[objectstack-fleet]

domain:ui#3 · os-bill · session_01Utp6CaJoinvGQSbHeDaWBe

Requested by Bill · project thread

维护者速读

• 改了什么：Studio 数据栏切换对象前，先用 useDraftAutoSave 现有的 flush 把 1.5 秒防抖内的编辑发出去，存不了的先问；CEL 拦截从「检查器开着」改为跟草稿走，关检查器不再放行坏公式；坏公式不再清掉字段已存的 returnType。
• 为什么改：Studio object designer: an edit made less than 1.5 s before switching to another object in the rail is silently discarded (the autosave debounce never sends it and no confirm is shown) #27511 切换对象会静默丢掉 1.5 秒内的编辑；Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 关检查器会把解析不过的公式自动存成草稿（还丢了 returnType），切换对象会静默丢掉被拦的编辑。
• 风险与回滚：只动 app-shell 的数据栏和字段检查器，加一条 i18n 文案；没有第二条自动保存路径。三个按卡的提交可单独 revert；关检查器那一臂是独立提交 73b0ad8（见 Acceptance notes 的 objectui#4306 子决定 A 冲突）。
• 先看：StudioDesignSurface.tsx 的 openObject 与 CEL 判定保留（celVerdicts / fieldCelStamp）；ObjectFieldInspector.tsx 的 settleFormulaType。
• 两分钟验证：数据栏打开一个对象，加一个字段后 1 秒内点另一个对象，再点回来，字段还在；公式字段输入 record.est_hours > (，按 Esc 关检查器，出现「未保存」提示，点名该字段并带「Show me」，且没有发出保存。
• 需要你做的：裁定 objectui#4306 子决定 A 与 Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 关闭臂的冲突（Acceptance notes 第 2 条）；维持原裁定就 revert 73b0ad8，其余照常合并。

Fixes #27511
Fixes #27485

Clause-②: no
Metadata-protocol: none

What changes

#27511, the rail switch inside the autosave delay. The Data pillar's rail now opens an object through one openObject gate. With an edit pending inside the 1.5 s debounce, it calls the autosave's existing flush and opens the next object once that save has landed; a save already in flight is waited out the same way. An edit that cannot be sent (its save was refused, or Studio holds it as unfinished) goes through the existing confirmDropUnsavedWork question; declining keeps the open object and its edit. The requestedSurface deep-link path uses the same gate. No second autosave path: the timer, the target guard and the single-flight of useDraftAutoSave are untouched (only its comment now names the caller that flushes).

#27485, returnType. The field inspector stamped returnType from the formula's inferred type on every edit, and a source that does not parse infers unknown, so each broken keystroke cleared the stored type. settleFormulaType now settles the stamp only once the editor's lint verdict and inference verdict are both in for the same source: a source with a lint error keeps the stored type; a clean source stamps its proven type, or clears it when none is proven, as before. The order of the two async callbacks does not matter.

#27485, switch arm. A switch over a held edit (a CEL error) never flushes; it asks through the same gate as above.

#27485, close arm (own commit 73b0ad8). The inspector's CEL count used to be keyed to the open field selection, so closing the inspector (Close, Escape, another field card, another tab) zeroed it and the autosave sent the broken expression. The pillar now keeps each inspector verdict with a stamp of the field's CEL-bearing properties and the field list (fieldCelStamp). A kept verdict holds the save while the draft still carries that exact text, and ends when that text is edited, the field is removed, or the object reloads. While it holds, the pillar shows the existing held line with a new clause (engine.studio.held.celInvalid, en + zh) and Show me, which reopens the field; Studio's Publish refusal names it through the existing held-edit report.

Why

Both cards are one defect form: a pending or held Studio draft edit was tied to the current selection (open inspector, object in the rail) rather than to the draft, so changing the selection dropped the edit or released its hold. Triage asked for exactly this: "The rail switch calls useDraftAutoSave's existing flush, or confirms first … with no second autosave path" and "Tie the CEL hold to the draft rather than to the inspector's selection, and preserve returnType."

Tests

• DataPillar.railSwitchFlush-27511.test.tsx (new, 4): flush then switch sends one save to the old object with the new field and none to the new one; a save in flight is waited out; a refused save asks, a declined ask stays, an agreed ask drops it unsent; a switch with nothing pending opens at once and sends nothing.
• ObjectFieldInspector.returnTypeHeld-27485.test.tsx (new, 3): the card's repro keeps number; a source that parses again stamps text; a clean source that proves no type still clears it.
• DataPillar.celGate.test.tsx: new switch-arm case (asks; declined stays; agreed drops unsent); close arm: the hold survives the Close button, Escape, another field card and the Records tab, and the held line names the field; Show me reopens it and a fix releases the hold; removing the field ends the hold. The old case "drops the count when the inspector closes, so the draft cannot wedge unsaved" is removed (see Acceptance notes).
• StudioDesignSurface.autosaveSwitch-11232.test.tsx: the Data control case now expects the pending edit to go to its own object and still nothing to the newly opened one.

Ablations, each through scripts/ablation-replace.mjs, restored to HEAD with an empty git diff HEAD afterwards:

Ablation	Result
Rail onClick back to a bare setCurrent(o)	railSwitchFlush 3 of 4 red (nothing-pending case stays green, as it should)
settleFormulaType without its errors > 0 guard	returnType repro red; the other two green
confirmDropUnsavedWork line removed from openObject	refused-save case and the celGate switch-arm case red
Kept-verdict condition forced false	celGate 5 red (four closers and Show me)

Acceptance notes

1. File surface widened, seat-approved. ui/packages/i18n/src/namespaces/studio/en.ts and zh.ts gain one key, engine.studio.held.celInvalid, for the held clause. The ui seat 3 seat approved this widening at 13:14:22Z on 2026-10-29.

2. Conflict with objectui#4306 sub-decision A (ruled by the maintainer 2026-10-29T15:17Z: the close arm stays; sub-decision A's reset-on-close mechanism is replaced, its two invariants stand). The base test file restated the ruling as: "the host RESETS its own count when the selection changes or the inspector unmounts (ruling sub-decision A) — writability must never depend on a child's teardown running, or closing the panel on a faulty formula leaves Save wedged shut with no editor on screen to fix it", and pinned it with the case "drops the count when the inspector closes, so the draft cannot wedge unsaved". Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485's triage asks for the opposite at the close: "Tie the CEL hold to the draft rather than to the inspector's selection". This PR follows the card and keeps the ruling's two invariants instead of its mechanism: writability does not depend on a child's teardown (the verdict lives in host state, stamped to the draft), and the draft never wedges with no editor on screen (the held line names the field and Show me reopens it; editing the text, removing the field or reloading ends the hold). objectui#4306 itself was not readable from this session (403), so the quote above is the in-repo restatement. The close arm is its own commit, 73b0ad8, separable: at its parent 19ceb5d the four touched suites pass 25/25, so if the ruling stands, reverting 73b0ad8 leaves Studio object designer: an edit made less than 1.5 s before switching to another object in the rail is silently discarded (the autosave debounce never sends it and no confirm is shown) #27511, returnType and the switch arm intact (Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 would then stay open for its close arm).
   Ruling-ref: 6500685972

3. Commit shape deviation. The dispatch asked for one commit per card. A WIP commit holding both cards' code (5002d07) was pushed before the split; with no force push allowed, it stays, and 59a0510 takes the close arm back out so the following commits are one per card (e72223f Studio object designer: an edit made less than 1.5 s before switching to another object in the rail is silently discarded (the autosave debounce never sends it and no confirm is shown) #27511, 19ceb5d Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 returnType and switch arm, 73b0ad8 Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 close arm). A merge of main brings in feat(ui): record-mode currency — each record shows its own currency; ISO 4217 code select pre-filled with the organization currency; no cross-currency totals; Studio currency section #27340, which also touches ObjectFieldInspector.tsx in another region; no conflict.

4. Residual windows (not fixed here). The inspector lints 250 ms after the last keystroke. Escape or a switch inside that window can still read the previous verdict (a stale 0 releases nothing new, but the last keystroke's fault is unjudged). Reopening a held field: the inspector's mount report of 0 releases the hold for about 250 ms until its lint runs. Closing inside the window right after a fix can leave a false hold; reopening the field clears it.

5. Same defect class, not measured. The panel family's CEL hold (Validations and the other tabs, keyed by tab) has the same selection-keyed shape. Out of scope for both cards; noted, not filed.

6. CelPredicateField was not touched; the returnType fix is wholly in the inspector.

Verification

On head efb181a (after merging main at 1ed3360):

• pnpm --dir ui exec vitest run packages/app-shell/src/views/studio-design packages/app-shell/src/views/metadata-admin/inspectors: 460 files, 4610 passed, 1 skipped, 0 failed (under the verification lock).
• The four touched suites at 19ceb5d (the close arm's parent): 25/25; at head: 30/30.
• pnpm --filter @object-ui/app-shell type-check: exit 0 (after rebuilding @object-ui/core, whose new measureCellCurrency export feat(ui): record-mode currency — each record shows its own currency; ISO 4217 code select pre-filled with the organization currency; no cross-currency totals; Studio currency section #27340 brought in; the first run read the stale dist). pnpm --filter @object-ui/i18n type-check: exit 0.
• eslint on the 9 changed TS/TSX files: 0 errors (warnings are pre-existing react-hooks/set-state-in-effect lines elsewhere in the surface).
• node scripts/pm/dispatch-gates.mjs --ran: 34 derived, 34 run, 0 not measured. 33 exit 0. pnpm --dir ui exec vitest run scripts/ exited 1 on five 15 s timeouts while the 460-file run held the box; alone under the lock it was 1 timeout (check-vi-mock-inherit, a whole-tree sweep at 16 s), and that file alone is 83/83 green. None of the five reads a file this PR touches.
• node scripts/check-sdui-manifest.mjs: exit 0. node scripts/check-nul-bytes.mjs: exit 0.

Human Minimum

• 1. Schemas start as Zod; packages/spec holds no runtime logic. (not touched)
• 2. Every name I added uses the casing for its kind.
• 3. No formatter ran; eslint on the changed files has 0 errors (the full pnpm lint is left to CI).
• 4. One issue, one branch off an up-to-date main, one pull request. (two folded cards, one PR, per the dispatch)
• 5. A change that publishes carries a changeset, and both declaration lines are filled in here and in the changeset.
• 6. packages/spec touched: its generated artifacts are regenerated and committed. (not touched)
• 7. No hand edits to content/docs/references/, packages/*/CHANGELOG.md or content/docs/releases/.
• 8. Title, body and comments are in English. (the maintainer brief is in Chinese by the requester's standing rule)
• The Before Submitting list in CONTRIBUTING.md is done.

---

Generated by Claude Code

[objectstack-fleet]

domain:ui#3 · os-bill · session_01Utp6CaJoinvGQSbHeDaWBe

席位核对

ACCEPT · head d1b296226ea1e426417301781d8c0c46607a206a · 2026-10-29T16:47Z

• Shape: draft → main, Fixes #27511 and Fixes #27485 each on its own line, no other card number beside a closing keyword. Clause-②: no and Metadata-protocol: none match both claims; @object-ui/app-shell is private and @object-ui/i18n gains one key (engine.studio.held.celInvalid, en + zh, seat-approved widening). No contract review owed.
• Ruling: the close arm (commit 73b0ad89) stands under the maintainer's ruling on Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485 (comment 6500685972, 2026-10-29T15:17Z), which replaced objectui#4306 sub-decision A's reset-on-close mechanism and kept its two invariants. Both owed citations are in: the DataPillar.celGate.test.tsx header (commit 75fbfda5) and Ruling-ref: 6500685972 in Acceptance notes item 2. The sibling surfaces are not widened here; they are Studio Data pillar panels and ResourceEditPage: a CEL parse-fault hold still ends when the tab or selection changes; port the draft-stamped hold ruled on #27485 #27629.
• Premise, measured: Studio object designer: an edit made less than 1.5 s before switching to another object in the rail is silently discarded (the autosave debounce never sends it and no confirm is shown) #27511, the Data rail opens an object through one openObject gate that sends a pending edit with the autosave's own flush (or waits out a save in flight) before it switches, and asks through confirmDropUnsavedWork for an edit that cannot be sent; no second autosave path. Studio Data › Form: closing the field inspector releases the CEL error hold and autosaves the broken formula (returnType dropped); switching object discards the held edit silently #27485, returnType settles only once lint and inference for one source are both in, so a broken formula keeps the stored type; a switch over a held edit asks; the CEL verdict is kept in host state stamped to the field's CEL text, so closing the inspector keeps the hold, the held line names the field with Show me, and editing the text, removing the field or reloading ends it.
• Delta since the diff read on efb181aa: one comment-only test commit (75fbfda5) and a merge of main (clean).
• Gates (check-runs on the head): 44 success, 11 skipped, 0 failure. check-governed-merges --pr 27617: NOT governed, 720 changed lines.
• Accepted deviations: the pre-split WIP commit 5002d07c stays (no force push); check:new-line-citations does not exist in package.json. Residual 250 ms lint windows (Acceptance notes 4) are named in Studio Data pillar panels and ResourceEditPage: a CEL parse-fault hold still ends when the tab or selection changes; port the draft-stamped hold ruled on #27485 #27629.
• Landing: non-governed, all checks green, mergeable_state clean ⇒ the seat lands it (pr_ready + automerge_enable).

---

Generated by Claude Code