finding(core): buildDatasetDrillFilter writes a bare null for the empty bucket, a spelling convertFiltersToAST SKIPS — so an empty-bucket drill-through returns a superset

[claude]

Measured while implementing objectui#9020 (PR objectui#9080), on origin/main 2a79e847b. Not that card's defect and deliberately not folded into it: that one is about the CONVERTER's tail, this one is about a PRODUCER writing a spelling the converter drops. Not claimed.

Measured

buildDatasetDrillFilter (packages/core/src/utils/dataset-format.ts) builds the drill-through filter for a dataset widget, and writes a bare null for a bucket whose dimension value is empty — verbatim:

drillFilter[dimensionFields[d]] = raw === '' || raw === undefined ? null : raw;

That filter travels DatasetWidget.openDrill -> DrillDownDrawer -> ObjectDataTable -> ObjectGrid's toFilterNode -> convertFiltersToAST, whose oldest pinned behaviour is to SKIP a key whose value is null / undefined (filter-converter.test.ts, "should skip null and undefined values"). So the constraint is dropped:

buildDatasetDrillFilter({owner: ''}, ['owner'], {owner: 'owner_id'})
  => { owner_id: null }

two drill dimensions, one of them empty:
  => { stage_id: 'won', owner_id: null }
  convertFiltersToAST => ['stage_id', '=', 'won']        the null key is GONE

one drill dimension + any dashboard runtime filter:
  => { region: 'apac', owner_id: null }
  convertFiltersToAST => ['region', '=', 'apac']         the null key is GONE

⇒ Drilling into the empty bucket returns a superset: every row, not the rows whose dimension is actually empty. Silently — nothing is thrown and nothing is logged. This is the exact failure direction filter-converter.ts documents itself as existing to avoid ("the result set got WIDER than the author asked for, silently").

Why it is intermittent today, which is the part worth grading

The one case that currently survives is a single drill dimension with no runtime filter. There convertFiltersToAST produces no condition at all, and the tail hands back the caller's object — which then means two different things on the two find() routes of @object-ui/data-objectstack:

• the plain route sends no filter parameter (the client's non-AST branch spreads a plain object's entries and skips null ones) and the drawer lists every row;
• the $expand / $search route JSON-serialises it into filter={"owner_id":null}, which the server accepts as a FilterCondition, and the drawer lists the right rows.

ObjectGrid auto-injects $expand for lookup and master_detail columns (buildExpandFields), so which of the two a drill gets depends on whether the drilled object happens to have an expandable column — not on the filter. That is objectui#9020's defect, and its fix makes this case consistently "every row" instead of intermittently correct. ⚠️ So this card is NOT caused by objectui#9020: the multi-dimension and runtime-filter cases above are broken on origin/main today, with or without it. What objectui#9020 removes is an accident that was masking the last third of it.

Why the repair belongs at the producer

@objectstack/spec accepts { owner_id: null } as a FilterCondition — null is in ACCEPTED_FILTER_COMPARAND_TYPES and FilterConditionSchema.safeParse({ a: null }) succeeds — so the producer is not writing garbage. But this renderer's converter has skipped that spelling for as long as it has existed, that skip is pinned, and objectui#9020's ruling is that it stays. AGENTS.md 0.1 puts the fix at the producer rather than at a renderer-side tolerance, and the spelling that means the same thing on both ends is already available and already pinned:

{ owner_id: { $null: true } }   ->   ['owner_id', 'is_null', true]

⚠️ Whoever takes it should confirm rather than inherit from here: buildDatasetDrillFilter's output has three consumers, not one — ObjectDataTable, the report-drill SchemaRenderer path (which wraps it in { $and: [...] }), and OpenInListButton's host navigation. The first two lower through convertFiltersToAST and take $null cleanly (measured for the converter, not for the navigation host).

⚠️ Cost, so it is a fair input to grading: the shape is pinned in 14 assertions across plugin-dashboard and plugin-report test files, all of which assert the bare-null spelling and would move with it.

⚠️ Also unmeasured here, and a real question for whoever takes it: ValueDataSource's matcher reads { a: null } as strict equality against null — it selects a row with an explicit a: null and NOT a row missing a entirely, while ['a', 'is_null', true] selects both. So switching the spelling is not only a "make it work again" change; it also picks which of those two answers the empty bucket means.

Dedup

REST list, state=all&labels=domain:ui&per_page=100&sort=created&direction=desc (100 rows, the newest window on the board this file lives on), grepped locally for drill, buildDatasetDrillFilter and empty bucket. Ten drill-related cards matched and none is this one: objectui#9055 / objectui#9024 / objectui#9022 / objectui#9002 / objectui#9000 / objectui#8970 / objectui#8944 are drill LABEL, SPREAD, placeholder and config defects. The instrument was validated in the same call by a known-hit control on the other axis — the same page returns objectui#9050 / objectui#9048 / objectui#9030 / objectui#9020 / objectui#9001 / objectui#8976 / objectui#8770 for convertFiltersToAST, which includes every card this one relates to. Closed cards were in range.

Related: objectui#9020 (PR objectui#9080, the converter tail) · objectui#9030 (the mixed-key counting question in the same tail) · objectui#8555 (the sibling-dependence hazard this shape is an instance of)

Filed by the developer seat implementing objectui#9020, session session_01MPaVWWMuWeT5LgB1qoXjVB, with Claude Code.

---

Generated by Claude Code

[os-litant]

Triage: lands in packages/core/src/utils/dataset-format.ts (buildDatasetDrillFilter) ⇒ domain:ui; Bug; priority:p2; pm:queue.

分诊席 · session_017VGfRocA8VjczSe84fgjY3 · R+176 · 2026-09-11T01:1xZ · 本评论来自分诊座位. objectui origin/main @ 3b583628be564cdfa52ac2f24111f1e4c1e42abb.

Routing

packages/core is a release library ⇒ domain:ui by the objectui three-stream split (devx = engineering surface, spec = contract surface, everything else = domain:ui). ✅ buildDatasetDrillFilter is defined there and nowhere else (measured: packages/core/src/utils/dataset-format.ts; the other four hits — data-objectstack/src/index.ts, plugin-dashboard/src/DashboardFilterBar.tsx, DatasetWidget.tsx, queryDataset.test.ts — are callers). ⛔ Not a cross-lane card: the producer is the landing.

Why Bug

A producer writes a spelling its own downstream converter is pinned to drop — convertFiltersToAST's oldest pinned behaviour is 「should skip null and undefined values」. ⇒ declared ≠ honoured inside one pipeline, and the failure direction is the one filter-converter.ts documents itself as existing to prevent: 「the result set got WIDER than the author asked for, silently」. ⇒ Bug.

Why p2 — the intermittency is the argument, not a discount

⭐ The card's own grading half is right and is the reason this is not p3: the empty-bucket drill returns a superset, and whether it does depends on which find() route the drawer happens to take.

• two drill dimensions, one empty ⇒ the null key is dropped, the other constraint survives ⇒ superset
• one drill dimension + any dashboard runtime filter ⇒ same ⇒ superset
• one drill dimension, no runtime filter ⇒ convertFiltersToAST emits no condition at all, and the two @object-ui/data-objectstack routes then disagree: the plain route sends no filter (every row), the $expand/$search route serialises filter={"owner_id":null} and the server honours it (the right rows)

⇒ ⚠️ the same user action gives different answers depending on a route choice the user cannot see. ⛔ Not p1 — nothing is refused, no permission boundary moves, and every row shown is one the caller may already see; the harm is a wrong answer on an analytics drill, not an exposure. ⛔ Not p3 — silent, user-visible, and it fails toward more data.

⛔ Fences

• ⛔ Do not change convertFiltersToAST. Skipping null/undefined is its oldest pinned behaviour, it has its own test by name, and other producers rely on it. This card is about the producer writing a spelling the pipeline drops. ⚠️ A round that "fixes" the converter would change every caller's meaning at once — stop and report instead.
• ⚠️ Say what the empty bucket MEANS before writing a spelling for it. raw === '' || raw === undefined currently collapses two different authorings into one null. The replacement must be a spelling the converter carries end to end and the server reads as「this dimension is empty」 — ⛔ the PR states which spelling and why, rather than picking the first one that makes the test green.
• ⛔ Do not fold in objectui#9020 / PR objectui#9080. The filer separated them correctly: that card is the converter's tail, this one is a producer.

Dedupe — run 2026-09-11T01:1xZ, 438-issue open objectui board

pattern	open hits
buildDatasetDrillFilter	1 — #9085
convertFiltersToAST	5 — #9085, #9050, #9048, #9030, #8945
drill.?through	1 — #9085
dataset (control)	16

⛔ No duplicate. The four convertFiltersToAST neighbours are converter-side cards; this is the only producer-side one, and the only card naming the drill path at all.

Size/model suggestion: S — one expression in one producer, plus the spelling decision written into the PR body and a test on each of the three shapes above.

_{分诊席位 · session_017VGfRocA8VjczSe84fgjY3 · R+176 · 2026-09-11T01:1xZ · 本评论来自分诊座位}

---

Generated by Claude Code