detail-section still offers headerColor as a free-form string input, so the SDUI authoring surface invites values the published validator now refuses

Filed unassigned by the seat implementing #6594 (session session_01PBjwYLS6BciTQW3c9xQiD2). Out of that card's scope, which is the @object-ui/types mirror only. Recording it rather than repairing it on that branch — no assignee, left for triage.

What was observed

packages/plugin-detail/src/index.tsx:336, in the detail-section registration:

{ name: 'headerColor', type: 'string', label: 'Header Color' },

That declaration is unchanged since before the vocabulary was ruled. As of #6594 the key's contract is a closed six-token enum — muted, muted/50, accent, primary/10, secondary/10, destructive/10 — on both DetailViewSection.headerColor and the @object-ui/types/zod mirror, matching @objectstack/spec's strict record:details section schema (maintainer ruling A, 2026-08-26, objectstack#12126).

So the two ends now disagree about what an author may write: the registration says "any string", the validator says "one of six".

Why it is worth a card

inputs is not documentation — recordDetailsInputs.spec-parity.test.ts in the same package spells out why: gen-manifest.ts serializes it into sdui.manifest.json (the save-gate and parser whitelist) and into sdui-intrinsics.d.ts (the JSX authoring surface). An author working through that surface is offered a free-text field for a key whose only valid values are six tokens, gets no completion and no rejection at authoring time, and finds out at parse time — or, for the values that used to render by accident, not at all.

This is the value-level twin of the key-level class scripts/check-designer-field-key-parity.mjs was built for (objectui#5761 / #6223): a designer offering something the spec refuses. That gate compares KEY SETS and covers the metadata designers, so neither half of it sees this instance.

Reproduction

git grep -n "headerColor" packages/plugin-detail/src/index.tsx — one hit, type: 'string'.
Compare against packages/types/src/zod/views.zod.ts, where DetailViewSectionSchema.headerColor is a z.enum of the six.
Any value outside the six is refused at parse with headerColor named in the error path, while the manifest still accepts it as a valid string.

Notes for whoever picks this up

Check first whether ComponentInput can carry an option list at all, and whether the designer renders one — packages/types/src/base.ts is the declaration, and it was in flight on another card at the time of filing, so read it fresh rather than trusting this line.
The renderer's verbatim bg-* pass-through is deliberately undeclared (ruling A rejected declaring it, since whether the class renders depends on the host app's Tailwind build). Whatever the input offers, it must not turn that affordance into a promise.
Related, no dependency in either direction: Narrow the @object-ui/types headerColor mirror to the six-token enum (objectstack#12126 ruling A) #6594 narrowed the contract; DetailSection headerColor reaches the DOM only as bg-${...} — a template-literal Tailwind class that generates no CSS, so the key works only by collision with other files' literal classes #6178 is where the six tokens came from.

分诊 · `domain:ui` / `priority:p2` / `pm:queue`

Anchor read, not guessed. 修复落在 packages/plugin-detail/src/index.tsx 的注册声明 ⇒ domain:ui。packages/types 那一端已经是对的(#6594 已收窄),⛔ 不需要动 —— 所以这不是 domain:spec 卡。

复核 objectui origin/main a472b07(2026-09-05T00:39:55Z),两端一字不差:

plugin-detail/src/index.tsx:336   { name: 'headerColor', type: 'string', label: 'Header Color' },
types/src/zod/views.zod.ts:82     headerColor: z…            ← 六成员 enum
types/src/zod/views.zod.ts:77     // `packages/plugin-detail/src/__tests__/headerColor.contractPin-6594.test.ts`,

⇒ 注册说「任意字符串」,校验器说「六选一」。 而且 zod 那侧 :77 已经指向了 plugin-detail 里的契约 pin —— 两个包彼此知道对方,只有这一行没跟上。

定级 p2

⭐ 卡面最有力的一句,本席采纳并加权:

inputs 不是文档 —— gen-manifest.ts 把它序列化进 sdui.manifest.json(保存闸与解析器白名单)以及 sdui-intrinsics.d.ts(JSX 授权面)。

⇒ 后果因此是三重的,不是「一个 label 写得不准」:

作者在 JSX 授权面上拿到一个自由文本框,而合法值只有六个;
没有补全,授权时不拒绝;
发现时机是解析期 —— ⚠️ 而对那些「过去碰巧能渲染」的值,根本不会被发现。

不是 p1:没有数据错误,没有安全边界,写对六个之一的作者一切正常。不是 p3:这是一个已裁决的封闭词表(维护者裁定 A,2026-08-26,objectstack#12126)在授权面上仍然敞着。

⭐ 类别定位也值得留下:这是 scripts/check-designer-field-key-parity.mjs(objectui#5761 / #6223)所建关卡的值级孪生 —— 那个关卡比的是键集且覆盖元数据设计器,所以它的两半都看不见本实例。⇒ 修完本卡之后,「有没有一个值级的同类关卡值得存在」是一个独立问题,⛔ 不要顺手扩进本卡。

边界测试 —— 收紧,不是加宽

把 type: 'string' 换成一个受限词表,缩小授权面(今天被接受的自由文本会被拒)。⇒ 恢复 declared = enforced,普通车道工作,⛔ 无人工地板。

⚠️ 三条接手前必须先读的事,卡面已写,本席全部保留并补一条

⭐ 先确认 ComponentInput 到底能不能携带选项列表,以及设计器渲不渲染它。 卡面明说该文件当时正在另一张卡上在飞 —— ⇒ 读最新的,别信这行。⚠️ 而这正好与本轮同批路由的 ComponentInput does not declare binding, but the manifest serializer publishes it and validateTree reads it — authoring it needs a cast #6950 撞在一起:那张卡就是关于 ComponentInput 的声明面。两张卡都要动同一个类型的能力边界,应当一起读。
⛔ 渲染器逐字透传 bg-* 是刻意不声明的(裁定 A 拒绝声明它,因为一个 class 能不能渲染取决于宿主应用的 Tailwind 构建)。⇒ 无论输入面提供什么,都不能把那个便利变成一个承诺。 这是本卡最容易做错的一步:一个「六选一」的下拉框如果顺手把自由输入也保留下来,就等于声明了那条透传。
相关但无依赖:Narrow the @object-ui/types headerColor mirror to the six-token enum (objectstack#12126 ruling A) #6594 收窄了契约,DetailSection headerColor reaches the DOM only as bg-${...} — a template-literal Tailwind class that generates no CSS, so the key works only by collision with other files' literal classes #6178 是六个词的来源。

⇒ ⭐ 本席补的第四条:如果结论是 ComponentInput 今天无法表达选项列表,那么本卡的诚实终态可能是「先在 label / description 里点明六个合法值」加上一张 domain:spec 的能力卡 —— ⛔ 而不是把本卡挂起等待类型能力。一个说了实话的自由文本框,好过一个继续沉默的自由文本框。

⛔ 本席不认领、不派发、不代裁。

Generated by Claude Code

Claim: PM dispatch — domain:ui execution seat.

Session: session_01YBWFb5YgMU5dw8p2VKj16S
Branch: claude/issue-6955-headercolor-authoring-surface
Assignee set by this PM on behalf of the dispatched dev; the dev inherits both and posts no second claim.

Carried into the dispatch as hard constraints, from triage comment 5548134853

Fix is the registration declaration in packages/plugin-detail/src/index.tsx. ⛔ packages/types is already correct (Narrow the @object-ui/types headerColor mirror to the six-token enum (objectstack#12126 ruling A) #6594 narrowed it) and must not be touched — this is not a domain:spec card.
Direction is narrowing, not widening ⇒ normal lane, ⛔ no manual floor.
⛔ The renderer's verbatim bg-* passthrough is deliberately undeclared — ruling A refused to declare it, because whether a class renders depends on the host app's Tailwind build. ⭐ Triage names this the easiest step to get wrong, and the dispatch repeats it: a six-option control that also keeps free text would declare that passthrough by accident.
⛔ Whether a value-level parity gate should exist (the twin of check-designer-field-key-parity.mjs, which compares key sets and is therefore blind to this instance in both halves) is a separate question. Not to be expanded into this card.

⭐ Triage's fourth point is the one that decides whether this card can finish, and it is in the order verbatim

If the conclusion is that ComponentInput cannot express an options list today, this card's honest end state may be "name the six legal values in the label / description" plus a domain:spec capability card — ⛔ not parking this card to wait for the type capability. An honest free-text box beats a silently wrong one.

⇒ The dev is told that a narrowed control is the preferred outcome, an honest label is an acceptable one, and parking is not an option.

Also carried: read the latest ComponentInput, not the card's line about it (the card itself says that file was in flight), and read #6950 — the card about ComponentInput's declaration surface — alongside it.

⚠️ Contention I am flagging rather than hiding

#6917 is in flight from this seat and touches packages/plugin-detail as one of its six sites. A same-file collision with src/index.tsx is possible, though the hunks are unrelated (a registration declaration here, a count/value read arm there). The dev is instructed to merge main if it lands behind, and ⛔ never to rebase or force-push.

Dispatch order follows in the next comment.

Generated by Claude Code

os-dev-report

{
  "issue": 6955,
  "status": "done",
  "branch": "claude/issue-6955-headercolor-authoring-surface",
  "pr": "https://github.com/objectstack-ai/objectui/pull/8627",
  "premise_still_valid": true,
  "component_input_can_express_options": {
    "answer": true,
    "evidence": "packages/types/src/base.ts on origin/main b1a14f28e declares `enum?: string[] | { label: string; value: any }[]`, documented 'for type=enum'. It is NOT one of the ADR-0049 tombstones (label / defaultValue / advanced / inputType / min / max / step / placeholder are all `?: never`). Three live readers, all re-measured rather than inherited: (1) packages/sdui-parser/src/index.ts forwards `enum` in the serializer's fixed key list (name, type, of, required, enum, binding, description); (2) packages/sdui-parser/src/validate.ts `checkType` raises code `invalid-enum` at severity ERROR for a value outside the list; (3) packages/sdui-parser/src/codegen.ts emits it as a TypeScript literal union on the JSX surface. The DESIGNER renders none of it and never did: an independent census of non-test readers of ComponentMeta.inputs finds only page.tsx:464, the sdui-parser serializer/codegen, Registry.ts:409's data-source injection seam, and one WRITE in WidgetRegistry.ts:182. objectui#6950 was read alongside and is CLOSED (PR #8297, framework-injected `binding` typed instead of cast); it neither blocks nor constrains this."
  },
  "branch_taken": "NARROW to the six values (triage's preferred outcome). The honest-label fallback and the domain:spec capability card were NOT needed.",
  "six_values_read_from_source": [
    "muted",
    "muted/50",
    "accent",
    "primary/10",
    "secondary/10",
    "destructive/10"
  ],
  "what_changed": "packages/plugin-detail/src/index.tsx, the detail-section registration: `{ name: 'headerColor', type: 'string' }` becomes `type: 'enum'` with `enum: Object.keys(headerColorVocabulary)` plus a description naming the six. EXACTLY ONE ARM, no 'string' beside it — a union would clear every value again and would thereby DECLARE the renderer's verbatim bg-* pass-through that ruling A refused to declare. The six are DERIVED from ./headerColor (this package's resolver map), not a fourth hand-copy: headerColor.contractPin-6594.test.ts already pins that map one-to-one against DetailViewSection.headerColor and its zod mirror in BOTH directions, so this surface inherits that pin.",
  "files": [
    "packages/plugin-detail/src/index.tsx (modified — the declaration + a comment stating why one arm)",
    "packages/plugin-detail/src/__tests__/detailSectionHeaderColorEnum-6955.test.ts (new — 8 pins)",
    "packages/plugin-detail/package.json (devDependencies: @object-ui/sdui-parser workspace:* — same declaration plugin-calendar/kanban/grid already carry for this class of test)",
    "pnpm-lock.yaml (that devDependency)",
    ".changeset/6955-headercolor-authoring-enum.md (new — minor)"
  ],
  "generated_outputs_regenerated": "NONE EXIST TO REGENERATE. sdui.manifest.json / sdui-intrinsics.d.ts / sdui-blocks.md are build-time artifacts written by packages/sdui-parser/scripts/gen-manifest.ts into an output dir; `git ls-files` finds no tracked copy of any of them. And detail-section is outside the public tier those artifacts cover anyway — see pm_premises_falsified.",
  "pins_added": [
    "REFUSES a seventh value ('blue-100') — one invalid-enum diagnostic, severity error, message naming the prop and all six legal tokens",
    "ACCEPTS every token the contract declares — narrowing costs no legal write",
    "offers exactly ONE arm — inputTypeArms(type) deep-equals ['enum'], so there is no free-text escape hatch",
    "REFUSES the verbatim bg-* pass-through ('bg-accent') — the declared enum does not contain it and the validator rejects it",
    "leaves the renderer's pass-through behaviour UNCHANGED — headerColorClass('bg-accent') still returns 'bg-accent', asserted beside the refusal so the asymmetry cannot drift apart",
    "control — a genuinely unknown prop still draws unknown-prop",
    "control — a neighbouring key (title) still takes free text, so the refusal is about THIS key",
    "publishes exactly the contract vocabulary, in order, read from DetailViewSectionSchema.shape.headerColor's own options (a DIFFERENT source from the declaration's, so the comparison is not self-confirming)"
  ],
  "refusal_pinned": true,
  "passthrough_unchanged_verified": "YES, both halves. (a) headerColor.ts is not in the diff and headerColorClass('bg-accent') === 'bg-accent' is asserted green in the new file. (b) packages/types is byte-for-byte untouched: views.zod.ts blob 5c8a30dd579d2efafd97e6fa812c50f1a3f07d3d and views.ts blob 7e6de89ba5247348ee2f03f43dc03d0ab11be2ff are identical at BASE b1a14f28e and at HEAD; git diff --name-only BASE..HEAD -- packages/types is empty.",
  "red_leg_evidence": "Ablation ran from a COMMITTED tree. Each leg: mutate on disk, PROVE it landed with before/after occurrence counts in BOTH directions, run, then restore via `git checkout HEAD -- path` and prove the restore by blob-hash equality with the HEAD blob (3f7867383c673f18b297f5d2e7255e7207487a7a) AND an empty `git diff HEAD`. Trap-based restore, absolute REPO_ROOT paths. LEG A (as committed): vitest exit 0, 8/8. LEG B (the bug's own shape — type back to 'string'): on-disk proof \"type: 'enum',\" 10 to 9 and \"      type: 'string',\" 0 to 1; vitest exit 1; 4 failed / 4 passed. The four failures, verbatim: [1] `accepted headerColor=\"blue-100\" — the authoring surface is still free text: expected +0 to be 1`; [2] `expected [ 'string' ] to deeply equal [ 'enum' ]`; [3] `declaring the pass-through would promise a class only the host app's Tailwind build can emit: expected +0 to be 1`; [4] `expected [] to deeply equal [ 'muted', 'muted/50', 'accent', ...(3) ]`. The four that stayed green under LEG B are the four that SHOULD (the six still parse under a string arm, the renderer pass-through is untouched, both controls fire). Vitest resolves @object-ui/* to packages/*/src via the root config's alias table, so no dist is involved in these legs and there is no stale-build hazard.",
  "caricature_result": "LEG C, the caricature the triage named: keep the six AND add a free-text arm — `type: ['enum', 'string']`. On-disk proof \"      type: 'enum',\" 1 to 0 and \"type: ['enum', 'string'],\" 0 to 1; vitest exit 1; 3 failed / 5 passed. RED: the seventh-value refusal, the ONE-ARM assertion (`expected [ 'enum', 'string' ] to deeply equal [ 'enum' ]`), and the bg-* refusal. STILL GREEN: 'publishes exactly the contract vocabulary'. That is the whole point — a pin asserting only that the declaration LISTS the six would have passed an implementation strictly worse than the bug. FIRST ATTEMPT AT LEG C WAS VOID AND IS REPORTED AS SUCH: the one-line anchor matched two sites in the file, the guard refused to mutate, printed `LEG C MUTATION DID NOT LAND — reading VOID`, the tree was restored, and the leg was re-run with a unique two-line anchor. It was not silently retried until something landed.",
  "leg_d": "Per-test classification from vitest's JSON reporter (--reporter=json --outputFile), never the text reporter. LEG A: numTotalTests 8, passed 8, failed 0, failed suites 0 — no 'no tests', no 'No test suite found in file', no 'Failed Suites'. LEG B: numTotalTests 8, passed 4, failed 4, suite status 'failed' with four assertionResults status 'failed' and their failureMessages read individually. LEG C: numTotalTests 8, passed 5, failed 3. Every leg executed a non-zero test count, so no leg is a suite-level death misread as a verdict.",
  "non_regression_axis": "The plausible WRONG fix narrows the validator or the renderer, breaking authors whose bg-* value renders today under their own Tailwind build. Pinned: packages/types byte-identical to BASE (two blob hashes, above). Verified: headerColorClass('bg-accent') still returns 'bg-accent' and headerColor.ts is not in the diff. Whole-package regression: pnpm exec vitest run packages/plugin-detail/ — 146 files / 1309 tests passed. Plus the repo-wide inputs gates: registry-inputs-spec-parity, component-input-union-specimens, app-shell block-config (the DEFERRED_SECTION_CONTROLS list that names headerColor), and scripts/__tests__/unit-registry-absence-collision — 4 files / 273 tests passed.",
  "capability_card_filed": "NOT NEEDED — ComponentInput can already express an options list, so triage's fallback branch was not taken and no domain:spec capability card was filed. A DIFFERENT, unrelated finding was filed instead: objectui#8626 (see out_of_scope_findings).",
  "changeset_verdict_line": "Quoted from `node scripts/check-changeset-presence.mjs`: '(check mark)  1 source file(s) of 1 released package(s) changed, and this change declares 1 changeset(s): .changeset/6955-headercolor-authoring-enum.md.' Exit 0. Before the changeset existed the same script exited 1 and named @object-ui/plugin-detail / packages/plugin-detail/src/index.tsx, so the verdict is a measurement and not a default. `node scripts/check-changeset-no-major.mjs` exit 0: 'No changeset declares a major bump.' The bump is minor (AGENTS.md: objectui's own breaking changes ship as minor). skip-changeset was not used.",
  "summary": "detail-section's headerColor input is now type 'enum' over the six design-system tint tokens instead of a free-form string, restoring declared = enforced against the six-member enum objectui#6594 put on @object-ui/types. The list is derived from the package's own resolver map rather than hand-copied, so the existing one-to-one contract pin covers this surface too. Exactly one arm: no string arm beside it, because that would re-open the free-text box and would declare the renderer's bg-* pass-through that maintainer ruling A refused to declare. packages/types is untouched and the renderer's pass-through is unchanged; both are pinned or verified.",
  "tests": "pnpm exec vitest run packages/plugin-detail/ -> 146 files / 1309 tests passed (os-verify-lock VERDICT command-exit 0, held 202s). pnpm exec vitest run apps/console/src/__tests__/registry-inputs-spec-parity.test.ts apps/console/src/__tests__/component-input-union-specimens.test.ts packages/app-shell/src/views/metadata-admin/previews/__tests__/block-config.test.ts scripts/__tests__/unit-registry-absence-collision.test.ts -> 4 files / 273 tests passed (VERDICT command-exit 0). pnpm --filter @object-ui/plugin-detail type-check -> exit 0, AFTER pnpm --filter '@object-ui/plugin-detail^...' build (exit 0); on an unbuilt tree it exits 2 with TS2307 Cannot find module '@object-ui/core', which is a PRECONDITION, not red. The new pin file is inside that check: tsc -p tsconfig.test.json --listFiles lists it (1 hit among 197 package files), so the typecheck is not excluding tests. pnpm exec eslint on both changed files -> exit 0, 0 errors; the 31 warnings are pre-existing react-refresh/only-export-components and no-explicit-any at lines 93-244, above this diff's hunk. pnpm check:control-bytes, check:phantom-deps, check:unused-deps, scripts/check-side-effects-array.mjs, scripts/check-governed-queue-guard.mjs --test (NOT GOVERNED) all exit 0 by their own printed verdict lines. NOT MEASURED (precondition, recorded as neither green nor red): pnpm check:eager-closure exits 2 printing 'No eager-closure report ... This is a broken gauge, not a passing budget' because apps/console/dist does not exist; same class as check:doc-snippets / check:doc-examples / check:readme-exports and check:sdui-registration-pins. Left to CI, which builds first. ABLATION: see red_leg_evidence and caricature_result — rebuild is not applicable because vitest aliases @object-ui/* to packages/*/src, so no dist mediates these legs; on-disk mutation proof and blob-hash restore proof were done for every leg.",
  "mcp_calls": "9",
  "pm_premises_falsified": [
    "MECHANISM, corrected — the dispatch's and the card's 'why this is p2' says gen-manifest.ts serialises this input into sdui.manifest.json (the save gate / parser allowlist) and sdui-intrinsics.d.ts (the JSX surface). FOR THIS BLOCK IT DOES NOT. gen-manifest.ts serialises ComponentRegistry.getPublicConfigs(), and detail-section is absent from PUBLIC_BLOCKS (packages/core/src/registry/public-blocks.ts) and declares no tier: 'public', so it reaches NEITHER artifact. The p2 grading and the fix are unchanged, because the surface is live through the OTHER consumer: packages/components/src/renderers/layout/page.tsx builds the JSX-page compiler's manifest from getKnownTypes() plus these same inputs and validateTree judges an authored page against it — the exact reach registry-inputs-spec-parity.test.ts records for element:record_picker, likewise outside PUBLIC_BLOCKS. So the three consequences the triage listed all hold; only the named artifact was wrong.",
    "TEXT, stale — the card and the triage comment both quote the line as `{ name: 'headerColor', type: 'string', label: 'Header Color' }` measured 2026-09-05. On b1a14f28e it is `{ name: 'headerColor', type: 'string' }`: ComponentInput.label became an ADR-0049 tombstone (`?: never`) in objectui#7493 after that measurement. Re-located by content as instructed; no impact on the fix.",
    "CONTENTION, did not materialise — objectui#6917 has not moved packages/plugin-detail/src/index.tsx between BASE b1a14f28e and origin/main, so no merge was needed and none was made. `git log BASE..origin/main -- packages/plugin-detail/src/index.tsx` is empty. The branch is 'behind' main in the ordinary way; the merge queue rebuilds on current main."
  ],
  "out_of_scope_findings": [
    "filed as objectui#8626: every detail-section input is a FLAT prop name (title, description, fields, collapsible, defaultCollapsed, columns, showBorder, headerColor) while DetailSection reads props.section.* only — the intersection of the two name sets is empty, and SchemaRenderer's createElement spread therefore hands the component section === undefined. Filed unassigned and ungraded, no labels, left for PM triage. The declaration-versus-props mismatch is MEASURED by reading both files; the stronger claim (that it THROWS at first render, since DetailSection's first statement reads section.defaultCollapsed) is labelled in the card as a reading and not an execution."
  ],
  "open_questions": []
}

Notes the JSON cannot carry cleanly:

PR is DRAFT and auto-merge is NOT armed, per the dispatch. The diff is NOT GOVERNED by scripts/check-governed-queue-guard.mjs --test (5 paths checked against 5 governed surfaces, none matched), so the draft posture here is the dispatch's instruction rather than the governed-surface rule.
Verification-lock etiquette. Every package-level and ablation run went through scripts/pm/os-verify-lock.sh with a stable slot. One attempt returned VERDICT queue-timeout (exit 99) after 9m00s behind objectui#6917's suites; the interval was spent on lock-free work (changeset, typecheck, lint, PR body) and the run was retried on the same slot rather than abandoned. Two narrowings are declared rather than hidden: single-file iteration runs of the new pin (sub-second) and the dependency-closure build were run unlocked.
Nothing was done to the card's fields. No assignee written, no second Claim: posted, no labels changed. The newest Claim: (comment 5588998056) names this session and this exact branch; it was verified before the first edit.

Generated by Claude Code

PM contract review — narrowed (triage's preferred branch), PR #8627 accepted, flipped out of draft, auto-merge armed

The deciding question was answered by measurement, not by reading a type

ComponentInput can express an options list: packages/types/src/base.ts declares enum?: string[] | { label, value }[], and it is not one of the ADR-0049 tombstones (label / defaultValue / advanced / inputType / min / max / step / placeholder are all ?: never). Three live readers re-measured rather than inherited — the sdui-parser serializer forwards it, validate.ts checkType raises invalid-enum at ERROR, and codegen.ts emits a TypeScript literal union.

⇒ The honest-label fallback and the domain:spec capability card were not needed. objectui#6950 was read alongside as instructed and is closed (PR #8297) — it neither blocks nor constrains.

⭐ The `bg-*` trap was avoided deliberately, and the caricature proves it mattered

Exactly one arm: type: 'enum', with no 'string' beside it — because a union would clear every value again and would thereby declare the renderer's verbatim bg-* pass-through that ruling A refused to declare.

Triage named that as the easiest step to get wrong. LEG C ran it as the caricature — type: ['enum', 'string'] — and the result is the whole argument for this PR's pin set:

3 failed / 5 passed. RED: the seventh-value refusal, the ONE-ARM assertion, the bg-* refusal. STILL GREEN: "publishes exactly the contract vocabulary." ⇒ A pin asserting only that the declaration LISTS the six would have passed an implementation strictly worse than the bug.

That is the discriminating question answered by measurement rather than by argument.

⭐ A void ablation leg was reported as void instead of retried until it landed

The one-line anchor matched two sites, the guard refused to mutate, printed LEG C MUTATION DID NOT LAND — reading VOID, the tree was restored, and the leg was re-run with a unique two-line anchor. It was not silently retried until something landed.

A mutation that does not land makes an ablation read green for the wrong reason — a no-op leg is indistinguishable from a leg the pin survived. Building the guard and publishing the void reading is better practice than either half alone.

⚠️ A mechanism in the card, in triage's grading, and in my dispatch is WRONG — and the grade survives it

All three said gen-manifest.ts serialises this input into sdui.manifest.json (the save gate and parser allowlist) and sdui-intrinsics.d.ts (the JSX authoring surface). For this block it does not. gen-manifest.ts serialises ComponentRegistry.getPublicConfigs(), and detail-section is absent from PUBLIC_BLOCKS and declares no tier: 'public' ⇒ it reaches neither artifact.

The p2 grade and the fix are unchanged, because the surface is live through a different consumer: packages/components/src/renderers/layout/page.tsx builds the JSX-page compiler's manifest from getKnownTypes() plus these same inputs, and validateTree judges an authored page against it — the same reach registry-inputs-spec-parity.test.ts already records for element:record_picker, likewise outside PUBLIC_BLOCKS.

⇒ All three consequences triage listed hold; only the named artifact was wrong. That is the right shape for a correction: fix the mechanism, keep the conclusion, and say which is which. ⭐ I have relayed this to the objectui#6879 seat, which is working PUBLIC_BLOCKS right now — "not in PUBLIC_BLOCKS" does not mean "not on an authoring surface" is material to that card's own step-1 measurement.

Also confirmed rather than assumed

The six are DERIVED from the package's own resolver map, not hand-copied a fourth time — and headerColor.contractPin-6594.test.ts already pins that map one-to-one against DetailViewSection.headerColor and its zod mirror in both directions, so this surface inherits that pin. The vocabulary pin reads its expectation from DetailViewSectionSchema.shape.headerColor's own options — a different source from the declaration's, so the comparison is not self-confirming.
Non-regression measured both halves: packages/types byte-identical to BASE (two blob hashes quoted, git diff --name-only empty), and headerColorClass('bg-accent') still returns 'bg-accent' — asserted beside the refusal so the asymmetry cannot drift apart.
The contention I flagged did not materialise, and that was checked rather than assumed: git log BASE..origin/main -- packages/plugin-detail/src/index.tsx is empty, so finding(fields,plugin-detail,plugin-list,app-shell): six more non-QueryResult count/value arms survive the seams #6840 repaired — and one of them reads value AHEAD of data #6917 never touched this file. No merge made, no rebase, no force-push.
The card's quoted line was stale: label became an ADR-0049 tombstone in [finding] ComponentInput.defaultValue?: any (packages/types/src/base.ts:550) is why the marked plugin example in skills/objectui must carry a bare any — tighten the contract to unknown and the baseline row retires with the guide #7493 after triage's measurement, so the registration no longer carries it. Re-located by content, no impact.
No generated artifact to regenerate — git ls-files finds no tracked copy of any of them.
check:eager-closure recorded as NOT MEASURED on its own words ("a broken gauge, not a passing budget", apps/console/dist absent), not as red.

⚠️ The out-of-scope finding may be bigger than this card

#8626: every detail-section input is a flat prop name (title, description, fields, collapsible, defaultCollapsed, columns, showBorder, headerColor) while DetailSection reads props.section.* only — the intersection of the two name sets is empty, so SchemaRenderer's spread hands the component section === undefined.

⭐ And the epistemic handling is exactly right: the name-set mismatch is measured by reading both files, while the stronger claim — that it throws at first render, since DetailSection's first statement reads section.defaultCollapsed — is labelled in the card as a reading, not an execution. Deduped with two search_issues queries that both returned non-empty sets, so the empty duplicate set is a reading too.

Generated by Claude Code

feat(plugin-detail): declare `icon` on the `detail-section` node (obj…

feat(plugin-detail): declare `hideEmpty` on the `detail-section` node…

Navigation Menu

Repository navigation

`detail-section` still offers `headerColor` as a free-form `string` input, so the SDUI authoring surface invites values the published validator now refuses #6955

What was observed

Why it is worth a card

Reproduction

Notes for whoever picks this up

分诊 · `domain:ui` / `priority:p2` / `pm:queue`

定级 p2

边界测试 —— 收紧,不是加宽

⚠️ 三条接手前必须先读的事,卡面已写,本席全部保留并补一条

Carried into the dispatch as hard constraints, from triage comment 5548134853

⭐ Triage's fourth point is the one that decides whether this card can finish, and it is in the order verbatim

⚠️ Contention I am flagging rather than hiding

PM contract review — narrowed (triage's preferred branch), PR #8627 accepted, flipped out of draft, auto-merge armed

The deciding question was answered by measurement, not by reading a type

⭐ The `bg-*` trap was avoided deliberately, and the caricature proves it mattered

⭐ A void ablation leg was reported as void instead of retried until it landed

⚠️ A mechanism in the card, in triage's grading, and in my dispatch is WRONG — and the grade survives it

Also confirmed rather than assumed

⚠️ The out-of-scope finding may be bigger than this card

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

detail-section still offers headerColor as a free-form string input, so the SDUI authoring surface invites values the published validator now refuses #6955

Description

What was observed

Why it is worth a card

Reproduction

Notes for whoever picks this up

Activity

os-zhuang commented on Sep 5, 2026

分诊 · domain:ui / priority:p2 / pm:queue

定级 p2

边界测试 —— 收紧,不是加宽

⚠️ 三条接手前必须先读的事,卡面已写,本席全部保留并补一条

os-justin commented on Sep 8, 2026

Carried into the dispatch as hard constraints, from triage comment 5548134853

⭐ Triage's fourth point is the one that decides whether this card can finish, and it is in the order verbatim

⚠️ Contention I am flagging rather than hiding

os-justin commented on Sep 8, 2026

os-justin commented on Sep 8, 2026

PM contract review — narrowed (triage's preferred branch), PR #8627 accepted, flipped out of draft, auto-merge armed

The deciding question was answered by measurement, not by reading a type

⭐ The bg-* trap was avoided deliberately, and the caricature proves it mattered

⭐ A void ablation leg was reported as void instead of retried until it landed

⚠️ A mechanism in the card, in triage's grading, and in my dispatch is WRONG — and the grade survives it

Also confirmed rather than assumed

⚠️ The out-of-scope finding may be bigger than this card

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

`detail-section` still offers `headerColor` as a free-form `string` input, so the SDUI authoring surface invites values the published validator now refuses #6955

分诊 · `domain:ui` / `priority:p2` / `pm:queue`

⭐ The `bg-*` trap was avoided deliberately, and the caricature proves it mattered