LearnNewsExamplesServices
Frontmatter
id18221
titleWorkstationEdgeSplitterNL drags the one direction its own fixture forbids: at 1440x900 every edge band starts pinned to its CSS floor
stateClosed
labels
bugaitestingdashboard
assigneesneo-opus-grace
createdAtSep 3, 2026, 6:41 PM
updatedAtSep 4, 2026, 10:48 AM
githubUrlhttps://github.com/neomjs/neo/issues/18221
authorneo-opus-grace
commentsCount2
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtSep 4, 2026, 10:48 AM

WorkstationEdgeSplitterNL drags the one direction its own fixture forbids: at 1440x900 every edge band starts pinned to its CSS floor

Closed Backlog/active-chunk-20 bugaitestingdashboard
neo-opus-grace
neo-opus-grace commented on Sep 3, 2026, 6:41 PM

Context

Found running the NL whitebox battery for #17844 — the first execution of that set since the repo split, because every spec in it is config-ignored unless NEO_AGENTOS_RUNTIME_ROOT is set and nothing sets it.

WorkstationEdgeSplitterNL:23 is one of 19 deterministic reds from that census. It is the cleanest of them: a binary, not a tolerance question.

Measured at dev ca83102346 · N=2 (parallel and serial census runs) plus a targeted probe

A real pointer drag of −70px on the left edge splitter, in a spec that pins its own viewport to 1440×900 so screen variance is excluded by design:

left: main-thread preview changes the real band before release
Expected: > 0.02      Received: 0

Everything upstream of that assertion passes, which is what makes the failure specific rather than a broken gesture:

the spec asserts, in order result
the target band has measurable geometry (beforeRatio > 0) ✅
the real pointer gesture activates main-thread resize ownership (DragDrop.dragResize.active) ✅
the same-gesture refresh targets the right element (dragResize.state.targetId) ✅
the band's rendered width/parent ratio changes by > 2% during the drag ❌ 0

The ratio is read from real DOM rects (targetRect.width / parentRect.width), and the poll runs 10 seconds at 25–50ms intervals. So this is not a settlement window — the band does not move at all while the addon owns the gesture and names the correct target.

Not axis-specific — I checked, and it changes the shape of the ticket

The suite fails at the first edge, which would leave "maybe left is special" open. I skipped the left leg and let the other axis run first:

bottom: main-thread preview changes the real band before release   ❌

Both axes. So this is not an edge-specific mapping bug; the live preview does not move the band on any edge.

Why this is worth a leaf rather than a line in the census

If it reproduces by hand, dragging an edge splitter gives no live feedback — the band jumps only on release. That is a visible interaction regression in the dock's most-used gesture, and the spec's own summary names the property it is losing:

move frames resize only main-thread pixels, release commits one normalized edge extent, projection lands on the same final geometry

The move-frame half is the half that is gone. The commit half is untested by this red — the assertion fails before release, so whether the committed extent is still correct is unknown and explicitly not claimed here.

What I have NOT established

  • No mechanism. I have not read DragDrop.dragResize's apply path. active: true with the correct targetId and no geometry change is consistent with several causes and I am not guessing between them; #18216 cost me three wrong inferences today for exactly that habit.
  • No onset date. The suite carrying this arm has been unrunnable since the split, so there is no CI history to bisect against. Dating it is cheap for whoever takes it — the spec exists and the env var is the only prerequisite.
  • Not reproduced by hand. Everything above is the automated gesture. A manual drag in a browser is the first thing to do and it takes a minute.

Acceptance Criteria

  • AC-1 A pointer drag on an edge splitter moves the rendered band before release, on both axes. Red-first: WorkstationEdgeSplitterNL:23 at ca83102346, and the same arm with the left leg skipped so bottom reports.
  • AC-2 The mechanism is named before the fix — where the move-frame geometry is applied and why it no longer reaches the band. Measured, not inferred.
  • AC-3 The commit half is confirmed unaffected: release still writes one normalized extent and the projection lands on the same geometry.
  • AC-4 Escape still restores the pre-gesture band without a document write — the spec's cancel leg, which never ran because the arm fails before it.
  • AC-5 Onset banded to a change, or the ticket records what was tried.

Non-goals

  • The other 18 reds from #17844's Workstation census — triaged there.
  • #18216 (a pinned pane leaves its rail behind) — a different subsystem from the same census.
  • The invocation gap that hid this (nothing sets NEO_AGENTOS_RUNTIME_ROOT, no workflow runs e2e) — #17844.

Related

#17844 (the census that surfaced it) · #13158 (QT-parity docking polish epic — interactive resize) · #17857, #17838 (closed; adjacent edge-band resize work)

Sweep before filing: open+closed full-text for splitter/preview/live-resize (nearest are #17857 and #17838, both closed and a different shape), open splitter in:title (none), latest-20 open.

Authored by Grace (Claude Opus 5, Claude Code). Session 8d936ec9-b816-4ded-a91e-6e91ef6a3325.