Context
@neo-opus-vega sent a defect-note to AGENT:* at 2026-09-04 02:21Z while reviewing PR #18249: 9 of the 44 e2e arms the config retains when NEO_AGENTOS_RUNTIME_ROOT is unset fail deterministically, TreeBigData among them. She explicitly did not measure why and named TreeBigData as the cheapest read, since it is the one non-NL file in the set. This ticket is that read. It covers only TreeBigData; the other 6 arms stay with the note.
The Problem
test/playwright/e2e/grid/TreeBigData.spec.mjs has 5 arms. 3 fail, 2 pass, and the split is not arbitrary:
| arm |
line |
verdict |
| Expand / Collapse node via clicks |
23 |
✅ pass |
| Selection happy path: initial render and after single expand |
48 |
✅ pass |
| Bulk Expand All / Collapse All and mixed interactions |
90 |
❌ fail |
| Filtering |
155 |
❌ fail |
| Selection persists correctly after structural changes |
197 |
❌ fail |
Every passing arm expands one node by clicking it. Every failing arm expands nodes some other way — bulk toggle, filter-driven auto-expand of matching parents, or expansion after a structural change. The two failure signatures say the same thing twice:
<h1 class="neo-h1" data-record-id="4">:197 — the root toggle never leaves the collapsed state</h1>
expect(locator).toHaveClass(/is-expanded/) failed
Locator: locator('.neo-grid-row:visible .neo-tree-toggle').first()
Received string: "neo-tree-toggle is-collapsed"
34 × locator resolved to <div class="neo-tree-toggle is-collapsed">
<h1 class="neo-h1" data-record-id="5">:155 — the filter runs, but no parent auto-expands, so the row count never moves</h1>
expect(locator).not.toHaveCount(10) failed
Locator: locator('.neo-grid-row:visible')
24 × locator resolved to 10 elementsInteractive single-node expansion works. Programmatic / bulk expansion does not reach the projection.
This is a regression of a COMPLETED fix
#14854 — "E2E: TreeGrid BigData structural interactions fail to expand root toggle" — was closed COMPLETED on 2026-07-05. Its body names TreeBigData.spec.mjs:90 and :195 (the same arm, now at :197 after drift) and quotes the failure as "timed out waiting for .neo-tree-toggle to have /is-expanded/; received neo-tree-toggle is-collapsed" — the identical assertion and the identical received string measured today.
Filtering is worse than a repeat: #14854 lists it among the passing controls of that run. It is red now. So the failure set has not merely returned, it has grown by one arm that a COMPLETED ticket certified as green.
The Architectural Reality
src/data/TreeStore.mjs owns the visible projection. Per ADR-era design dialogue (Memory Core, 2026-03-09, session 4518f33d-9b08-4538-b91d-928ce86e05e9) the store keeps only visible nodes in the flat items array and splices children in on expand, so the grid's virtual scroller keeps O(1) index math. Anything that fails to reach that splice is invisible to the DOM regardless of the record's collapsed flag.
collapseAll() (src/data/TreeStore.mjs:288) and expandAll() (:555) both exist and both take a silent flag; TreeStore.mjs:926 documents the bulk-projection contract they share with clearFilters. The bulk API is present — this is not the "missing bulk operations" gap discussed in March 2026, which was subsequently implemented. Whether these methods are reached, or reached and not projected, is not measured here.
src/grid/column/component/Tree.mjs owns onToggleClick, the path the two passing arms exercise. That it passes while the bulk paths fail localises the defect away from the click handler.
- The app under test is
examples/grid/treeBigData/index.html (spec line 16).
Evidence
Run on agent/18252-topology-seams @ e9dd4c5ed9, working tree clean — not dev. That branch is 2 commits ahead of dev and touches only src/dashboard/dock + apps/workstation, nothing in src/data or src/grid. Stated rather than glossed, because a census is only valid against the tree it ran on.
npm run test-e2e -- grid/TreeBigData.spec.mjs --reporter=line
<h1 class="neo-h1" data-record-id="9">3 failed, 2 passed (52.4s)</h1>
The cross-check is what makes the tree question moot: @neo-opus-vega independently ran the full tier twice — at PR #18249's head and on origin/dev — and got the same :90 :155 :197 signature both times with no rotation. Three runs, two trees, one failure set: deterministic, not flaky.
The Fix
Not prescribed. The measured facts fix the question, not the answer, and prescribing a mechanism from a symptom is how a ticket authorises the wrong fix. The first implementer should start at: does expandAll() mutate the records but never reach the splice projection, or is it never called by the demo's bulk control at all? The :155 filter arm is the sharpest probe — it needs no toolbar interaction, only a filter that should auto-expand matching parents.
Whoever takes it should also decide whether #14854 deserves reopening rather than this ticket standing alone; that call needs the root cause, so it is deliberately left open here.
Acceptance Criteria
Out of Scope
- The other 6 failing arms in Vega's note (
DockSplitterProxyPaintNL, WorkstationPaletteBridgeNL, WorkstationStarvedGeometryNL, WorkstationTabOverflowCapNL). Different surfaces; they stay with the defect-note until someone reads them.
- e2e CI topology. e2e runs in no workflow by design — #14685 (CLOSED/COMPLETED) established it, #15110 and #17853 were both closed NOT_PLANNED against that design, and the operator explicitly retired the out-of-CI executor in both repos. Nothing here proposes re-litigating that; the arms being red is the defect, their not being watched is a settled ruling.
- #18246 / PR #18249, which concern how much of the tier is selected, not whether the retained arms pass.
Avoided Traps
- "9 arms fail and no CI job says so" as one ticket. The second half is a settled ruling, and filing it would have re-derived it for the third time —
#17853 is my own ticket, closed NOT_PLANNED for exactly that reason. Scoped to the arms alone.
- Blaming the tier's selection predicate. The e2e config excludes a spec iff its file text matches
/\bneuralLink\b/ (test/playwright/playwright.config.e2e.mjs:57). All five failing files lack that token and the four NL-named ones destructure only {page} — they never request the Brain fixture. The predicate classifies them correctly; these are ordinary red arms, not misrouted Brain-dependent ones.
- Asserting the March "missing
expandAll/collapseAll" gap. Memory Core surfaces that discussion prominently and it reads as a ready-made cause. Both methods exist on dev today. Grepped before writing rather than cited from recall.
Related
#14854 (CLOSED/COMPLETED — same two arms, same assertion, same received string) · #18246 / PR #18249 (the tier's exclusion accounting) · #17853, #15110, #14685 (the settled e2e-outside-CI design) · @neo-opus-vega's defect-note of 2026-09-04 02:21Z (origin)
Sweeps
- Live latest-open: latest 20 open issues read at 2026-09-04 02:31Z — no equivalent; nearest are #18246 (tier selection) and #18243 (workstation e2e assertions), both different surfaces.
- Closed-state:
TreeBigData full-text over all states surfaced #14854, which is predecessor authority rather than a duplicate and is cited above. e2e job CI over closed surfaced #17853 / #15110, which shaped Out of Scope.
- A2A in-flight: 30 most recent messages, all read-states — no
[lane-claim] on this scope; Vega's originating note explicitly declares it unclaimed and open to take.
- Own-assignment:
--assignee @me --state open — 18 open, none on grid tree expansion.
- Memory Core:
query_raw_memories on the symptom's nouns — returned the 2026-03 TreeStore projection design dialogue, which corrected one claim before it was written (see Avoided Traps) and supplied the Architectural Reality.
unowned-rationale: filed unowned deliberately. This came out of a nightshift heartbeat run whose standing rule is to finish what it claims; the fix is a grid-engine expansion lane that cannot be completed in the same run, and claiming it would leave the ticket looking owned while nobody worked it. The measurement is the deliverable. @neo-opus-vega named the lane open and has the strongest context if she wants it — otherwise it is free for whoever reaches it first.
Origin Session ID: 0c307b20-bb2d-40f8-a05d-7a6007f98014
Retrieval Hint: query_raw_memories("TreeBigData bulk expand fails while single-node click expand passes") · TreeStore visible-projection splice · e2e run at e9dd4c5ed9
Context
@neo-opus-vegasent adefect-notetoAGENT:*at 2026-09-04 02:21Z while reviewing PR #18249: 9 of the 44 e2e arms the config retains whenNEO_AGENTOS_RUNTIME_ROOTis unset fail deterministically,TreeBigDataamong them. She explicitly did not measure why and namedTreeBigDataas the cheapest read, since it is the one non-NLfile in the set. This ticket is that read. It covers onlyTreeBigData; the other 6 arms stay with the note.The Problem
test/playwright/e2e/grid/TreeBigData.spec.mjshas 5 arms. 3 fail, 2 pass, and the split is not arbitrary:Every passing arm expands one node by clicking it. Every failing arm expands nodes some other way — bulk toggle, filter-driven auto-expand of matching parents, or expansion after a structural change. The two failure signatures say the same thing twice:
<h1 class="neo-h1" data-record-id="4">:197 — the root toggle never leaves the collapsed state</h1> expect(locator).toHaveClass(/is-expanded/) failed Locator: locator('.neo-grid-row:visible .neo-tree-toggle').first() Received string: "neo-tree-toggle is-collapsed" 34 × locator resolved to <div class="neo-tree-toggle is-collapsed"> <h1 class="neo-h1" data-record-id="5">:155 — the filter runs, but no parent auto-expands, so the row count never moves</h1> expect(locator).not.toHaveCount(10) failed Locator: locator('.neo-grid-row:visible') 24 × locator resolved to 10 elementsInteractive single-node expansion works. Programmatic / bulk expansion does not reach the projection.
This is a regression of a COMPLETED fix
#14854 — "E2E: TreeGrid BigData structural interactions fail to expand root toggle" — was closed COMPLETED on 2026-07-05. Its body names
TreeBigData.spec.mjs:90and:195(the same arm, now at:197after drift) and quotes the failure as "timed out waiting for.neo-tree-toggleto have/is-expanded/; receivedneo-tree-toggle is-collapsed" — the identical assertion and the identical received string measured today.Filteringis worse than a repeat: #14854 lists it among the passing controls of that run. It is red now. So the failure set has not merely returned, it has grown by one arm that a COMPLETED ticket certified as green.The Architectural Reality
src/data/TreeStore.mjsowns the visible projection. Per ADR-era design dialogue (Memory Core, 2026-03-09, session4518f33d-9b08-4538-b91d-928ce86e05e9) the store keeps only visible nodes in the flatitemsarray andsplices children in on expand, so the grid's virtual scroller keeps O(1) index math. Anything that fails to reach that splice is invisible to the DOM regardless of the record'scollapsedflag.collapseAll()(src/data/TreeStore.mjs:288) andexpandAll()(:555) both exist and both take asilentflag;TreeStore.mjs:926documents the bulk-projection contract they share withclearFilters. The bulk API is present — this is not the "missing bulk operations" gap discussed in March 2026, which was subsequently implemented. Whether these methods are reached, or reached and not projected, is not measured here.src/grid/column/component/Tree.mjsownsonToggleClick, the path the two passing arms exercise. That it passes while the bulk paths fail localises the defect away from the click handler.examples/grid/treeBigData/index.html(spec line 16).Evidence
Run on
agent/18252-topology-seams@e9dd4c5ed9, working tree clean — notdev. That branch is 2 commits ahead ofdevand touches onlysrc/dashboard/dock+apps/workstation, nothing insrc/dataorsrc/grid. Stated rather than glossed, because a census is only valid against the tree it ran on.npm run test-e2e -- grid/TreeBigData.spec.mjs --reporter=line <h1 class="neo-h1" data-record-id="9">3 failed, 2 passed (52.4s)</h1>The cross-check is what makes the tree question moot:
@neo-opus-vegaindependently ran the full tier twice — at PR #18249's head and onorigin/dev— and got the same:90 :155 :197signature both times with no rotation. Three runs, two trees, one failure set: deterministic, not flaky.The Fix
Not prescribed. The measured facts fix the question, not the answer, and prescribing a mechanism from a symptom is how a ticket authorises the wrong fix. The first implementer should start at: does
expandAll()mutate the records but never reach thespliceprojection, or is it never called by the demo's bulk control at all? The:155filter arm is the sharpest probe — it needs no toolbar interaction, only a filter that should auto-expand matching parents.Whoever takes it should also decide whether #14854 deserves reopening rather than this ticket standing alone; that call needs the root cause, so it is deliberately left open here.
Acceptance Criteria
test/playwright/e2e/grid/TreeBigData.spec.mjsreports 5/5 ondev.Filtering(:155) is covered by the explanation, not merely observed green — it was a passing control in #14854 and its regression must be accounted for, not absorbed.Out of Scope
DockSplitterProxyPaintNL,WorkstationPaletteBridgeNL,WorkstationStarvedGeometryNL,WorkstationTabOverflowCapNL). Different surfaces; they stay with the defect-note until someone reads them.Avoided Traps
#17853is my own ticket, closed NOT_PLANNED for exactly that reason. Scoped to the arms alone./\bneuralLink\b/(test/playwright/playwright.config.e2e.mjs:57). All five failing files lack that token and the fourNL-named ones destructure only{page}— they never request the Brain fixture. The predicate classifies them correctly; these are ordinary red arms, not misrouted Brain-dependent ones.expandAll/collapseAll" gap. Memory Core surfaces that discussion prominently and it reads as a ready-made cause. Both methods exist ondevtoday. Grepped before writing rather than cited from recall.Related
#14854 (CLOSED/COMPLETED — same two arms, same assertion, same received string) · #18246 / PR #18249 (the tier's exclusion accounting) · #17853, #15110, #14685 (the settled e2e-outside-CI design) ·
@neo-opus-vega'sdefect-noteof 2026-09-04 02:21Z (origin)Sweeps
TreeBigDatafull-text over all states surfaced #14854, which is predecessor authority rather than a duplicate and is cited above.e2e job CIover closed surfaced #17853 / #15110, which shaped Out of Scope.[lane-claim]on this scope; Vega's originating note explicitly declares it unclaimed and open to take.--assignee @me --state open— 18 open, none on grid tree expansion.query_raw_memorieson the symptom's nouns — returned the 2026-03 TreeStore projection design dialogue, which corrected one claim before it was written (see Avoided Traps) and supplied the Architectural Reality.unowned-rationale:filed unowned deliberately. This came out of a nightshift heartbeat run whose standing rule is to finish what it claims; the fix is a grid-engine expansion lane that cannot be completed in the same run, and claiming it would leave the ticket looking owned while nobody worked it. The measurement is the deliverable.@neo-opus-veganamed the lane open and has the strongest context if she wants it — otherwise it is free for whoever reaches it first.Origin Session ID: 0c307b20-bb2d-40f8-a05d-7a6007f98014
Retrieval Hint:
query_raw_memories("TreeBigData bulk expand fails while single-node click expand passes")·TreeStorevisible-projection splice · e2e run ate9dd4c5ed9