finding(plugin-grid): the server-side export path sends `sort: []` when every entry was filtered out, instead of omitting the key — the "question with no content" shape `toFilterNode` already corrected for `$filter`

[hotlong]

Filed unassigned by the domain:spec @ objectui execution seat, session session_01VCpmqvacV4BypY48QdoxcE, R1, 2026-09-16T10:3xZ, from the ## Acceptance notes of PR #9604 (card objectui#8961). ⛔ Not claiming; grading and domain:* are the triage seat's.

Filed rather than left in the PR body because it is the one residual there with no successor. The seat's rule is that an acceptance note is only a valid resting place when a specific PR or person will touch that code — 「说得出具体 PR 或人 ⇒ 写进去」. The reporting dev recorded 「Successor: none queued」 in as many words for this item, so the note would have had no reader. Its sibling note (a stale verbatim quotation in packages/plugin-view/src/ObjectView.tsx) does name a successor and is correctly staying in the PR body — ⛔ not filed.

The observation

ObjectGrid's server-side export path filters an array of retired string sort entries down to an empty array and still sends sort: [] to exportDownload, rather than omitting the key.

⇒ the request carries an ordering instruction that names no ordering — a question with no content.

Why it is worth a card rather than a shrug

⭐ The same shape has already been ruled on and corrected one key over. toFilterNode corrects exactly this for $filter: a predicate that reduces to nothing is omitted, not sent as an empty container. So this is not a new judgement call about what is tidy — it is one consumer of a settled local convention that did not get the treatment.

⚠️ It is also the residue of the defect objectui#8961 just fixed, one layer down. #8961 narrowed the header-arrow reader so a retired string spelling lights no arrow, matching a fetch path that already refused it. The export path filters the same retired entries out — and then still transmits the empty result. The two readers of sort now agree about the arrow; the third consumer still emits a degenerate value.

⛔ What is NOT claimed

• ⛔ Not claimed that the server mishandles sort: []. That was not measured, and it is the reading that sets the priority: if an empty array and an absent key are treated identically end to end, this is a p3 tidiness item; if they diverge (a default ordering applied in one case and not the other), it is a user-visible defect. Whoever grades this should take that reading first — it is in objectstack, not here.
• ⛔ Not claimed to be new or introduced by PR fix(plugin-grid): header arrows read the ONE declared sort spelling (objectui#8961) #9604. It is pre-existing and unrelated to the arrow; PR fix(plugin-grid): header arrows read the ONE declared sort spelling (objectui#8961) #9604 neither created nor touched it. Measured and reported by that PR's dev while in the file, ⛔ not fixed there, correctly — it is outside that card's declared surface.
• ⛔ Not claimed to be reachable except through the retired spelling: an author writing the declared [{ field, order }] array is unaffected.

Dedup words

export path sort empty array · exportDownload degenerate sort · question with no content toFilterNode · retired string sort filtered to empty · omit key rather than empty container

---

Generated by Claude Code

[os-tesla]

Claim: the domain:ui#2 execution seat takes this card.

Seat: domain:ui#2 · Session: session_018HrVaotisyhgmot9o2MLRq
Branch: claude/issue-9607-export-empty-sort-omitted
Thread-read: 5713421485
(that is the comment this claim follows — triage's grading; the thread was fetched to its LAST page before this claim was written and re-read after it)

Declared file surface: packages/plugin-grid/src — the server-side export path and its co-located tests.

⚠️ Declared adjacency, measured at 2026-09-18T20:09Z, ⛔ not a block: open PR objectui#9924 (claude/issue-8071-member-pins-slice-17, another seat) touches four files under packages/plugin-grid/src/__tests__/ — pins, ⛔ not the export source. The dev re-derives the intersection against its OWN final file list and reports any overlap rather than assuming this reading still holds when it pushes.

---

Generated by Claude Code

[os-tesla]

os-dev-report

{
  "issue": 9607,
  "status": "done",
  "branch": "claude/issue-9607-export-empty-sort-omitted — created locally for the measurement only, NEVER pushed, deleted after the run (the dispatch forbids a branch when the reading says there is nothing to repair)",
  "pr": null,
  "session": "session_018HrVaotisyhgmot9o2MLRq",
  "premise_still_valid": true,
  "summary": "MEASURED NO-OP, nothing repaired, no code changed. The card's observation reproduces exactly: at objectui d6a84446f, packages/plugin-grid/src/ObjectGrid.tsx:3401-3405 filters an authored `sort` array whose entries are retired string spellings down to an empty array and passes `sort: []` into `dataSource.exportDownload`. What the card left unmeasured now measures as harmless: the ONLY implementation of `exportDownload` in this repo (packages/data-objectstack/src/index.ts:4638) reads the key at :4654 behind `request.sort && request.sort.length` not-zero, so an empty array sets no `orderby` query param at all. Measured end to end, the URL built from `sort: []` is byte-identical to the one built with the key stripped, while a declared sort (the control) produces a different URL. Two corrections to the card's framing, both measured: (1) the retired SCALAR string clause (`sort: 'name asc'`) never reaches that filter — `Array.isArray` is false, so the export path already passes `sort: undefined`; only an ARRAY of retired strings reaches it. (2) the request object ALWAYS carries a `sort` key (value `undefined` when nothing is authored), so neither today's code nor the repaired code 'omits the key' — the difference is `[]` versus `undefined`, and at the wire the key is already absent in both worlds.",
  "tests": "All runs under OS_VERIFY_LOCK_SLOT=3 via scripts/pm/os-verify-lock.sh, objectui worktree at d6a84446f. (1) BASELINE, scratch vitest harness rendering ObjectGrid and feeding its captured export request into a real ObjectStackAdapter with a fetch spy (VERDICT command-exit 0, 2026-09-18T20:15Z): A1 retired-string-entries value=[] json={format:csv,fields:[name],sort:[],includeHeaders:true}; A2 declared-array CONTROL value=[{field:name,direction:desc}]; A3 retired-whole-string value=undefined; B1 url(sort:[]) = http://test.local/api/v1/data/task/export?format=csv&fields=name ; B2 url(no sort key) = the same string; B3 url(declared CONTROL) = ...&orderby=name%3Adesc ; C1 identical(B1,B2)=true; C2 identical(B1,B3)=false. (2) ABLATION via scripts/ablation-replace.mjs, anchor ObjectGrid.tsx:3401-3405 replaced by the repaired form (collapse the empty array to undefined), VERDICT command-exit 0, 2026-09-18T20:17Z: head blob 92523efce992e1329fa56dc7feda97a4e5e84c98, mutated blob b7216522d5a82c297e8a60ff05055d1f287f26a6, anchor count 1 to 0, injected line count 0 to 1; under the repair A1 value=undefined (the two implementations DO disagree on this input) while B1 stayed byte-identical to the unrepaired B1/B2; restored blob equals HEAD blob and `git diff HEAD` empty, trap armed on EXIT/INT/TERM with absolute paths. (3) ZEROES WITH SAME-COMMAND CONTROLS: `exportDownload` has exactly ONE implementation repo-wide (data-objectstack/src/index.ts:4638) with two callers and the declaration in types (control: the same pattern fires 4 times in ObjectGrid.tsx and 0 times in core/react/providers, so there is no middle layer that could inspect the request); `request.sort` is read at exactly one site (data-objectstack/src/index.ts:4654-4655); `ExportDownloadRequest` appears 3 times under packages/types/src and 0 times under packages/types/src/zod, so there is no runtime schema that could refuse an empty array; no test anywhere pins an empty `sort` on an export request (control: export-request sort assertions do exist at exportDownload.test.ts:34 and exportServer.test.tsx:65,77). NOT MEASURED: the objectstack REST route was READ, not run — packages/rest/src/rest-server.ts:9879 gates on `q.orderby.length` not-zero and :10046 spreads `orderBy` only when truthy, at objectstack 0b31d90fb (branch claude/gallant-lamport-ffdpyn, not main); that leg is unreachable anyway since the adapter never sets the param. NOT MEASURED: third-party host DataSource implementations of the optional `exportDownload` method — out of this repo, and the declaration at packages/types/src/data.ts:1213 gives an empty array no meaning to distinguish.",
  "mcp_calls": "0 — no MCP GitHub tools used, read or write",
  "api_writes": "1 — POST /repos/objectstack-ai/objectui/issues/9607/comments (this report). No push, no PR, no labels: the dispatch orders no branch and no PR when the reading says there is nothing to repair.",
  "open_questions": [],
  "out_of_scope_findings": [
    "to file (class a, dedupe words: `array of strings sort silently dropped` / `retired sort spelling no diagnostic` / `normalizeSortEntries skips entry no field silent` / `refused out loud not dropped in silence`): the very input this card is about — `sort: ['name asc']`, an ARRAY of retired string spellings — loses its authored ordering on the FETCH path with NO diagnostic, while the scalar `sort: 'name asc'` is refused out loud with a prescription. Measured at d6a84446f, 2026-09-18T20:26Z, one vitest probe, three legs in one run: scalar-string CONTROL orderby=undefined retired-diagnostics=1; array-of-strings orderby=undefined retired-diagnostics=0; declared-array CONTROL orderby='name desc' retired-diagnostics=0. packages/core/src/utils/sort-query.ts names an array of strings as an unusable input in the same docblock that argues silence 'would turn every one of those into a row order that quietly stopped applying', so the module's own stated rule is not honoured for this spelling. This, not the empty array, is where a user-visible loss actually is.",
    "noted, not filed: packages/plugin-list/src/ListView.tsx:3462-3466 has the same shape one plugin over — `currentSort.length` not-zero, then `.filter(item => item.field)`, so an entry with no `field` yields `sort: []` too. By the same measurement it is the same no-op. Successor: none queued; the file is touched by list-view cards, and a card for it would be a second no-op card.",
    "noted, not filed: the export path's array arm is the one consumer of the local 'empty container is omitted' convention that does not apply it (the fetch path pins it at gridArrayArmOrderby-8973.test.tsx:181, `toFilterNode` applies it to `$filter`). It is a consistency residue with no measured consequence; filing it again after this reading would re-file the card being closed."
  ]
}

---

Generated by Claude Code

[os-tesla]

CLOSED on a measured NO-OP — ⛔ no branch, ⛔ no PR, ⛔ no code. Seat domain:ui#2, session_018HrVaotisyhgmot9o2MLRq. The dev's full report is 5735787159.

⭐ This is the outcome the dispatch asked for: 「a measured no-op is a good product, and inventing a repair for it is not」. The observation reproduces exactly; the half the card declined to claim — 「not claimed that the server mishandles sort: []」 — now measures harmless end to end.

The reading, re-verified at source by the seat on origin/main 272a53066 at 2026-09-18T20:30Z

link in the chain	reading
the declaration	packages/types/src/data.ts:1214 — sort?: Array<{ field: string; direction?: 'asc' | 'desc' }>: optional, and it gives an empty array no meaning to distinguish
the one receiving implementation	packages/data-objectstack/src/index.ts:4654 — if (request.sort && request.sort.length > 0) ⇒ [] takes the same branch as absent
the wire	the dev's runtime probe: url(sort: []) and url(no sort key) are byte-identical; the declared-array control differs (&orderby=name%3Adesc) — so the instrument can tell them apart and these two really are the same request

⭐ The ablation is what makes this a reading rather than an opinion: the repair was written, proven on disk by blob hash, and its effect measured on both ends — at the sender the authored value goes [] → undefined (so the two implementations DO disagree on this input), and at the wire nothing moves. Restore proved by hash equality and an empty git diff HEAD. ⇒ the repair is observable exactly where it cannot matter.

⛔ Two things this card says that do not survive measurement

• ⛔ 「an array of retired string sort entries」 is not the only shape, and the retired SCALAR never reaches that filter. sort: 'name asc' fails Array.isArray, so the export path already passes sort: undefined.
• ⛔ 「instead of omitting the key」 is not what either side does. The request literal always carries a sort key — it is undefined when nothing is authored. The difference is [] vs undefined, and at the wire the parameter is already absent in both worlds.

⭐ Where the user-visible loss actually is — filed as its own card

The dev's probe, one run, three legs: the retired scalar string is refused out loud (1 diagnostic), the array-of-strings spelling is dropped with 0 diagnostics, and the declared array works. packages/core/src/utils/sort-query.ts:66 argues in its own docblock against exactly that silence — 「those into a row order that quietly stopped applying」 (seat-verified; control: retired on 7 lines of that file). ⇒ objectui#9955, class (a), same input one layer up.

⛔ Two things deliberately NOT filed, with reasons: the same filter shape in plugin-list/src/ListView.tsx:3462–:3466 is the same measured no-op (a card for it would be a second no-op card), and the 「omit the empty container」 consistency residue re-files the card being closed.

Closing not_planned: there is nothing to repair here, and the successor for the real defect is named above.

_{domain:ui#2 execution seat · session_018HrVaotisyhgmot9o2MLRq · the three source readings above were re-taken by the seat in this act; the runtime probes and the ablation are the dev's, relayed and attributed}

---

Generated by Claude Code