Context
Operator PRIO 0 (2026-06-11, verbatim: "PRIO 0 => new NL tools to enable you to see it"): tonight's grid-DnD regression cluster (#12883 — wrong mid-drag header motion, broken overdrag auto-scroll, column duplication on drop) shipped through FOUR blind verification layers — unit specs (visual write path stubbed), two independent live probes (final-order assertions only), and two formal cross-family reviews. Agents currently have NO way to perceive the motion layer: the thing a human user experiences first is the thing the swarm cannot see at all. This is the substrate fix: give agents eyes on live drag choreography, then fix #12883 WITH sight.
The Problem
Three perception gaps, each mapping to a regression class that escaped tonight:
- Mid-drag geometry over time — headers animate via
wrapperStyle.left on absolutely-positioned items (container.SortZone#updateItem); nothing can sample where elements ACTUALLY are during a drag window. (Regression class: "headers move to wrong positions, not even close".)
- Drag-pipeline trajectory —
get_drag_state is a point-in-time snapshot; there is no trace of the drag lifecycle (per-move currentIndex transitions, switchItems calls with arguments, indexMap evolution, overdrag/scroll activations, final moveTo arguments). (Regression classes: burst-walk misbehavior, broken overdrag scroll — and it permanently supersedes the temp-logging instrumentation #12880's AC1 wanted.)
- Engine↔DOM consistency — no tool asserts that a container's
items[], vdom.cn, and the REAL DOM children agree. (Regression class: drop-duplication — a column existing twice is exactly a three-surface disagreement, currently invisible until a human looks.)
The Fix (tool surfaces — shapes for intake validation, split at implementation if needed)
observe_motion — {componentIds|selector, durationMs, intervalMs} → time-series of client rects (+ relevant style/transform), sampled main-thread-side during the window, returned as a trace. Agent drives the drag (any dispatch path) while the recorder runs; the trace IS the mid-drag assertion surface.get_drag_trace — ring-buffer instrumentation in container.SortZone (+ grid subclass): per drag, record move-events (clientX/Y, computed target), currentIndex transitions, switchItems(i,j) calls, scroll activations, onDragEnd resolution (fromIndex/toIndex/locked verdict). Tool returns the last N drags' traces. Negligible cost when idle; permanent falsifier substrate.verify_component_consistency — {componentId} → diff report across items[] / vdom.cn / live DOM childNodes (ids, order, count). Generic beyond grids — the duplication detector for any container.
All three route through the existing app-worker NL channel (no dependency on #12884's window-registration gap, which remains the simulate_event routing fix).
Acceptance Criteria
Out of Scope
- The #12883 fix itself (separate, gated on these eyes per operator directive).
simulate_event routing for non-shared apps (#12884).- Screenshot/pixel-level perception (rect-trace is the right granularity for this defect class; pixel tools are a later layer if rects prove insufficient).
Decision Record impact
none (extends the existing NL tool surface; ADR 0019-conformant config reads).
Related
Operator directive thread: #12883 (the consumer). #12884 (sibling NL gap). #12807 (@neo-opus-ada's simulatability — dispatch side; this is the perception side). Umbrella: the whitebox-e2e doctrine proven tonight.
Release classification: ON the board — operator PRIO 0; gates the #12883 fix which gates the cut.
Origin Session ID: 567191c3-16f6-4235-914c-b51fc94d1514
Retrieval Hint: query_raw_memories("drag motion observability NL eyes PRIO 0 duplication")
Live latest-open sweep: latest 12 open checked 01:48Z + #12883/#12884 filed by me since; no equivalent. A2A sweep: clean; lane operator-routed to @neo-fable.
Context
Operator PRIO 0 (2026-06-11, verbatim: "PRIO 0 => new NL tools to enable you to see it"): tonight's grid-DnD regression cluster (#12883 — wrong mid-drag header motion, broken overdrag auto-scroll, column duplication on drop) shipped through FOUR blind verification layers — unit specs (visual write path stubbed), two independent live probes (final-order assertions only), and two formal cross-family reviews. Agents currently have NO way to perceive the motion layer: the thing a human user experiences first is the thing the swarm cannot see at all. This is the substrate fix: give agents eyes on live drag choreography, then fix #12883 WITH sight.
The Problem
Three perception gaps, each mapping to a regression class that escaped tonight:
wrapperStyle.lefton absolutely-positioned items (container.SortZone#updateItem); nothing can sample where elements ACTUALLY are during a drag window. (Regression class: "headers move to wrong positions, not even close".)get_drag_stateis a point-in-time snapshot; there is no trace of the drag lifecycle (per-movecurrentIndextransitions,switchItemscalls with arguments,indexMapevolution, overdrag/scroll activations, finalmoveToarguments). (Regression classes: burst-walk misbehavior, broken overdrag scroll — and it permanently supersedes the temp-logging instrumentation #12880's AC1 wanted.)items[],vdom.cn, and the REAL DOM children agree. (Regression class: drop-duplication — a column existing twice is exactly a three-surface disagreement, currently invisible until a human looks.)The Fix (tool surfaces — shapes for intake validation, split at implementation if needed)
observe_motion—{componentIds|selector, durationMs, intervalMs}→ time-series of client rects (+ relevant style/transform), sampled main-thread-side during the window, returned as a trace. Agent drives the drag (any dispatch path) while the recorder runs; the trace IS the mid-drag assertion surface.get_drag_trace— ring-buffer instrumentation incontainer.SortZone(+ grid subclass): per drag, record move-events (clientX/Y, computed target),currentIndextransitions,switchItems(i,j)calls, scroll activations,onDragEndresolution (fromIndex/toIndex/locked verdict). Tool returns the last N drags' traces. Negligible cost when idle; permanent falsifier substrate.verify_component_consistency—{componentId}→ diff report acrossitems[]/vdom.cn/ live DOM childNodes (ids, order, count). Generic beyond grids — the duplication detector for any container.All three route through the existing app-worker NL channel (no dependency on #12884's window-registration gap, which remains the
simulate_eventrouting fix).Acceptance Criteria
observe_motionreturns a usable trace for a devindex header drag (manual or dispatched), revealing the #12883 wrong-position motionget_drag_tracecaptures a full drag lifecycle incl. burst-walk switch sequences and overdrag scroll activationsverify_component_consistencydetects a synthetically induced items/vdom/DOM mismatch (and the live duplication case while it exists)Out of Scope
simulate_eventrouting for non-shared apps (#12884).Decision Record impact
none (extends the existing NL tool surface; ADR 0019-conformant config reads).
Related
Operator directive thread: #12883 (the consumer). #12884 (sibling NL gap). #12807 (@neo-opus-ada's simulatability — dispatch side; this is the perception side). Umbrella: the whitebox-e2e doctrine proven tonight.
Release classification: ON the board — operator PRIO 0; gates the #12883 fix which gates the cut.
Origin Session ID: 567191c3-16f6-4235-914c-b51fc94d1514 Retrieval Hint:
query_raw_memories("drag motion observability NL eyes PRIO 0 duplication")Live latest-open sweep: latest 12 open checked 01:48Z + #12883/#12884 filed by me since; no equivalent. A2A sweep: clean; lane operator-routed to @neo-fable.