐字保留的置信缺口，本席原样尊重、⛔ 不替它补：

Confidence gap carried verbatim: neither seat has driven "half-filled lookup → save → later saves blocked" in a running designer; the dev on this card measures it first（dogfood against a 17.3.0 backend）。The urgency may move; the direction should not.

⇒ schema 层的 ACCEPT@17.2.0 / REFUSE@17.3.0 是跑出来的；"后续保存被永久挡住"是门的文档描述的失败类，⛔ 不是本卡驱动出来的。⇒ p2。

⭐ 提级条件：dogfood 一旦在 17.3.0 后端上复现出"保存后该对象再也存不了"，立即 p1。⇒ ⭐ 这也是派工的第一步（裁决明写 the dev on this card measures it first），⛔ 优先于写任何代码。

⛔ 三条执行禁令，来自裁决，⛔ 接手人不得变更

1. ⛔ 不得把 MetadataService.specKeyReference.test.ts 弱化成绿。裁决明写「The pin is not flipped green」——钉子必须陈述新的事实：半填 lookup 留在客户端；PUT body 永不携带 reference 为空的 lookup。⭐ 这一条最容易被"让测试过"的本能违反，故置顶。
2. ⛔ 不是选项 B（把不完整的关系字段从 PUT body 里剥掉、却仍显示"已保存"）—— 那会让设计器显示一个服务端从未收到的字段，即 objectstack#4001 关掉的静默丢弃形状。
3. ⛔ 不是选项 C（去上游 contest）。裁决对此给了理由，值得接手人读一遍以免重提：spec 的散文早已声明关系类型必须有 reference，17.3.0 是故意关掉一个 declared≠enforced 缺口；且草稿状态的正确归宿是客户端，不是元数据存储。

一句方向性的话，值得留在卡上

⭐ 裁决那句「the correct home for draft state is the client, not the metadata store」比本卡本身更通用。设计器把"用户还没填完的东西"写进服务端，本质上是把编辑会话状态存进了持久化契约。⇒ 修法不只是"别 PUT 这一个形状"，而是半成品不出客户端。⛔ 接手时请按这条写钉子，否则下一个必填键落地时会重演。

---

Generated by Claude Code