fix(deps): triage the 14 rc.6 spec-symbol collisions and refresh the console parity exemptions (#4167)

[yinlianghui]

Fixes #4167.

⛔ Stacked on #4169 (claude/issue-7100-global-nav-studio-retire), not on main. This card's premise — the @objectstack/spec family resolved at 17.0.0-rc.6 — exists only on that branch, so CI here runs against that base and inherits its state. It retargets automatically if #4169's base evolves.

Both of the card's items are complete and their gates are green:

#	Item	Bar	State
1	13 check:spec-symbols name collisions, per-symbol triage	check:spec-symbols green	✅ (14, not 13 — see below)
2	12 stale parity exemptions + 3 record-picker inputs	registry-inputs-spec-parity green	✅ 51/51

There is also a third finding that changes how this PR's CI must be read, in §3. It is not mine to fix and it is measured, not assumed.

---

1. The collisions — 14, and each triaged on its own

The card named 13. The guard reports 14: @object-ui/types:GlobalFilterSchema appeared during the rc.6 bump and is an artifact of it, so it belongs here rather than to a future card. Details in its row.

Verdicts follow the objectstack#4115 method — bind, derive, rename, or ALLOW with a written reason — and the answer genuinely differs per symbol. Nothing was blanket-renamed.

#	package	symbol	what the spec's same-named export is	verdict
1	app-shell	MetadataState	metadata item LIFECYCLE state — 'draft' | 'active' | 'deprecated' | 'archived'	rename → MetadataCacheState
2	app-shell	resolveI18nLabel	resolver for the INLINE per-locale map, against a BCP-47 locale	rename → resolveKeyedI18nLabel
3	core	DATE_RANGE_PRESETS	the 13-preset dashboard vocabulary, extracted by objectstack#4614	re-export (+ satisfies completeness pin)
4	fields	AddressValue	z.input< typeof AddressValueSchema > — seven parts	derive (import + re-export)
5	layout	BreakpointColumnMap	identical $strict six-key map	re-export
6	layout	BreakpointOrderMap	identical $strict six-key map	re-export
7	react	MetadataState	as row 1	rename → MetadataCacheState
8	react	NavigationMode	z.input< typeof NavigationModeSchema > — the same 7-member union	derive (+ both-directions pin)
9	react	resolveI18nLabel	as row 2	rename → resolveKeyedI18nLabel
10	types	DateRangePreset	13 HISTORICAL windows; this one adds 8 FUTURE ones	rename → FilterBuilderDateRangePreset
11	types	KanbanConfig	identical $strict 3-key config	re-export
12	types	CalendarConfig	identical $strict 4-key config	re-export
13	types	GanttConfig	17 keys; the copy declared 6	derive + one declared local key
14	types	GlobalFilterSchema	the authored global-filter schema	ALLOW, with the reason

The three renames, and why each is a rename rather than an ALLOW

The AuthProvider / AuthProviderConfig split from batch 5 is the test applied: a JSX element cannot be mistaken for a zod enum, so it was ALLOWed; AuthProviderConfig names the spec's OAuth registration shape in the same domain and the same words, so it was renamed. All three below fail that test the same way.

• MetadataState — an object of five arrays vs a string union, mutually unassignable, so nothing would have caught the misreading at the point it was made. Cache because that is what it is: MetadataProvider is a TTL cache and these are its materialized contents. Six references repo-wide.
• resolveI18nLabel — the one where the collision had already started costing something, and the only one that is dangerous rather than untidy. rc.6 widened I18nLabel from string to string | Record< string, string >, so the same authored value now reaches either resolver, and each answers wrongly and silently for the other's input: the keyed one returns undefined for { en: 'Owner' } (no key, no defaultValue); the spec's reads key / defaultValue / params as locale tags and picks one as display text. retire(studio): drop the global_nav designer surfaces with the rc.6 spec bump (objectstack#7100) #4169 met this and aliased the spec's import as resolveInlineI18nLabel in five files, with hand-written comments at two of them — a review convention, which is exactly what objectstack#4115 exists to replace with a rule. Keyed is now the counterpart of that Inline. A fourth copy in apps/console/src/utils.ts is renamed with them; it is outside the guard's scan surface (packages/*/src only) but plants the identical premise.
• DateRangePreset — two vocabularies that are not even nested by accident. The spec's 13 are historical windows and its superRefine on GlobalFilterSchema rejects everything else by name; this one adds 8 future windows (tomorrow, next_week, …) for the FilterBuilderConfig surface, where a forward-looking range is the point. So a reader who took the spec's 13 off this name would have been told 8 windows exist that the dashboard schema rejects, and vice versa. FilterBuilderCondition (batch 5) named its sibling the same way. Zero internal consumers outside DateRangeFilter in the same file, so no call-site churn.

No tripwire tests were added for the new names, deliberately. check:spec-symbols already is that tripwire: the day the spec publishes MetadataCacheState, resolveKeyedI18nLabel or FilterBuilderDateRangePreset, the guard reds naming it. The ObjectUiLocal…-plus-tripwire form in #4115 exists for symbols that stay in the spec's own vocabulary; these three left it.

Four copies were losing information, not just duplicating it

• GanttConfig declared six keys and called itself canonical; rc.6 declares seventeen. The eleven it never mentioned — parentField, typeField, baselineStartField, baselineEndField, groupByField, resourceView, assigneeField, effortField, capacity, quickFilters, autoZoomToFilter — are all read today by plugin-gantt/src/ObjectGantt.tsx, through a local GanttConfigEx intersection that existed only because this type did not carry them. timeSegments (shift segmentation) is the one genuinely local key and stays declared as objectui's; GanttConfigSchema is $loose upstream, so it is legal metadata rather than a second dialect — the spec's decision for this vocabulary, not a widening taken here.
• GanttConfig.tooltipFields carried the comment "not part of the upstream GanttConfigSchema". It is, as of rc.6.
• AddressValue declared five of seven parts (countryCode, formatted missing) under a comment already claiming to be "the part names of AddressSchema". The widget still renders five inputs — formatted is derived and countryCode is not a part a human types — but the type no longer asserts the platform cannot store them, and the { ...address } write-through now says so.
• DATE_RANGE_PRESETS was Object.keys(PRESET_RANGES). The spec's own doc comment names this module as one of the three copies objectstack#4614 extracted the vocabulary to collapse. It is now the spec's array by reference (toBe, not toEqual — a faithful copy passes every value comparison, which is why test(plugin-charts): serve ObjectChart's option-color probe from a double, not the network #4115 makes reference identity the test), and the local date-macro bounds table is pinned complete against it with satisfies.

That satisfies is load-bearing, not decoration. With the list coming from the spec and the bounds staying local, the two can drift in exactly the direction the spec's comment names as the failure mode — a preset the schema knows and the bounds table lacks "validates clean and then resolves to nothing". The annotation form const PRESET_RANGES: Record< DateRangePreset, … > would not catch it: a string index signature satisfies every literal key, so the check passes vacuously.

GlobalFilterSchema — the 14th, and why it is an ALLOW

It collided because of #4169's own forced rewrite. rc.6 put a refinement on the spec's schema, which makes .extend() throw at module load and types every override as never under .safeExtend(); the only remaining spelling is a .shape spread. That moved a derivation the guard could see (a call target) into an object literal it deliberately does not descend into — the guard's header says why: a spread inside z.object({…}) is indistinguishable, to an AST pass, from a hand-typed shape.

So the dialect did not change; its spelling did. The precedent is exact and one entry above it in the same list: SelectOptionSchema composes the identical way (...SpecSelectOptionSchema.shape) and is ALLOWed for the identical reason. Nothing about #4165 is encoded here — the refinement is still not carried, the standstill is still pinned, and the entry says which issue owns the decision.

---

2. The parity exemptions

registry-inputs-spec-parity: 49 tests, 2 failing → 51 tests, 0 failing.

2a. All twelve off-spec exemptions deleted — the list is now empty

Not eleven or thirteen: the carries no stale exemption test named all twelve at once, and each was verified against the resolved ComponentPropsMap[type].shape at this pin rather than from the issues' wording.

• objectstack#6776 — page:header now declares recordChrome / showStar / showCopyId; page:accordion declares variant; page:tabs declares tabStyle. Five keys the renderers had read all along are contract now.
• objectstack#5775 / PR objectstack#6281 — element:record_picker declares labelField / valueField / label; page:card declares children (replacing the retired body); page:section / page:footer / page:sidebar carry the shared PageContainerProps. These seven are the objectui 跟进(#5775):page:card 的 registry inputs 仍把已退役的 body 发布为设计器槽;page:section/footer/sidebar 三个注册项没有 inputs,设计器无法授权 children #4027 stale-pin set, which predicted its own deletion ("Delete this entry when the pin moves") and got it.

page:tabs is the interesting one: the spec declares both tabStyle and type, so the carrier collision the deleted entry wrote up was resolved upstream by declaring the alias rather than renaming. That is why page:tabs.type stays in UNPUBLISHED_EXEMPTIONS (spec-declared, unpublishable in a flat carrier) while tabStyle needs no cover at all.

The map is kept declared-and-empty rather than removed, with a tombstone test — otherwise "empty because everything resolved" and "deleted by accident along with the divergences it covered" look identical and every forward assertion passes either way. That is the misdiagnosis #4027 records for the SPEC_SHAPE_EMPTY_ON_THE_PIN carve-out, which also left an assertion behind rather than just a comment.

Note that the UNPUBLISHED_EXEMPTIONS entries did not go stale, exactly as their own reasons predicted: ADR-0087 D2 retirement replaces an entry with z.never().optional() rather than deleting it, so displayField / searchFields / multiple / page:card.body are still "declared" to this gate. They resolve via objectui#3809, not via the pin.

2b. Three record-picker inputs declared

sort, limit and emptyText were already read by the renderer and declared by the contract, but absent from inputs — so renderers/layout/page.tsx, which builds the JSX-page compiler's prop whitelist from getKnownTypes() plus these inputs, drew an unknown-prop warning on keys the renderer then honoured. objectui#3407's shape, and the same repair #3808 and #3830 made before it. The red demanding them was predicted in writing by the exemption covering the retired trio, which called it "correct and wanted".

Each description documents renderer behaviour rather than restating the schema, per the #3830 / #4041 shape — because that is the half an author cannot read off the contract:

• sort — 'array' is the spec's shape, verified by safeParse: [{ field, order }] parses and the terse "name asc" does not. Overridden outright by a node-level dataSource (dataSource.sort ?? sort), not merged.
• limit — default 50 comes from the renderer (record-picker.tsx:107), not the schema; a record outside the limit cannot be picked at all, with nothing in the control saying more exist.
• emptyText — published as string against a contract of string | Record< string, string >. rc.6 widened it to I18nLabel, and this renderer passes the value straight into a text node with no locale resolution, so only the plain-string form renders. The description says so rather than advertising a shape the renderer drops — the narrowed-type treatment objectui#3832 describes, with the render-site gap belonging to §3.

Both sets are pinned by name as DECLARED, not merely "not failing", because the derived assertions would go green just as readily if the keys had been exempted instead — which is the move #3808 exists to rule out.

---

3. ⚠️ How to read this PR's CI — 13 type errors that are not mine

First, what actually runs here: almost nothing, and that is structural. Every gate workflow in .github/workflows/ filters pull_request to branches: [main, develop] — ci.yml, lint.yml, changeset-guard.yml, changeset-presence.yml, control-bytes.yml, skills-paths.yml, docs-links.yml, performance-budget.yml, live-e2e.yml, all of them. This PR's base is a feature branch, so none of them trigger. Its checks converge at two, both success (Auto Label PRs, Vercel Preview Comments), with nothing queued and nothing red. A green tick here is therefore evidence of very little, which is why every gate below was run locally instead and its real output quoted.

Second, what will run when the chain reaches main. check:spec-symbols is green, but the TypeScript Type Check job it lives in will be red on the merged chain, and this card cannot make it green. The issue's premise that this gate is "currently the main red" of that job is accurate about check:spec-symbols and incomplete about the job.

turbo run type-check --continue reports 13 errors in 5 packages, and every one is the same fact: rc.6 widened I18nLabel from string to string | Record< string, string >, and these sites take it into a ReactNode, an input value, or objectui's keyed resolver.

layout           src/AppSchemaRenderer.tsx:169,173      NavigationArea.label
plugin-list      src/components/TabBar.tsx:140,200,216  ViewTab.label
plugin-dashboard src/DashboardGridLayout.tsx:398,399    DashboardWidget.title
plugin-designer  src/DashboardEditor.tsx:143,231,341    widget title editing
app-shell        src/utils/resolveActionParams.test.ts:245,345,358   ActionParam.label

#4169 repaired eight such render sites; these are the ones its pass did not reach, and turbo run build cannot see them — these packages build through vite/rolldown, so only type-check compiles them. That is why a bump reporting a green 44/44 build and 14,609 passing tests still has them.

This is objectui#4163 part 1 verbatim ("the I18nLabel audit"), which my scope guard fences off, and it is not mechanical: @object-ui/layout deliberately has no i18n dependency ("enables convention-based i18n auto-resolution without coupling the layout package to i18n"), so choosing where its locale comes from is a design call in that package, not a find-and-replace. Left alone, with the inventory posted to #4163 so it is a named worklist rather than an open-ended audit.

Measured, not assumed

A second worktree at the untouched base (db9dad961) was type-checked with the identical command:

	base db9dad961	this branch
tasks	71 successful / 78	73 successful / 78
failing packages	7	5
type errors	16	13

The 13 are byte-identical to 13 of the base's 16. This branch introduces none and removes three.

The three it removes

All pre-existing, all rc.6 adaptation in this card's own family (not #4163's), and all found only by running type-check — which #4169 did not. Two are one-line re-points; the third is a member-list update. They are included because without the first, turbo stops at @object-ui/types and never reaches a single package this PR touches, so verifying my own work required them.

• types/…/page-nav-misc-spec-parity.test.ts — ThemeInput no longer exists (the …Input alias retirement, item 1 of the card's own rc.6 inventory). Re-pointed by side, not by name: ThemeInput → Theme and Theme → ThemeParsed, because rc.6 moved the bare name onto the input side. Following the names would have made the second pin compare the authoring side against itself and go green on nothing — the same swap retire(studio): drop the global_nav designer surfaces with the rc.6 spec bump (objectstack#7100) #4169 avoided in spec-report.ts.
• data-objectstack/spec-symbol-batch6.test.ts — rc.6 added a third DroppedFieldsEvent.reason arm, primary_key. This pin fired exactly as designed, and its firing is the point: the hand copy it replaced had widened reason to bare string, and under that the arm would have arrived silently. The member list is updated; the consumer gap is objectui#3935, which was filed before the arm shipped and named this precise moment as its activation condition. emitWriteWarning still branches binary, so a primary_key drop is currently narrated as "Read-only" — false, and the same mis-wording ADR-0056 的 Consequences 自相矛盾:「Capabilities 在 Setup 里作者化」与 P2/P4「在 Studio 设计、Setup 里没有 facet 可编辑」冲突 #3794 fixed for readonly_when. Not fixed here: it needs an exhaustive map plus a third wording in both locale tables, and writeWarningToast branches on DroppedFieldsEvent.reason with a binary ternary, so every future reason is labelled "Read-only" #3935 already prescribes the shape.

3b. And one inherited test failure, in a package this PR does not touch

pnpm exec vitest run is 14,618 / 14,620 with one failure:
plugin-detail/src/__tests__/recordDetailsInputs.spec-parity.test.ts → "the spec really takes OBJECT sections — the id-list spelling is gone, not unioned in".

Proved to be a pure spec fact, with a probe that runs no objectui code at all:

RecordDetailsProps.safeParse({ layout: 'custom', sections: [{ name: 'contact_info', … }] })
// success: false — issue at path ["layout"], code invalid_type, expected "never"
// "`record:details` property `layout` was removed in @objectstack/spec 17.0.0 (#6946, ADR-0087 D2)"

rc.6 turned record:details.layout into a D2 tombstone, and the fixture carries layout: 'custom'. The test's own premise is untouched — the sections object form still parses, member keys name / label / columns / fields.

The interesting half is that the sibling assertion in the same test has gone VACUOUS. It reads expect(idList.success).toBe(false) to prove the retired id-list spelling of sections is rejected — and that is now green because of layout, not because of sections. A rejection assertion passing for the wrong reason is worse than a failing one, so this needs the fixture repaired and the rejection re-pointed at the key it names.

Not fixed here: objectui#3818 already owns record:details.layout (it is the card whose upstream resolution objectstack#6946's tombstone is), the renderer's own layout read point needs the same audit, and none of it is in this card. Posted to #3818 rather than filed as a twin.

---

Reverse verification

Three, each predicted red before running and each red for the named reason. The second and third are the ones that matter: in both, the repo's own derived assertion and the pin this PR adds fire together, which is what shows the new pins are not vacuous.

1. Un-fix one collision — git checkout origin/claude/issue-7100-global-nav-studio-retire -- packages/layout/src/ResponsiveGrid.tsx:

❌  a spec-named symbol is hand-written, not derived:
    • @object-ui/layout declares 2 spec-named symbols the spec already owns:
        interface `BreakpointColumnMap`  packages/layout/src/ResponsiveGrid.tsx:16
        interface `BreakpointOrderMap`   packages/layout/src/ResponsiveGrid.tsx:29

2. Restore one deleted exemption (page:header.recordChrome):

× carries no stale exemption — a declared key must lose its entry
× the twelve rc.6-obsoleted off-spec exemptions are gone, and their keys are contract now
  AssertionError: expected [ 'page:header.recordChrome' ] to deeply equal []
  AssertionError: expected [ 'page:header.recordChrome' ] to not include 'page:header.recordChrome'
Tests  2 failed | 49 passed (51)

3. Withdraw one newly-declared input (emptyText):

× element:record_picker publishes every top-level key its spec props schema declares
× the three keys the rc.6 bump added to element:record_picker are discoverable
  AssertionError: expected [ 'emptyText' ] to deeply equal []
  AssertionError: element:record_picker does not publish emptyText: expected [ 'object', 'filter', …(6) ] to include 'emptyText'
Tests  2 failed | 49 passed (51)

Case 3 is the one the pin exists for: the derived assertion alone would have gone green again if emptyText had been added to UNPUBLISHED_EXEMPTIONS instead of declared. The named pin refuses that, which is the move objectui#3808 exists to rule out.

---

Verification

Run locally, because §3 explains that this PR's own CI runs no gate workflow — its two checks are Auto Label PRs and Vercel Preview Comments, both green, and neither judges any of this.

• node scripts/check-spec-symbol-derivation.mjs — ✅ 1224 files scanned against 4834 spec export names; 14 declared dialects, 3 untriaged collisions in 1 packages (the 3 are the pre-existing @object-ui/types DEBT ledger, untouched)
• pnpm exec turbo run build --concurrency=2 — 44/44 successful
• pnpm exec turbo run type-check --concurrency=2 --continue — 73/78, the 5 failures being §3's inherited I18nLabel set
• pnpm exec vitest run (whole repo) — 14,618 passed / 14,620, 1 skipped, 1 failed: plugin-detail/…/recordDetailsInputs.spec-parity.test.ts, inherited and unrelated — see §3b
• apps/console/…/registry-inputs-spec-parity.test.ts — 51 passed (was 49 with 2 failing)
• check-changeset-fixed ✅ · check-changeset-no-major ✅ · check-changeset-presence ✅ (3 changesets) · check-control-bytes ✅
• turbo run lint over the 10 touched packages — 10/10 successful, 0 errors (warnings pre-existing)

Changesets: minor for the seven packages with changed public exports (major is banned by check-changeset-no-major — objectui's own breaking changes ship as minor with the break spelled out, since its major is pinned to @objectstack's), plus patch for plugin-designer, which only follows the rename at its call sites.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectui	Ignored		Aug 10, 2026 7:07pm

[yinlianghui]

Train note (objectui seat PM, session session_017Qqyix2QcnpUC9XeYVDzx3): auto-merge disarmed on purpose. Retargeting this stacked PR to main did not trigger CI (base-edit fires no synchronize), and the merge queue would have bounced it anyway: this tree lacks #4208's resolveActionParams typetest fixes and the post-train quick-reference-current-release-4143 doc test, both red on the combined tree while main is already red from the #4169 landing.

This PR's content lands via PR #4208 instead — that branch is stacked on this one and already contains every commit here; the train-finisher agent is adding the rc.6 residual fixes on top, retargeting #4208 to main, and landing the tail as one squash that closes #4167 (this PR's issue) and #4163 part 1 together. Once it merges, this PR will show an empty diff and will be closed without merge — the work is not lost, it is riding the tail car. Main-health context: fastest single landing back to green.

---

Generated by Claude Code