fix(app-shell): an edit a Studio pillar holds unsent guards departure like a dirty pillar - #26839

Requested by Warren · project thread

Fixes #26705

Clause-②: no
Metadata-protocol: none

维护者速读

• 改了什么:Studio 把「暂存未发的编辑」(还不能保存的字段、步骤、组件)也算作未保存工作;离开支柱、刷新页面、在左侧栏切换到别的对象/流程/页面时,会先弹出与 Access 相同的确认框。
• 为什么改:之前 Data、Automations、Interfaces 支柱暂存的编辑只拦发布,不拦离开,一次误点就静默丢失;分诊裁决是「confirm 加 beforeunload,共用一个守卫」。
• 风险与回滚:只多问一句,不改保存与发布逻辑;暂存与否沿用已有判定。有问题直接 revert 本 PR。
• 先看:StudioDesignSurface.tsx 里的 confirmDropUnsavedWork 和 unsavedPillarWork,然后三处左侧栏 onClick / onPick 的守卫行,最后 StudioDesignSurface.heldDataLeave-26705.test.tsx。
• 两分钟验证:pnpm --dir ui exec vitest run packages/app-shell/src/views/studio-design/StudioDesignSurface.heldDataLeave-26705.test.tsx,预期 4 个测试全绿。
• 需要你做的:无,席位落地。

What changes

Base 607d3bb1, then origin/main 55b5042e merged in (no conflict; nothing it brought touches these files). Then d33bfcf7: the zh catalogue preload in two metadata-admin suites (test-only, scope widened by the seat, Claim 6452362760). One shared guard, read by every departure that drops a pillar's buffer:

• confirmDropUnsavedWork(unsaved, locale), a module function beside useHeldEditReport: the one question (engine.edit.unsavedLeaveConfirm, the text Access already asks) or true when nothing would be lost. Access's own confirmDiscardEdits now calls it too, so the question has one spelling.
• Surface: heldEdit moves up beside pillarDirty, and unsavedPillarWork = pillarDirty || heldEdit !== null replaces pillarDirty in confirmLeavePillar, the beforeunload effect and useRegisterUnsavedEdits. That covers the pillar links (the shell seam's link guard asks while registered), Home, the package switcher, the cross-pillar ?surface= request, Create app, the shell's own departures and reload. Pillars already take their held report back on unmount, so a confirmed leave clears the guard.
• Rail item switch (a pillar-local path the surface cannot see; another item's load replaces the one buffer the held edit lives in): the Data object rail asks on objHeld, the Automations flow rail on flowHeld, the Interfaces nav rail on leafHeld. Re-picking the open item (same object, same flow, an entry for the same leaf) drops nothing and asks nothing, mirroring Access's rail.
• Unchanged by design: how a held edit is judged and saved, and Publish's refusal while one is held.

Why

Triage first grade on the card, verbatim: "Guard departure the way the Access pillar and the Interfaces navigation already do (confirm plus beforeunload), with one shared guard." A held edit lives only in the pillar's buffer, so before this change one stray click dropped a half-added field, step or widget that the strip had just promised was "kept here".

Tests

Three new files, one per pillar, each through the real StudioDesignSurface on the real route shape (the held edit is produced the way an author does it: a Picklist with no option, a Notify step with no Title, a Metric widget with no dataset):

• StudioDesignSurface.heldDataLeave-26705.test.tsx, StudioDesignSurface.heldFlowLeave-26705.test.tsx, StudioDesignSurface.heldWidgetLeave-26705.test.tsx, four cases each: leaving the pillar asks (Cancel keeps the held edit, OK leaves, and walking back asks nothing); the rail switch asks (Cancel keeps it, re-picking the open item asks nothing, OK opens the other item and the hold is gone); a reload asks only once the edit is held; and the CONTROL: with nothing held, the rail, the reload and the pillar link never ask.

Acceptance notes

• Not guarded, measured and left as they were: creating a new object, flow or Interfaces item while another item's edit is held opens the new item (the create has already been written), and a same-pillar ?surface= request from the pending-changes sheet opens the requested item without asking. Both drop the held edit the same way a rail switch did. The card names leaving the pillar, the rail switch and reload; these two are left for a follow-up if wanted.
• Browser Back/Forward is still not vetoed, as the shell seam's docblock already states for every page.
• @object-ui/app-shell is private: true, so no changeset.

Verification

Gates derived by node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack: 24 families, all run, reconciled with --ran (24 run, 0 NOT-MEASURED). pnpm --dir ui exec vitest run scripts/ reported one timeout in check-vi-mock-inherit.test.ts (15 s, shared box under load) and passed 83/83 when that file was re-run alone. pnpm check:new-line-citations (ui): 0 new citations.

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; pnpm lint passes. (eslint on the four files: 0 errors; StudioDesignSurface.tsx carries the same 16 warnings as on the base, the three tests none.)
• 4. One issue, one branch off an up-to-date main, one pull request.
• 5. A change that publishes carries a changeset, and both declaration lines are filled in here and in the changeset. (Publishes nothing: @object-ui/app-shell is private.)
• 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, apart from the maintainer brief above.
• The Before Submitting list in CONTRIBUTING.md is done.

Generated by Claude Code