Repository navigation
finding(plugin-form): object-form.customFields 注册描述说 merge，渲染器 replace —— 逐成员合并的代码存在但不可达 #9778
Description
Activity
Ruling: batch #161 item 5 · the code is wrong:
object-form.customFieldsMERGES over the metadata-generated field set, as its published registration description says — the per-member merge branch inObjectForm.tsxbecomes reachable whencustomFieldsis non-empty; the registration description is unchanged · maintainer 「其他同意」 2026-09-18T12:07ZDirector seat, summon #24,
session_01Wj1HUjzyeiBQ8atRf1ZhaL. Presented in detail with the recommendation 代码错; the maintainer agreed. Facts (this card; triage 5725870821): the registration description reads 「Field definitions merged over the set generated from object metadata」; the renderer replaces; the merge lookupschema.customFields?.find((f) => f.name === name)runs only whenhasInlineFieldsis false, i.e. only on an emptycustomFields— dead code that records the intent; PR objectui#9774 pinned the current (replace) behaviour deliberately so any settlement re-reads it.Ruling
ObjectForm.tsx: with a non-emptycustomFields, the metadata-generated set is rendered with each named member overridden by itscustomFieldsentry and unnamed members kept; acustomFieldsentry naming a field absent from metadata is appended. The registration description stands as the contract.- PR objectui#9774's member pin flips to the merge behaviour, with one case per direction (override, keep, append).
- Changeset
@object-ui/plugin-formminor with a behaviour note: hosts that usedcustomFieldsas a full replacement now see the metadata-generated members too; ⛔ no 「replace mode」 option is added (a host wanting a bespoke set declares its own form). - ⛔ 散文错 would narrow the published author surface below what it promised, on a surface whose own code shows the wider intent.
Four-facet reading: ① the declared contract wins and the existing merge code shows it was the intent; ② 「override a few, keep the rest」 is the everyday use of a custom-fields hook; ③ an AI author reading the description gets what it expects; ④ nothing declared, an unreachable branch made reachable.
Execution
needs-user-decision→pm:queue;domain:spec,priority:p2stay (lands in objectui's plugin-form).Clause-②: no.
Generated by Claude Code
Claim: PM loop round R+301 — objectui#9778
object-form.customFieldsmerges, per the 12:07Z ruling
Session:session_01UanLVj6xvbS6puBCewLr8L
Branch:claude/issue-9778-customfields-merge
Worktree:objectui-issue-9778
Domain:domain:spec
Seat:domain:spec#1
File surface:packages/plugin-form/src/ObjectForm.tsx·packages/plugin-form/src/__tests__/objectFormCustomFieldsMembers-8071.test.tsx· one.changeset/entry
Container & model: M,mode:subagent,model: opus
Clause-②: no
Thread-read: 5729750439Why this card is dispatchable now, when its own body says it is not
⚠️ The body ends 「⇒ 需要定口径后才可派发」 and triage5725870821agreed, filing itneeds-user-decision. That is superseded: the director seat ruled it at5729750439(batch #161 item 5, summon #24) with the maintainer's 「其他同意」 at 2026-09-18T12:07Z — 代码错, the renderer merges, the registration description stands as the contract — and executedneeds-user-decision→pm:queue. ⇒ the口径 exists, and the ruling is quoted verbatim into the dispatch order as the binding constraint. ⛔ This seat did not pick the side.Staleness check — all three faces, read this act
- 动作面 —
packages/plugin-form/src/ObjectForm.tsxonmain, last 8 commits: newest is6df914164f(2026-09-18T14:29Z, thescale/precisionstep fix). ⛔ The ruled action is not already done: atmainthe replace path still stands at:724(if (hasInlineFields && schema.customFields)→setFormFields(schema.customFields.map(...))) and the merge lookup still sits at:793(schema.customFields?.find(f => f.name === name)) behind:786'sif (!field && !hasInlineFields) return;. Premise holds. - 卡引用面 — objectui#9774 is MERGED ⇒ its pin is in
main, and rows 2 and 3 ofobjectFormCustomFieldsMembers-8071.test.tsx(:119,:134) are the two that assert the replace behaviour and must flip. objectui#8071 is open andpm:dispatchedto this seat. - 工作项面 — the two fences the seat post carried are both gone: objectui#9873 and objectui#9896 are merged ⇒
ObjectForm.tsxandObjectGrid.tsxare FREE.
Serial constraints
✔
packages/plugin-form/src/ObjectForm.tsx— FREE, held by no open PR.
⛔apps/console/src/__tests__/registry-inputs-spec-parity.test.ts— HELD by open PR objectui#9924 (slice 17, parked). objectui#9774 touched both that ledger and the member pin, so the order fences the ledger explicitly: flipping what a pin asserts does not change whether the member is pinned, so the ledger's counts should not move — and if the dev finds it must, it stops and hands back rather than editing a held file.Read at 2026-09-18T19:05Z.
Generated by Claude Code
- 动作面 —
os-dev-report
{
"issue": 9778,
"status": "done",
"branch": "claude/issue-9778-customfields-merge",
"pr": "#9948",
"session": "session_01UanLVj6xvbS6puBCewLr8L",
"premise_still_valid": true,
"summary": "Implemented the 12:07Z ruling:object-form.customFieldsnow MERGES over the metadata-generated field set in ObjectForm.tsx, and the registration description is untouched. The replace early-return is gone, the object schema is fetched whenever an objectName and a data source are present (without it there is nothing to merge over and the existing per-memberfindstays dead code), and members naming nothing in the generated set are appended in authored order. Two deliberate consequences: the loading flag now drops only when nothing will be fetched (otherwise an empty form flashed for the duration of getObjectSchema), and a failing getObjectSchema falls back to the members-only source when inline members are present instead of replacing a form that used to render with an error panel. Premise re-verified at branch base 00c4df5: the replace path and the unreachable merge lookup were both still there. The fenced ledger apps/console/src/tests/registry-inputs-spec-parity.test.ts was NOT edited and its test stays green.",
"tests": "All at 096496e (final commit), each exit code captured before any pipe, every run through scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=dev-9778. (1) pnpm --filter @object-ui/plugin-form test -> VERDICT command-exit 0, 'Test Files 100 passed (100) / Tests 940 passed | 1 skipped (941)'. (2) pnpm --filter @object-ui/plugin-form type-check -> exit 0, after pnpm --filter '@object-ui/plugin-form^...' build (exit 0); without that dependency-closure build it emits TS2307 'Cannot find module @object-ui/*' on every workspace import - a missing dist, not a type error. (3) consumers that pass customFields through: packages/plugin-view/ + packages/plugin-designer/ + registry-inputs-spec-parity.test.ts + public-block-binding-reach.test.tsx -> exit 0, 'Test Files 66 passed (66) / Tests 757 passed (757)'. (4) The pin file alone: 6 selected, 6 passed (selected count asserted; the whole file was run, no -t regex). (5) ABLATION, fix committed first, then ObjectForm.tsx alone reverted to base 00c4df5 with a trap-guarded absolute-path restore: mutation proven on disk (objectui#9778 markers 5 -> 0; the replace line setFormFields(schema.customFields.map(normalizeVisibility)) back at 1 occurrence; blob 85d6c1c4 vs HEAD cd4bd869) and the flipped pins went '4 failed | 2 passed (6)' - row 1 expected ['zz','note'] to equal ['customer','note','amount','zz'], row 2 expected +0 to be 1 (getObjectSchema), row 3 expected null to be 'Customer', row 4 expected ['zz','yy'] to equal ['customer','note','amount',...]; rows 5-6 stayed green because neither is about the merge. Restoration proven by state, not exit code: blob back to cd4bd869 (identical to HEAD), git diff HEAD empty, markers back to 5. No ablation file left behind; one scratch test file used to measure thefieldswhitelist migration route was deleted and its absence verified via git status. (6) Gates, all exit 0: check:control-bytes, check:new-line-citations, check:changeset-claims, check:pending-changeset-literals, check:test-path-roots, check:vi-mock-specifiers, check:vi-mock-inherit, check:vi-mock-override-shape, check:element-data-source-declaration, node scripts/check-changeset-presence.mjs (1 changeset declared for 3 published source files), node scripts/check-changeset-no-major.mjs. eslint on the three changed files: 0 errors, 64 pre-existing no-explicit-any warnings. Control-byte self-scan of all four changed files: no match, with a lit control (a crafted 0x01 byte) firing on the same instrument. (7) NOT MEASURED: check:sdui-registration-pins exits 2 with its own prerequisite refusal ('No console build to weigh at apps/console/dist/assets') - a prerequisite, not a red; this diff adds and removes no registration, CI builds the console and runs it. CI convergence on PR 9948 is in_progress and not waited on.",
"mcp_calls": "0 - no MCP GitHub tool was called, read or write; all GitHub traffic went through the REST proxy with curl",
"api_writes": "2 - POST /repos/objectstack-ai/objectui/pulls (draft PR 9948), POST /repos//issues/9778/comments (this report). 0 label writes: the dispatch reserves labels for the PM, and the standing objectui clause forbids this seat applying the changeset-exemption label there in any case, so the two do not actually conflict. The PR carries no labels at read-back, including no needs:contract-review.",
"open_questions": [],
"out_of_scope_findings": [
"to file (class b, dedupe words:DrawerForm customFields replace/ModalForm customFields replace/presentation arms merge parity/formType drawer inline fields/customFields merge unreachable): the same registered description still covers the drawer and modal presentations of this block and both still REPLACE. ObjectForm spreads the schema - customFields included - into DrawerForm / ModalForm for formType 'drawer' | 'modal', and each runs its ownif (schema.customFields?.length) { setFormFields(schema.customFields); }. One declared key, a merge on the flat presentation and a replacement on the other two. Outside this card's file surface (the ruling names ObjectForm.tsx) and it needs its own pins.",
"noted, not filed: the console ledger entry for object-form.customFields still describes the replace behaviour in prose (they REPLACE the metadata-generated set ... getObjectSchema asserted at zero calls ... handed back as a finding rather than repaired). It is reviewed prose, not a computed assertion - the ledger's counts do not move and its test stays green (measured) - and the file is fenced/held by open PR objectui#9924. Successor: whoever lands objectui#9924, or the next objectui#8071 slice that edits that ledger."
]
}
Generated by Claude Code
os-dev-report
{
"issue": 9778,
"status": "done",
"branch": "claude/issue-9778-customfields-merge",
"pr": "#9948",
"session": "session_01UanLVj6xvbS6puBCewLr8L",
"premise_still_valid": true,
"summary": "Implemented the 12:07Z ruling:object-form.customFieldsnow MERGES over the metadata-generated field set in ObjectForm.tsx, and the registration description is untouched. The replace early-return is gone, the object schema is fetched whenever an objectName and a data source are present (without it there is nothing to merge over and the existing per-memberfindstays dead code), and members naming nothing in the generated set are appended in authored order. Two deliberate consequences: the loading flag now drops only when nothing will be fetched (otherwise an empty form flashed for the duration of getObjectSchema), and a failing getObjectSchema falls back to the members-only source when inline members are present instead of replacing a form that used to render with an error panel. Premise re-verified at branch base 00c4df5: the replace path and the unreachable merge lookup were both still there. The fenced ledger apps/console/src/tests/registry-inputs-spec-parity.test.ts was NOT edited and its test stays green. ADDENDUM (second commit 0750ec3, one extra file added to the surface by the PM):.changeset/8738-object-form-fields-description.mdhad a ROTTED claim and was repaired prose-only - dated, falsifier named, frontmatter and semver level untouched. See out_of_scope_findings item 3 for the measurement.",
"tests": "All at 096496e (final commit), each exit code captured before any pipe, every run through scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=dev-9778. (1) pnpm --filter @object-ui/plugin-form test -> VERDICT command-exit 0, 'Test Files 100 passed (100) / Tests 940 passed | 1 skipped (941)'. (2) pnpm --filter @object-ui/plugin-form type-check -> exit 0, after pnpm --filter '@object-ui/plugin-form^...' build (exit 0); without that dependency-closure build it emits TS2307 'Cannot find module @object-ui/*' on every workspace import - a missing dist, not a type error. (3) consumers that pass customFields through: packages/plugin-view/ + packages/plugin-designer/ + registry-inputs-spec-parity.test.ts + public-block-binding-reach.test.tsx -> exit 0, 'Test Files 66 passed (66) / Tests 757 passed (757)'. (4) The pin file alone: 6 selected, 6 passed (selected count asserted; the whole file was run, no -t regex). (5) ABLATION, fix committed first, then ObjectForm.tsx alone reverted to base 00c4df5 with a trap-guarded absolute-path restore: mutation proven on disk (objectui#9778 markers 5 -> 0; the replace line setFormFields(schema.customFields.map(normalizeVisibility)) back at 1 occurrence; blob 85d6c1c4 vs HEAD cd4bd869) and the flipped pins went '4 failed | 2 passed (6)' - row 1 expected ['zz','note'] to equal ['customer','note','amount','zz'], row 2 expected +0 to be 1 (getObjectSchema), row 3 expected null to be 'Customer', row 4 expected ['zz','yy'] to equal ['customer','note','amount',...]; rows 5-6 stayed green because neither is about the merge. Restoration proven by state, not exit code: blob back to cd4bd869 (identical to HEAD), git diff HEAD empty, markers back to 5. No ablation file left behind; one scratch test file used to measure thefieldswhitelist migration route was deleted and its absence verified via git status. (6) Gates, all exit 0: check:control-bytes, check:new-line-citations, check:changeset-claims, check:pending-changeset-literals, check:test-path-roots, check:vi-mock-specifiers, check:vi-mock-inherit, check:vi-mock-override-shape, check:element-data-source-declaration, node scripts/check-changeset-presence.mjs (1 changeset declared for 3 published source files), node scripts/check-changeset-no-major.mjs. eslint on the three changed files: 0 errors, 64 pre-existing no-explicit-any warnings. Control-byte self-scan of all four changed files: no match, with a lit control (a crafted 0x01 byte) firing on the same instrument. (7) NOT MEASURED: check:sdui-registration-pins exits 2 with its own prerequisite refusal ('No console build to weigh at apps/console/dist/assets') - a prerequisite, not a red; this diff adds and removes no registration, CI builds the console and runs it. CI convergence on PR 9948 is in_progress and not waited on. ADDENDUM at 0750ec3 (the changeset-prose commit): gates re-run, all exit 0 - check:changeset-claims, check:pending-changeset-literals, check:control-bytes, check:new-line-citations, node scripts/check-changeset-presence.mjs, node scripts/check-changeset-no-major.mjs, and node scripts/check-changeset-overwrite.mjs, which reports this edit as its own case 2 ('correcting a declaration on purpose ... legitimate'), report-only, exit 0. No code changed in that commit, so no test re-run was owed. CI, read once per head and NOT waited on - a snapshot of a running matrix is not a verdict, and cancelled/timed_out were counted as not-success: at 096496e, 36 check runs, 30 completed (27 success, 3 skipped), 0 not-success, 6 still running (Test shards 1-4, Type Check, Lint); at the new head 0750ec3, 24 check runs, 1 completed (success), 21 in progress, 2 queued, 0 not-success. Nothing has failed at either head; neither reading is a pass.",
"mcp_calls": "0 - no MCP GitHub tool was called, read or write; all GitHub traffic went through the REST proxy with curl",
"api_writes": "3 - POST /repos/objectstack-ai/objectui/pulls (draft PR 9948), POST /repos//issues/9778/comments twice (the first report, issuecomment-5735127215, and this superseding one). Reads of the gate comment, the PR and the two heads' check-runs were GETs. 0 label writes, 0 PATCHes: the PR body was written once at creation and NOT amended, so the changeset repair is described here instead - if the body should carry it, the seat writes that; the dispatch reserves labels for the PM, and the standing objectui clause forbids this seat applying the changeset-exemption label there in any case. The PR carries no labels at read-back, including no needs:contract-review.",
"open_questions": [],
"out_of_scope_findings": [
"to file (class b, dedupe words:DrawerForm customFields replace/ModalForm customFields replace/presentation arms merge parity/formType drawer inline fields/customFields merge unreachable): the same registered description still covers the drawer and modal presentations of this block and both still REPLACE. ObjectForm spreads the schema - customFields included - into DrawerForm / ModalForm for formType 'drawer' | 'modal', and each runs its ownif (schema.customFields?.length) { setFormFields(schema.customFields); }. One declared key, a merge on the flat presentation and a replacement on the other two. Outside this card's file surface (the ruling names ObjectForm.tsx) and it needs its own pins.",
"noted, not filed: the console ledger entry for object-form.customFields still describes the replace behaviour in prose (they REPLACE the metadata-generated set ... getObjectSchema asserted at zero calls ... handed back as a finding rather than repaired). It is reviewed prose, not a computed assertion - the ledger's counts do not move and its test stays green (measured) - and the file is fenced/held by open PR objectui#9924. Successor: whoever lands objectui#9924, or the next objectui#8071 slice that edits that ledger.",
"in scope by PM extension, DONE in commit 0750ec3:.changeset/8738-object-form-fields-description.md(flagged by the changeset-claim-re-read gate, comment 5735116663, because it namesObjectForm.tsx). MEASURED independently at this PR's head, not inherited: both named read sites callwarnUnresolvedTopLevelField- inSimpleObjectForm'sfieldNames.forEachloop in ObjectForm.tsx under the comment 'Warn, don't accept', and inbuildFlatFieldsin flatFields.ts under 'warn, don't accept' - and the helper in sectionFields.ts is a de-duplicatedconsole.warnkeyed by object plus member shape. Both still SKIP the member (return/continue), and neither throws or renders an empty state. DATED, not rewritten:git logat this head gives the body's landing commit fd9bf26 at 2026-09-09T14:28:48Z and the falsifier 8fda009 (objectui#8859, route 1 of the same card) at 2026-09-09T16:00:31Z, 92 minutes later; both are ancestors of my head (merge-base --is-ancestorexit 0, the self-evidencing direction). The clause dated is 'is silently skipped ... no throw, no warning, no empty-state': 'silently' was dropped from the present-tense sentence and the original reading is preserved verbatim in a dated note that names the falsifier and states what did NOT change (still skipped, still no throw, still no empty-state). Read as a paragraph, not a line: the same fact also dated the opening 'route 1, a diagnosticconsole.warn, is a separate ruling still pending', which now says it was pending when written and landed 92 minutes later. 'Behaviour is unchanged' was JUDGED in the same pass and found STILL TRUE - it is a claim about that entry's own change, a registration description, and my diff does not touch it; it now reads 'Behaviour is unchanged by this entry' so a CHANGELOG reader cannot take its scope for the release. Frontmatter and thepatchlevel untouched; no other pending changeset touched."
]
}
Generated by Claude Code
ACCEPT — objectui#9778, delivered by PR objectui#9948 at head
0750ec3e309bd674843c3f0c1c5f8b46f557d27b.Implemented-by:
claude/issue-9778-customfields-mergeVerified by this seat, ⛔ not taken from the dev's report
check reading clause ② no— 0 added lines whose first non-space token isexport, against a lit control of 18 added lines containingcustomFieldson the same diff. ⛔ Tested against this diff, not inherited from the cardchangeset level @object-ui/plugin-form: **minor**— the level the ruling named, with the behaviour note in the ruling's own words and ⛔ no 「replace mode」 optionserial fence apps/console/src/__tests__/registry-inputs-spec-parity.test.tsabsent from the file list ⇒ the ledger held by open PR objectui#9924 was not touchedgoverned surface 0 of 5 paths hit the register ⇒ ordinary queue landing CI at this head 36 of 36 completed — 33 success · 3 skipped, nothing running, ⛔ zero cancelledand zerotimed_out. All four test shardssuccessmergeable_statebehind→clean, self-resolved after objectui#9919 landed underneath ⇒ ⛔ no sync commit was pushed⚠️ The CI row above is the settled matrix. Two earlier readings on this PR (27/36 and 33/36 with shards still running) were ⛔ not treated as passes — a snapshot of a running matrix says only that nothing has failed yet.The ruling was implemented, ⛔ not reinterpreted
The director-seat ruling (
5729750439, maintainer 「其他同意」) is carried verbatim in the PR body: a non-emptycustomFieldsnow merges over the metadata-generated set — named members override, unnamed members are kept, members naming nothing in metadata are appended. The registration description is unchanged, which is the half the ruling explicitly protected.⭐ Two consequences the dev declared rather than hid: the loading flag now drops only when nothing will be fetched (otherwise an empty form flashed for the duration of
getObjectSchema), and a failinggetObjectSchemafalls back to the members-only source when inline members are present. Both follow necessarily from fetching the schema at all; ⛔ neither is a second ruling.A rotted claim was repaired on the way through
The
changeset-claim-re-readgate flagged.changeset/8738-object-form-fields-description.md, whose body publishes verbatim into the CHANGELOG. Its sentence 「is silently skipped … no throw, no warning, no empty-state」 is false:ObjectForm.tsx:781andflatFields.ts:66both callwarnUnresolvedTopLevelField.⭐ ROTTED, ⛔ not born false — the body landed in
fd9bf26df0and its falsifier8fda009057(objectui#8859) landed 92 minutes later ⇒ the reading was TRUE when written, so it is dated with the falsifier named, ⛔ never rewritten as though always wrong. The still-true halves were kept and only scoped. ⭐ Read as a paragraph rather than a line, a second rotted half surfaced in the same entry — 「route 1 … is a separate ruling still pending」, when route 1 is the falsifier — which is exactly the objectui#8617 shape where repairing one half leaves the paragraph asserting the same wrong thing.⚠️ Created by this change, owed as its own cardDrawerFormandModalFormstill replace: each runs its ownif (schema.customFields?.length) { setFormFields(schema.customFields); }, whileObjectFormnow merges. ⇒ one registered description, one merge arm and two replace arms — class (b), and ⛔ outside this card's ruled file surface. ⛔ Unfileable while REST issue creation is refused for this session. Dedupe words:DrawerForm customFields replace·ModalForm customFields replace·presentation arms merge parity.Proceeding: ready → enqueue → follow to MERGED.
Clause-②: no⇒ ⛔ no contract-review route is owed.Read at 2026-09-18T19:52Z.
Generated by Claude Code
Release: objectui#9778 — delivered by PR objectui#9948, MERGED and verified on
origin/mainBY SYMBOL, COMMENTS STRIPPED.⛔ Not read from
merged/merged_at/merge_commit_sha— all three lie in both directions.⚠️ And ⛔ not from the tip:mainmoved five times on other PRs while this sat in the queue (00c4df509c→05a49f2ee→e22fa12335→0c789a4024→d6a84446fb→ …). A moved tip proves nothing.⭐ The first reading was WRONG, and the repair is the discipline itself
packages/plugin-form/src/ObjectForm.tsx, subjectsetFormFields(schema.customFields.map:reading at the branch base on origin/mainnowcomments STRIPPED — the judgement 1 0 ✅ comments included — ⛔ the wrong instrument 1 1 lit control hasInlineFields(stripped)17 17 merge lookup customFields?.find(stripped)1 1 The comments-included count still reads 1 because the landed change documents what it replaced:
:750–:762explains that a non-emptycustomFields「used to return from here withsetFormFields(schema.customFields.map(normalizeVisibility))BEFORE the generated set existed, so that lookup only ever ran over an empty array」.⇒ Judged on the unstripped count this landing would have read NOT LANDED. ⭐ 「by symbol, comments stripped, with a lit control」 is not a flourish — here it is the whole difference between a true and a false verdict, and the seat's first probe omitted the stripping.
⚠️ The control is what makes the0a reading:hasInlineFieldsholds at 17 on both trees, so the instrument was alive when it returned zero. AndcustomFields?.findstands at 1 on both — that is the point of the card: the lookup was never absent, it was unreachable, and it is now reached.What landed
The early-return replace path is gone; the effect now runs
if (!objectSchema) return;and merges. The three ruled directions are carried in the code's own comment and pinned one case each inobjectFormCustomFieldsMembers-8071.test.tsx: override (a member naming a declared field replaces it in the generated set's position), keep (a declared field no member names still renders), append (a member naming nothing in metadata is added after the generated set, in authored order). The registration description is unchanged — the half the ruling explicitly protected.The rotted changeset repair landed with it:
.changeset/8738-object-form-fields-description.mdonmaincarries the dated correction (Dated correctionpresent, falsifier8fda009057named). ⇒ the CHANGELOG this publishes into will ⛔ no longer assert a silent drop that stopped being silent 92 minutes after it was written.Half-state cleared in this same act: the keyword close left
pm:dispatchedand the assignee standing; both go with this release.⚠️ Owed and ⛔ unfileable while REST issue creation is refused:DrawerFormandModalFormstill replace — each runs its ownif (schema.customFields?.length) { setFormFields(schema.customFields); }— so one registered description now spans one merge arm and two replace arms. Class (b), created by this landing. Dedupe:DrawerForm customFields replace·ModalForm customFields replace·presentation arms merge parity.Read at 2026-09-18T20:15Z.
Generated by Claude Code
- added a commit that references this issue
on Sep 28, 2026
由
domain:spec@ objectui 执行席立卡（sessionsession_01UanLVj6xvbS6puBCewLr8L）。由 objectui#8071 slice 11 的 dev 在 PR objectui#9774 上测出并交回，⛔ 未顺手修 —— 那张卡只写 pin。事实
object-form.customFields的注册描述说的是合并：渲染器替换，不合并。
⭐ 而且逐成员合并的代码确实存在，只是到不了：
ObjectForm.tsx的 metadata 分支里有而那个分支只在
hasInlineFields为假时运行 —— 也就是只在customFields缺席或为空时运行。⇒ 那次查找永远是在空集上做的。⇒ 声明面说「merged over」，实现面在有值时走的是另一条路；写下合并语义的那段代码是死代码。
判据
类 (b) —— 违背已声明契约：注册描述是已发布的作者面，作者按它写
customFields会以为自己在覆盖若干个字段、其余沿用 metadata，实际拿到的是整套被替换。⛔ 本卡不裁哪一边错
散文错（实现才是意图，改描述）还是代码错（描述才是意图，让合并分支可达）—— 这是 spec / 渲染器的取舍，⛔ 不是 dev 的、也⛔ 不是本席的。⇒ 需要定口径后才可派发。
object-form.customFields成员 pin 钉的是当前行为（替换）。⇒ 无论往哪边settle，那条 pin 都会变红，并强制重读一次。⭐ 这是故意的，⛔ 不是要修绿的东西。查重词：
customFields merge unreachable·ObjectForm inline fields·registration description wrong·hasInlineFields branch·merged over metadataRefs: objectui#8071 · objectui#9774
Generated by Claude Code