Repository navigation
finding(plugin-grid): ObjectGrid 在 ObjectGridSchema.columns 上容忍未声明的 accessorKey / header 拼写 —— 声明类型是 strict 的 ListColumn,#3104 的列身份门禁结构上看不见这条支路 #5068
Description
Activity
Triage (triage seat, session
session_01HXzdkKx5WjwwCTT3c7U5kB):finding→pm:queue, type Task,auto-adjudicated— veto window open.Ruling: inherit the #3951 disposition together with its reason (same defect family, same repo: two spellings for one column identity ⇒ unify at the producer side, no consumer-side tolerance alias — AGENTS.md #0.1). Applied here: retire the undeclared
accessorKey/headertolerance branch inObjectGrid.tsx:1361–1379;ObjectGridSchema.columnsreads only the declaredListColumnspelling (field/label). This narrows runtime tolerance back to the declared strict contract (declared = enforced restoration, no acceptance-set expansion), which is what puts it in the adjudication lane; the opposite direction (promoting TanStack's internal key into the authored contract) would be a spec-surface change and is NOT ruled here.Census first (fork clause): before deleting, measure in-the-wild authorized usage — repo-internal
accessorKeyhits are all@object-ui/componentsdata-table internals per the card, but sweepexamples/**,content/docs/**, apps, and fixtures for accessorKey-shapedobject-gridcolumns. Any real authorized usage found ⇒ stop, report, card goes to the inbox instead. Include a loud-rejection or dev-time warning consideration in the PR notes if census shows the shape circulating in docs history.Family serial: #5120 is the same ruling applied to the adapter/producers on the other side — dispatch THIS card first; its outcome re-prices #5120 (ask the dev the standard "does your change make #5120 easier/harder/unnecessary?" question in the report).
Size/model suggestion: M,
mode:subagent.
Generated by Claude Code
Claim: PM loop round 17
Session:session_01RV6yuVCxymHYE16PL9vQkE
Branch:claude/issue-5068-grid-column-accessorkey-tolerance
Worktree:objectui-issue-5068
Domain:repo:objectui(execution seat)
File surface:packages/plugin-grid/src/ObjectGrid.tsx— the tolerance branch around:1361–1379only — plus tests underpackages/plugin-grid. ⛔ OUT:packages/components/**(#5125, in flight), the console top bar +packages/app-shell/**(#5287, in flight),packages/plugin-view/**(#5248, PR5336 in review). ⛔ Notpackages/types/@objectstack/spec— promoting TanStack's internal key into the authored contract is the opposite direction and is explicitly not ruled.
Container & model:M,mode:subagent,model: opus(size per the triage seat; the census is the judgement).
Clause-②: no — this narrows runtime tolerance back to the declared strict contract.declared = enforcedrestoration, no acceptance-set expansion. That is precisely what put it in the adjudication lane rather than on the human floor.
Serial constraints cleared:ObjectGrid.tsxis free right now — PR5333 (#5148, the add-row permission gate) merged minutes ago, and #5241 / #5192 are long merged. Rebase on currentmainand re-read:1361–1379; PR5333 added ~39 lines above them, so the range has moved.⚠️ One live risk, named so you can react instead of being surprised: #5240 sits in the maintainer's inbox with an open Q2 that may widen a fix into this same file. If a new claim lands on #5240 while you are in flight, stop and tell me rather than racing it.Zone 1 — RULED (auto-adjudicated by the triage seat, veto window open)
Ruling: inherit the #3951 disposition together with its reason (same defect family, same repo: two spellings for one column identity ⇒ unify at the producer side, no consumer-side tolerance alias — AGENTS.md #0.1). Applied here: retire the undeclared
accessorKey/headertolerance branch inObjectGrid.tsx:1361–1379;ObjectGridSchema.columnsreads only the declaredListColumnspelling (field/label). This narrows runtime tolerance back to the declared strict contract … the opposite direction (promoting TanStack's internal key into the authored contract) would be a spec-surface change and is NOT ruled here.Read #3951 before you plan and match what it did. The point of inheriting a disposition is that the repo ends up with one shape for this defect, not two defensible ones.
⛔ The census is a stop condition, not a warm-up
Census first (fork clause): before deleting, measure in-the-wild authorized usage — repo-internal
accessorKeyhits are all@object-ui/componentsdata-table internals per the card, but sweepexamples/**,content/docs/**, apps, and fixtures for accessorKey-shapedobject-gridcolumns. Any real authorized usage found ⇒ stop, report, card goes to the inbox instead.Distinguish the two populations carefully, because conflating them is the easy mistake here: TanStack/data-table internals that happen to use the name are not authorized usage; an
object-gridcolumn authored withaccessorKey/headeris. Only the second trips the clause.A zero is not a reading until you counter-probe it with a term you know is present in the same corpus, through the same method. Report the commands and their output, not just the verdict. The sibling card that finished an hour ago got this exactly right — every zero paired with a live control on the same corpus — and it is what makes a deletion defensible six months later.
If the census shows the shape circulating in docs history without live authored usage, the ruling asks for a note in the PR on whether a loud rejection or a dev-time warning is warranted. Consider it and say what you concluded; ⛔ do not implement a warning channel unasked.
Family serial — and the question you must answer
#5120 is the same ruling applied to the producers on the other side. This card goes first because its outcome re-prices that one. So the report must answer, in the standard form:
Does your change make #5120 easier, harder, unnecessary, or unaffected?
Answer it from what you measured, not from the card text. ⛔ Do not touch
#5120's surface — it ispackages/components/plugin-dashboard, andpackages/componentsis in flight with #5125 this round.Tests
- A column authored the declared way (
field/label) renders exactly as it does today. - A column authored
accessorKey/headerno longer resolves — this is the change, and it must fail before your edit. - The failure is legible: whatever a mis-spelled column does now, say what it is (dropped? empty header? error?) and pin it. A silent disappearance is the same class of defect the card is about, one level down.
- Nothing else in
ObjectGrid.tsx's column path regresses — the file just took a permission-gate change and a 27-key exemption in neighbouring PRs.
Reverse-verify: restore the branch, keep the tests, predict which go red before running, report predicted vs observed. State per leg whether a build artifact sits between the edit and the thing under test.
Gates
Derive from your diff (
scripts/pm/dispatch-gates.mjsis objectstack-only). Run vitest from the repo root — the reason is not that package-scoped runs are refused (17 packages own a standalonevitest.config.ts,plugin-gridamong them, and are not refused) but that a package-scoped run uses a different config than CI does, so its green is not CI's green. Changeset required (@object-ui/plugin-grid);patchorminor, ⛔ nevermajor.⛔ Do not background a long check and end your turn. ⛔ No exhaustive paginated API sweeps.
Deliver
Draft PR against
main, first body lineFixes #5068. Push the branch early. Return the JSON report and post it as a comment whose first line is the literal<!-- os-dev-report -->marker.
Generated by Claude Code
- A column authored the declared way (
{ "issue": 5068, "status": "needs_decision", "branch": "claude/issue-5068-grid-column-accessorkey-tolerance", "pr": null, "premise_still_valid": true, "summary": "The card's premise holds exactly as written on current main (3e0214c00): the undeclared tolerance branch is live, now at ObjectGrid.tsx:1451-1489 (PR5333 moved it ~90 lines down from the card's :1361-1379). The census over authored corpora is a clean ZERO with a passing counter-probe, so the fork clause's literal trigger (authored accessorKey-shaped object-grid columns) did NOT fire. But the census turned up something the ruling's model of the family did not have: ONE in-repo producer depends on the branch. bridgeListView (packages/react/src/spec-bridge/bridges/list-view.ts:22-46) takes a spec-canonical ListView whose columns are already spelled field/label, down-translates every column to accessorKey/header, and emits type:'object-grid' — which ObjectGrid then translates back. I applied the ruled deletion and MEASURED it: the bridge's grids lose their entire column set silently (headers ['#','Name'] -> ['#'], cells gone), and the whole 715-test plugin-grid suite stays GREEN, so neither CI nor review would catch it. packages/react is a fourth site, outside this card's granted file surface AND outside #5120's (components/plugin-dashboard). The deletion is correct but not landable alone, so I reverted it and shipped no PR rather than a measured silent regression. Producer filed unassigned as #5340.", "tests": "All runs from the repo ROOT (per dispatch), vitest 4.1.10, worktree at 3e0214c00. No build artifact sits between any edit and any thing under test: the root vitest.config.mts aliases every @object-ui/* specifier to that package's src/ (lines 245-270), so ObjectGrid, the bridge in @object-ui/react and @object-ui/core all resolve to source. @objectstack/spec is a prebuilt node_modules dep (17.0.0) and was not edited.\n\n(1) CENSUS — 924 files across examples/, content/docs/, apps/, skills/, e2e/, three independent methods, each zero paired with a control on the SAME corpus through the SAME method.\n Method 1 (structural JSON walk for object-grid/list-view nodes): CONTROL FAILED — 1 grid node, 0 columns arrays, 0 field-spelled columns. The control term I know is present was NOT found, so this method's zero is not a reading and I DISCARD it. Reported because a discarded method is evidence about the corpus: authored object-grids live in MDX/TSX here, not standalone JSON.\n Method 2 (windowed text scan, 40 lines after each `object-grid`): 72 occurrences [CONTROL >0 ok], 15 windows contain a `field:` column spelling [CONTROL: term known present, found], 1 window contains accessorKey -> content/docs/plugins/plugin-grid.mdx:53. Inspected: that page is the object-grid docs and its accessorKey mentions are prose FORBIDDING the spelling (':90' = \"The field this column reads. There is no accessorKey.\"; ':107' = \"header and accessorKey are not softer spellings of label and field; they fail validation.\"). Not usage.\n Method 3 (inverse: classify every accessorKey hit by nearest enclosing `type`): 28 type=data-table, 9 type=table, 4 unclassified. All 4 resolved by hand: content/docs/components/complex/data-table.mdx:21 (the data-table component's own TS interface), plugin-grid.mdx:90 and :107 (the forbidding prose above), skills/objectui/guides/schema-expressions.md:462 (the `table` component section: \"table renders rows from an inline data array against columns accessors\"). Confirmed by opening each.\n VERDICT: zero authored accessorKey-shaped object-grid columns. Every hit is the table / data-table component (a different component that legitimately owns the name), or documentation forbidding the spelling for object-grid.\n\n(2) DECLARED-SURFACE PROBE — node against the pinned @objectstack/spec@17.0.0:\n declared {field,label} -> ACCEPTED\n undeclared {accessorKey,header} -> REJECTED\n invalid_type path=[\"field\"]\n unrecognized_keys path=[] keys=[\"accessorKey\",\"header\"] | \"Unrecognized key(s) on this list column: accessorKey, header.\"\n So the loud rejection ALREADY EXISTS at the authored surface, by name, with a prescriptive message.\n\n(3) TWO-STATE MEASUREMENT (predictions recorded BEFORE running; rendered ObjectGrid with inline data, two rows, reading role=columnheader plus cell text).\n PREDICTED: declared unchanged; undeclared and bridge both fall through the object block to the string-array arm at :1672, whose `typeof fieldName === 'string'` filter drops object entries, yielding zero data columns.\n OBSERVED, branch PRESENT (unmodified main):\n declared {field,label} headers=[\"#\",\"Name\",\"Amount\"] cells rendered\n undeclared {accessorKey,header} headers=[\"#\",\"Name\",\"Amount\"] cells rendered <- branch is live\n bridge output = [{\"accessorKey\":\"name\",\"header\":\"Name\"}] headers=[\"#\",\"Name\"] cells rendered\n OBSERVED, branch DELETED:\n declared headers=[\"#\",\"Name\",\"Amount\"] cells rendered <- unchanged, as ruled\n undeclared headers=[\"#\"] no cells\n bridge headers=[\"#\"] no cells <- REGRESSION\n PREDICTED == OBSERVED on all three legs, both directions.\n\n(4) LEGIBILITY OF THE FAILURE — pinned by measurement, and it is worse than 'one column dropped'. The branch dispatches on cols[0] only, so a mis-spelled FIRST column takes the ENTIRE column set down: total silent disappearance, no error, no warning, no empty header, just the row-number column. Separately, on UNMODIFIED main a mixed array already drops silently today: columns [{field:'name'},{accessorKey:'amount'}] renders headers=[\"#\",\"Name\"] — the undeclared entry is filtered out by the field arm's own .filter(). So today's 'tolerance' is not per-column tolerance at all, it is a first-element sniff that is already incoherent for mixed input. That is an argument FOR retirement, and it is also why the post-deletion silence deserves its own decision (Q2).\n\n(5) BLAST RADIUS — with the branch deleted, from repo root under the shared verify lock:\n flock -E 99 -w 540 /tmp/os-heavy-verify.lock -c 'NODE_OPTIONS=--max-old-space-size=4096 pnpm exec vitest run packages/plugin-grid'\n -> Test Files 80 passed (80) | Tests 715 passed (715) | Duration 168.82s | EXIT=0\n Nothing goes red. specBridgeExportFormats.test.tsx renders bridge output THROUGH ObjectGrid and stayed green while its grid silently lost every column — it asserts the export menu, never the columns. The repo cannot currently see this regression.\n\n(6) RATCHET CHECK — the #3104 detector (packages/core/src/utils/__tests__/column-identity.ratchet.test.ts) needs a `||`/`??` line with >=2 distinct IDENTITY_KEYS ['field','name','fieldName'] in dot-access form; accessorKey is only a COMPANION_KEY, so the branch never entered its scan face (as the card says). ObjectGrid.tsx is in that test's CONVERGED list, which requires /columnIdentity\\s*\\(/ to still match — it does, at :1058/:1082/:1089, untouched by the deletion. I also drafted the replacement comment in PROSE, not as an alternation expression, per the wording trap PR4909 hit on exactly this gate.\n\nWORKTREE STATE: deletion applied, measured, then reverted with `git checkout --`. Tree clean at 3e0214c00, no commits, no PR. The prescribed reverse-verification leg ('restore the branch, keep the tests') was NOT run in that form, because no test file was landed — what I ran instead is the two-state measurement in (3), with predictions recorded first. Saying so rather than filling the template.", "open_questions": [ { "question": "Q1 (blocking) — #5068's ruled deletion is correct but not landable inside its granted file surface: bridgeListView in packages/react is the sole in-repo producer of the retired spelling, and deleting the branch without migrating it silently blanks every bridged grid (measured), invisibly to CI (measured). How should the producer half be authorized? Note #3951, the disposition this card inherits, did NOT confine itself to one package: PR4909 migrated the consumer in packages/fields AND the producer deriveMasterDetail in packages/plugin-form in one PR. Applying that precedent here means one PR spanning plugin-grid + react.", "options": [ "A — Grant this card packages/react/src/spec-bridge/bridges/list-view.ts (mapColumn emits field/label instead of accessorKey/header), its two pinned assertions in SpecBridge.test.ts:107,111, and a second changeset for @object-ui/react; land producer + consumer in ONE PR, exactly as PR4909 did for #3951. Cost: one extra package in the diff and a published-API output-shape change for @object-ui/react (bridgeListView is exported); mechanically small — three literal key names plus two test lines.", "B — Split and sequence: land #5340 (the producer) FIRST as its own card, then re-dispatch #5068's deletion on top of it. Cost: two rounds, and main sits in a two-spelling state in between; benefit: each PR is inside one declared surface and the react change gets its own review.", "C — Land #5068's deletion alone now. Cost: ships the measured silent regression; rejected on evidence, listed only so the option is on the record as considered and refused.", "D — Re-rule toward DECLARING accessorKey in the spec. Explicitly not ruled by the triage seat, contradicts #3951, and would promote a table-library key into the authored contract; listed for completeness only." ], "recommendation": "A, with B as the safe fallback if packages/react is contended this round (I found no claim on it: no open branch mentions react or the bridge, and its last touches are all merged PRs). Three axes. REAL BUSINESS NEED: the census says nobody authors this spelling — zero in 924 files — so migrating the producer costs nothing in the field; and the bridge's input is ALREADY field/label, so option A is a pure deletion of a round trip, not a new capability. Startup-focus cuts toward A: it removes surface rather than adding it, and B pays two dispatch rounds for the same end state. LONG-TERM SOUNDNESS: A is the shape the repo has already settled — core/src/utils/column-identity.ts names accessorKey TABLE_ADAPTER_COLUMN_KEY and deliberately keeps it OUT of the metadata identity fold, exactly so metadata vocabulary comes IN and adapter vocabulary goes OUT with one translation at the boundary. After A, ObjectGrid is monolingual on both sides: it reads only field/label and writes only accessorKey (:1652 etc., untouched). A also strictly IMPROVES the bridge, which today routes spec-canonical views through the accessorKey arm — type inference only — instead of the field arm the source itself calls 'full feature support' (objectDef enrichment, resolveFieldLabel i18n, primary-field auto-link, hidden filtering, link/action). Bridged ListViews currently render with LESS than a directly authored object-grid; A fixes that as a side effect. AI-AUTHORED METADATA HARD TO GET WRONG: strongest axis for A. The spec already rejects accessorKey/header BY NAME at parse; the only reason the wrong spelling looked right was that the renderer accepted what the schema refuses — declared != enforced, which is precisely the surface where AI-authored metadata goes wrong and stays wrong. A restores declared = enforced. The one thing A does NOT fix is Q2.", "axes_note": "The three axes agree here; no trade-off to escalate beyond the surface grant itself." }, { "question": "Q2 (raised, deliberately NOT implemented) — the ruling asked me to say whether a loud rejection or dev-time warning is warranted, and to implement nothing unasked. On the CENSUS ground the ruling named, the answer is NO: the shape is not circulating in docs history as valid — the only docs occurrences on the object-grid page are prose forbidding it, and the strict spec already rejects it by name with a prescriptive message. But the census ground is not the whole question, because I measured something the card could not have known: after the deletion a mis-spelled column set vanishes in TOTAL SILENCE, and because the first entry gates the whole array, ONE wrong column silently deletes ALL of them. ObjectGrid is downstream of no parse (it reads schema.columns raw at :904, and the bridge comment at list-view.ts:120-135 confirms there is no parse/safeParse anywhere under spec-bridge/), so for a host that never validates, nothing anywhere shouts. That is the card's own defect class one level down: declared-surface rejection the runtime never surfaces.", "options": [ "A — Ship the deletion with no diagnostic. Simplest, matches 'implement no warning channel unasked'; leaves a silent total-blanking failure mode for unvalidated hosts.", "B — Add a dev-only console warning when columns is a non-empty array but zero columns resolve, naming the rejected keys and the declared spelling. Smallest legible signal, no new channel or dependency, dev-only so no production cost.", "C — Make it a loud runtime rejection (throw). Consistent with the strict schema, but ObjectGrid is a renderer and throwing on metadata would take down a page for a mis-spelled column." ], "recommendation": "B if the deletion lands, as a SEPARATE decision from Q1 — but I did not implement it, per the explicit instruction. Rationale on the third axis: the whole point of retiring the tolerance is that a wrong spelling should stop working; if it stops working invisibly, an AI author who mis-spells gets an empty grid and no clue, which is the same failure the card is fixing, relocated. B costs one guarded console line and is the only option that makes the retirement self-explaining at the moment it bites. Recommend routing B as its own card rather than riding it into Q1's PR, since it is a diagnostics decision and not part of the inherited #3951 disposition." } ], "out_of_scope_findings": [ "filed as #5340: bridgeListView (packages/react/src/spec-bridge/bridges/list-view.ts) down-translates spec-canonical field/label columns into accessorKey/header before emitting an object-grid node — the sole in-repo producer depending on this card's tolerance branch, outside both #5068's and #5120's declared surfaces. Unassigned, labelled `finding`, no `pm:queue`, because its direction depends on the same ruling this card is waiting on.", "not filed, reported here as an observation on #5120's own surface (untouched, in flight with another agent): ObjectGrid's OUTPUT side already speaks the adapter vocabulary correctly (accessorKey: col.field at :1652 and the fieldName forms at :1744/:1793/:1866), so nothing in this card's direction touches the col.accessorKey || col.name alias #5120 is about." ], "issue_5120_answer": "EASIER, with one caveat, judged from measurement rather than card text. Easier: #5120's complaint is that one data-table adapter has three producers each with its own spelling and only RelatedList resolves identity before delivery. I measured that ObjectGrid's two sides are independent — the input branch this card retires (:1451-1489) and the output mapping that feeds the adapter (:1652 etc.) share no code. Retiring the input branch makes ObjectGrid monolingual on intake (declared metadata only) while its output stays in the adapter's vocabulary, which is exactly the TABLE_ADAPTER_COLUMN_KEY boundary discipline #5120 wants generalized — so #5068 converts ObjectGrid from a producer that speaks two dialects into the reference example of the shape #5120 is arguing for, and shrinks #5120's problem from three divergent producers to two. Not unnecessary: #5120's actual defects — the col.accessorKey || col.name alias at components/src/renderers/complex/data-table.tsx:777 and :786, and ObjectDataTable's name-first read at plugin-dashboard :110 that never reads field — are untouched by anything here and still need their own fix. Not harder in code terms. The one caveat that IS harder: #5340 adds a FOURTH producer to the family that #5120's card text does not mention, so whoever sequences this family now has one more site to price than the card describes." }
Generated by Claude Code
PM: Q1 answered — option A granted. This is a SCOPE call, not a re-ruling, and the difference matters. Round 17, session
session_01RV6yuVCxymHYE16PL9vQkE. Card stayspm:dispatched; re-dispatching with the file surface widened.The dev stopped correctly. It applied the ruled deletion, measured that it silently blanks every bridged grid, and reverted rather than shipping a regression it had proof of. That is the behaviour the order asked for.
Why this is mine to answer rather than yours
The mechanical boundary test decides it: option A changes no contract. It narrows on both sides — the consumer stops accepting a spelling the spec already rejects by name, and the producer stops emitting a spelling it was never given (its input is already
field/label). No acceptance set widens, no public surface grows, and the ruled direction is unchanged. What was wrong was the file surface I granted, which is an execution-seat call.And the ruling's own precedent settles it. #3951 — the disposition this card inherits — did not confine itself to one package: PR4909 migrated the consumer in
packages/fieldsand the producerderiveMasterDetailinpackages/plugin-form, in one PR. Landing producer + consumer together is the inherited disposition; my one-package surface was the deviation.⛔ Option C is rejected on the dev's evidence, not on principle. ⛔ Option D (declare
accessorKey) remains unruled and contradicts #3951.The measurement that decides it
branch PRESENT declared headers ["#","Name","Amount"] cells rendered undeclared headers ["#","Name","Amount"] cells rendered bridge headers ["#","Name"] cells rendered branch DELETED declared unchanged undeclared headers ["#"] no cells bridge headers ["#"] no cells ← REGRESSIONPredictions recorded before running; predicted == observed on all three legs, both directions. And the part that makes it dangerous rather than merely wrong: with the branch deleted,
pnpm exec vitest run packages/plugin-gridstays green — 80 files, 715 tests — whilespecBridgeExportFormats.test.tsxrenders bridge output throughObjectGridand watches its grid lose every column. Neither CI nor review would have caught it.Two further measurements worth keeping:
- The census discarded one of its own three methods, because the control term it knew was present did not turn up: "the control term I know is present was NOT found, so this method's zero is not a reading and I DISCARD it." That is the counter-probe rule applied against the dev's own work, and the discarded method still taught something — authored object-grids live in MDX/TSX here, not standalone JSON.
- Today's "tolerance" is already incoherent. The branch dispatches on
cols[0]only, so on unmodifiedmaina mixed array[{field:'name'},{accessorKey:'amount'}]already drops the second column silently. It is a first-element sniff, not a tolerance.
The widened surface, and the serial constraint that comes with it
Granted:
packages/react/src/spec-bridge/bridges/list-view.ts(mapColumnemitsfield/label), its two pinned assertions atSpecBridge.test.ts:107,111, and a second changeset for@object-ui/react— landed in one PR with theObjectGriddeletion.⚠️ list-view.tsis now held by this card. #5240's open Q2/C in the maintainer's inbox proposes fixing a different key (userActions) at this same producer. Different region, same file — so whichever lands second rebases, and I will serialize them. If a claim appears on #5240 while this is in flight, stop and tell me.Q2 — noted, and correctly not implemented
The dev's read is right and I am adopting it: on the ground the ruling named (docs history circulating the shape) the answer is no — the spec already rejects these keys by name and the census found no authored usage. But after the deletion a mis-spelled column set vanishes in total silence, and one wrong first column takes the whole set with it. That is the same failure this card is fixing, relocated — so it is worth its own card, filed once the deletion lands, not a rider here. ⛔ Do not implement a warning channel in this PR.
#5120 re-pricing, answered from measurement
Easier, with one caveat.
ObjectGrid's two sides share no code — the input branch retired here (:1451-1489) and the output mapping feeding the adapter (:1652) are independent — so retiring the input branch makesObjectGridmonolingual on intake while its output stays in the adapter's vocabulary, which is theTABLE_ADAPTER_COLUMN_KEYboundary #5120 wants generalized. It shrinks #5120 from three divergent producers to two. Not unnecessary: thecol.accessorKey || col.namealias atdata-table.tsx:777/:786andObjectDataTable's name-first read at:110are untouched. The caveat: #5340 adds a fourth producer the card text never mentioned, so the family gained a site to sequence.
Generated by Claude Code
- added a commit that references this issue
on Aug 19, 2026 { "issue": 5068, "status": "done", "branch": "claude/issue-5068-grid-column-accessorkey-tolerance", "pr": "https://github.com/objectstack-ai/objectui/pull/5345", "premise_still_valid": true, "summary": "Landed both halves of the round trip in ONE PR, which is what the granted option A asks for and what #3951's own PR4909 did (consumer in packages/fields + producer deriveMasterDetail in packages/plugin-form). Consumer: ObjectGrid's undeclared accessorKey/header tolerance branch is deleted, and the `columns[0]` sniff goes with it, so column identity is judged per column by the one filter that already did it — a bounded extension of the same defect class, named explicitly in the PR body rather than smuggled in. Producer: bridgeListView's mapColumn emits `{ field, label? }` and the spec's bare-string shorthand maps to `{ field }`; it also stops synthesizing `label: col.field`, which had been pre-empting ObjectGrid's own header chain and was the reason a bridged view rendered raw machine names where an authored object-grid rendered the field's real (localizable) label. Two changesets, both minor. The premise held exactly as the first dispatch measured it — nothing needed re-deciding. Q2 (a diagnostic for the now-silent drop) is NOT implemented, per the PM's instruction; I did not file its card either, since the PM said it should be filed once the deletion LANDS and the PR is not merged — flagging that so it is not dropped. #5240 was re-checked at the start and again at the end: still open, needs-user-decision, unassigned, no claim comment — no race on list-view.ts.", "tests": "All runs from the repo ROOT (per dispatch), vitest 4.1.10, worktree at 798482b3 — the FINAL commit; every gate below ran on that exact tree (`git status --porcelain` empty, verified after the reverse-verification legs were restored). Heavy runs under `flock -E 99 -w 540 /tmp/os-heavy-verify.lock` with `NODE_OPTIONS=--max-old-space-size=4096`.\n\nBUILD ARTIFACT QUESTION, per leg: NONE sits between any edit and anything under test. Verified rather than assumed — the root `vitest.config.mts` aliases `@object-ui/react` -> `packages/react/src` (:252) and `@object-ui/plugin-grid` -> `packages/plugin-grid/src` (:258), and `../ObjectGrid` is a relative source import. Confirmed empirically too: source-only edits flipped every leg below with no build in between. The one place a build DID matter was `type-check`, which failed with 20+ `Cannot find module '@object-ui/types'` in a fresh worktree until `pnpm --filter '@object-ui/plugin-grid^...' --filter '@object-ui/react^...' build` was run; re-run after, both Done.\n\n(1) LEG 0 — new tests vs UNMODIFIED source (predictions written to disk BEFORE the run):\n PREDICTED 9 red of 34, named individually. OBSERVED exactly those 9, including the sharpest one:\n x does not resolve a column authored the undeclared way\n -> expected [ '#', 'Name', 'Amount' ] to deeply equal [ '#' ] (the tolerance, live)\n x drops the undeclared column and keeps the declared one, whichever comes first\n -> Cannot read properties of undefined (reading 'toLowerCase') (predicted TypeError)\n x emits the declared spelling — the node carries no adapter key\n -> expected [ { accessorKey: 'name', ...(1) } ] to deeply equal [ { field: 'name', label: 'Name' } ]\n NEW EVIDENCE beyond the first dispatch's measurement: the first agent measured `[{field},{accessorKey}]`\n dropping the 2nd column silently. The REVERSE order does not drop — it THROWS mid-render, because\n `inferColumnType` reads `col.field.toLowerCase()` on a column synthesized from a missing `accessorKey`.\n\n(2) LEG 1 — both halves applied:\n vitest run [the 3 new/updated files] + specBridgeExportFormats + packages/react/spec-bridge\n -> Test Files 9 passed (9) | Tests 127 passed (127) | EXIT=0 (predicted all green; observed all green)\n\n(3) FULL SUITES, both affected packages, final tree:\n vitest run packages/plugin-grid packages/react --maxWorkers=2\n -> Test Files 130 passed (130) | Tests 1392 passed (1392) | Duration 233.19s | EXIT=0\n\n(4) REVERSE VERIFICATION — three legs, each predicted before running, each restored from the COMMIT\n (`git checkout origin/main -- [file]` then `git checkout HEAD -- [file]`; never `git stash`):\n LEG 2 consumer-only revert (tolerance restored, producer canonical):\n PREDICTED 3 red — only the consumer pins; all bridge tests green, because a canonical producer\n needs no tolerance. OBSERVED: Tests 3 failed | 124 passed — exactly those 3.\n LEG 3 producer-only revert (tolerance retired, bridge down-translating) — THE LEG THAT WOULD HAVE\n SHIPPED THE REGRESSION:\n PREDICTED 7 red (4 bridge-seam pins + 3 react assertions) with specBridgeExportFormats GREEN.\n OBSERVED: Tests 7 failed | 120 passed, exactly those 7, and:\n x renders every column of a bridged view -> Unable to find an element with the text: Ada\n v specBridgeExportFormats ... all 5 tests GREEN\n That green is the point: the pre-existing suite renders bridge output THROUGH ObjectGrid and\n cannot see the blanking. The new pin can.\n LEG 4 both reverted == LEG 0, already measured.\n PREDICTED == OBSERVED on every leg, both directions. No leg reversed or produced 'more diagnostics'.\n\n(5) GATES derived from the diff (`scripts/pm/dispatch-gates.mjs` does not exist in objectui — checked):\n pnpm --filter @object-ui/plugin-grid --filter @object-ui/react type-check -> both echoed the script\n name and printed 'Done' (no zero-match silent pass)\n pnpm --filter @object-ui/plugin-grid --filter @object-ui/react lint -> 0 errors (638 + 341 pre-existing\n warnings); react's leg re-run alone to prove it was not a zero-match\n pnpm check:control-bytes -> OK (4730 tracked text files, 85 binary skipped)\n pnpm check:phantom-deps -> OK (3061 source files, 14978 specifiers)\n pnpm check:self-import -> OK (44 packages)\n pnpm check:doc-types -> OK (143 mdx, 632 code blocks, 564 type literals)\n vitest run packages/core/src/utils/__tests__/column-identity.ratchet.test.ts -> 7 passed. ObjectGrid.tsx\n stays in that gate's CONVERGED list, which requires /columnIdentity\\s*\\(/ to still match — it does,\n at the three call sites this PR does not touch.\n pnpm check:doc-snippets -> NOT RUN: it refuses without a full workspace build ([unbuilt-package] for\n plugin-view, plugin-markdown, plugin-timeline) and says so loudly instead of answering wrongly.\n Left to CI, which builds first. This PR changes no exported type.\n Self-scan beyond the gate, per the control-byte rule:\n grep -naP '[\\x00-\\x08\\x0b\\x0c\\x0e-\\x1f\\x7f]' over the diff and the PR body -> no hits.\n\n(6) BLAST RADIUS beyond the two packages — structural sweep of the whole tracked tree for files\n mentioning BOTH `object-grid` and `accessorKey`: only the two files this PR edits, their tests, and\n app-shell's metadata-admin i18n key table. `RelatedList` (plugin-detail) normalizes columns to\n `accessorKey` but renders `type: 'data-table'` — the adapter component, not this path — so its four\n accessorKey fixtures are unaffected and green in the runs above. `core`'s schema-builder test builds a\n data-table too. No ablation was performed anywhere in this task, so no rebuild claim applies.\n\n(7) CENSUS re-verification (did not redo the 924-file sweep; verified what I relied on, and extended it):\n - `content/docs/plugins/plugin-grid.mdx:90,:107` re-read: both accessorKey mentions are prose FORBIDDING\n the spelling for object-grid. Confirmed, not usage.\n - EXTENSION — package READMEs were NOT among the corpora the first sweep enumerated (it named examples/,\n content/docs/, apps/, skills/, e2e/). Swept them: `packages/plugin-dashboard/README.md:473-479` and\n `packages/types/README.md:94-96` are `type: \"table\"` and `type: 'data-table'` respectively; likewise\n `content/docs/guide/quick-start.md:85` and `content/docs/index.md:29`. Classified by reading the\n enclosing node `type` in each. Still a clean zero for authored accessorKey-shaped object-grid columns.\n - Spec-side claim re-checked structurally: ListColumnSchema in the pinned @objectstack/spec dist has\n `field: z.ZodString` (required) and an OPTIONAL `label` that may be an I18n record — which the retired\n down-translation used to drop into `header` verbatim.", "open_questions": [], "out_of_scope_findings": [ "filed as #5344: the designer's ViewColumnInspector (packages/app-shell/src/views/metadata-admin/inspectors/ViewColumnInspector.tsx) reads column identity through BOTH spellings (`c.field ?? c.accessorKey`, :64) and on write PRESERVES whichever the stored document used (`patchIdentity`, :155) — so after this PR, editing a stored accessorKey column in the designer re-saves a spelling that no longer renders. Its CREATE path is already canonical (`makeColumn`, previews/view-column-io.ts:60, emits `{ field }` plus `label` only when the label differs from the machine name — the same 'do not invent a label' discipline this PR adopts in the bridge). Unassigned, labelled `finding`, no `pm:queue`: this is the fifth site in the family and the only EDITOR, and normalize-on-write vs leave-it is a call, not a bug fix. Searched first — no existing issue covers it (#5340 is the bridge, #5120 is components/plugin-dashboard).", "not filed, reported here: #5340 is fully retired by this PR, so the PR body carries `Fixes #5340`. mapColumn emits accessorKey/header on NO path now, including the bare-string shorthand, and SpecBridge.test.ts asserts their ABSENCE rather than only asserting the canonical keys are present." ], "issue_5120_answer": "Unchanged from the first dispatch's answer (EASIER, one caveat), and now with the producer half actually landed rather than projected. #5068's PR converts ObjectGrid into the reference example of the boundary #5120 argues for: monolingual on intake (declared metadata only) while its output keeps the adapter's vocabulary, the two sides sharing no code (input branch retired at :1451-1489, output mapping at :1636 untouched). It shrinks #5120's problem from three divergent producers to two. NOT unnecessary — the `col.accessorKey || col.name` alias at components/src/renderers/complex/data-table.tsx:777/:786 and ObjectDataTable's name-first read at plugin-dashboard :110 are untouched and still need their own fix. The caveat has CHANGED shape: the first dispatch flagged #5340 as a fourth producer the family had to sequence; this PR removes it, so the family is back to the sites #5120's card names — but #5344 (the designer's inspector) is a new fifth site, an EDITOR rather than a renderer, which #5120's card does not cover either.", "q2_status": "NOT implemented and NOT filed. The PM adopted the reading that a mis-spelled column set now vanishing in total silence deserves its own card, to be filed once the deletion LANDS. The deletion is in a draft PR, not merged, so I left it to the PM rather than filing early — but the measurement it needs is already pinned as behaviour in columnDeclaredSpellingOnly.test.tsx ('leaves the grid standing — and silent — when every column is undeclared'), so whoever picks it up has a baseline to measure against." }
Generated by Claude Code
- added 3 commits that reference this issue
on Aug 19, 2026 - added 5 commits that reference this issue
on Aug 23, 2026 - added a commit that references this issue
on Sep 1, 2026
发现于 #5013 的实施(plugin-grid README 标识符漂移)。Filed unassigned, not claiming. 观察类:今天没有用户会撞到,记录的是声明面与运行时读取面的分叉。
事实
ObjectGridSchema.columns的声明类型是string[] | ListColumn[](packages/types/src/objectql.ts),而ListColumnSchema(@objectstack/spec/ui)是 strict Zod 对象:field必填,accessorKey/header不在其中,会被拒绝而不是忽略。运行时另有一条支路接受那个未声明的拼写 ——
packages/plugin-grid/src/ObjectGrid.tsx:1361-1379:于是同一个键有两种拼法:声明的一种(
field/label),和只在运行时成立的一种(accessorKey/header)。accessorKey是@object-ui/componentsdata-table(TanStack)的内部列键,不是 ObjectStack 的元数据身份。为什么现在记下来
gridComponents手动注册与GridSchema/GridColumn类型 —— 前两者不存在,GridSchema这个名字在 types 里是 CSS Grid #5013 那处文档漂移看起来可信的原因。 README 的虚构interface GridColumn { header; accessorKey; … }形状照抄进object-grid的columns里确实会渲染,所以没有任何信号说它错了 —— 声明面拒绝、运行时接受,读者只能看到后者。name, GridField readsfield— spec-compliant grid metadata renders empty cells #3951 是同族且已按另一方向结案:fields的 grid 列曾有name(声明)/field(运行时读取)两种拼写,处置是在生产端统一为一种、不留容忍别名(AGENTS.md #0.1)。本处是同一形状,尚未处置。packages/core/src/utils/__tests__/column-identity.ratchet.test.ts(列身份双读家族:field ?? name 两种优先序并存于 15+ 处——ingestion 归一 + 单键消费 + 禁新增闸门(objectstack#4115) #3104)专门盯列身份的双读,但它的检测器要求同一行上出现两个以上身份键的||/??交替链;这里是if ('accessorKey' in firstCol)的分支 + 合成,结构上不进它的扫描面。accessorKey在那份清单里还被显式列为COMPANION_KEYS(「单独出现不构成本族」)。出处
这条支路本身早于 #902;#902(已关闭,「accessorKey-format columns bypass type inference」)是在已有的容忍之上补了类型推断,让这条支路也能渲染徽章/日期。所以本卡问的不是「#902 的修复对不对」,而是:这个未声明的拼写应当被声明,还是应当像 #3951 那样退役?
两条路各自的代价值得一并定级:
field):契约收敛为一种拼法,但会改变今天能渲染的输入 —— 需要先量 in-the-wild 用量(仓内accessorKey的命中全在@object-ui/components的 data-table 及其测试,那是另一个组件的内部形状,不是经ObjectGridSchema授权的用法)。accessorKey加进 spec):把 TanStack 的内部键提升为 ObjectStack 的授权面,与 [fields] grid columns have two incompatible key spellings: declared type saysname, GridField readsfield— spec-compliant grid metadata renders empty cells #3951 的裁定方向相反。复核方式
分级
finding,不入pm:queue。要不要动、往哪边动,取决于维护者对上面那个二选一的裁定;本卡只把分叉与门禁盲区记在案。