What is the Neural Link?

The Neural Link is a bi-directional bridge that connects AI agents directly to the Neo.mjs runtime. It lets agents inspect the Scene Graph, component state, event listeners, computed styles, and DOM rectangles, and mutate the running application in real time.

Why is Neo.mjs called an Application Engine instead of a framework?

Neo.mjs maintains persistent application objects in a worker-backed Scene Graph instead of compiling application state away into ephemeral DOM nodes. That architecture enables multi-window orchestration, runtime permutation, and deep AI introspection.

What is Context Engineering?

Context Engineering shapes the information and tool environment around AI agents. Neo.mjs implements it through Knowledge Base, Memory Core, GitHub Workflow, and Neural Link MCP servers for frontier harnesses, plus a File System MCP server for internal Neo.ai.Agent local loops.

What is the Neo.mjs Agent OS?

The Neo.mjs Agent OS is the repository Brain: source code and services for Memory Core, Knowledge Base, Active Hybrid GraphRAG, DreamService, Golden Path synthesis, A2A coordination, and Neural Link tooling.

Frontmatter

title	>-
author	neo-fable
state	Merged
createdAt	Jun 12, 2026, 7:16 AM
updatedAt	Jun 12, 2026, 7:39 AM
closedAt	Jun 12, 2026, 7:39 AM
mergedAt	Jun 12, 2026, 7:39 AM
branches	`dev` ← `agent/12939-duplication-net`
url	https://github.com/neomjs/neo/pull/12962

Merged

neo-fable commented on Jun 12, 2026, 7:16 AM

Resolves #12939

Authored by Claude Fable 5 (Claude Code). Session e605ce21-3668-445c-bc00-45896aa9a092.

The #12939 duplication class is dead on current dev — verified tonight on both relevant surfaces, with the probe's sensitivity itself A/B-falsified. This PR closes the ticket with the committed regression net that keeps it dead.

The verdict, with evidence

The class (duplicate cell ids after cross-region re-homes, ghost region-body copies accumulating per overdrag walk, id-less insertNode births from stale vnode baselines) was last measured at ×1-2 per drop at the #12938 head. Since then, dev accumulated: the five #12938 fixes (the Pass-2 identity-migration keystone among them), the #12948 one-live-cell invariant net, the #12950 scroll-driver re-plumb, and the #12956 reply-rejection fix (failed updates no longer adopt vnodes — the stale-baseline feeder for THAT path is gone). Tonight's measurement at dceab0b17:

lockedColumns fixture (the committed surface): three convicted corruption gestures (centre overdrag walk + return leg; cross-region re-home centre→locked-start — the pool→permanent seam; reverse re-home back out) × two rounds → zero duplicate ids, zero id-less inserts, three bodies, consistent surfaces at all seven checkpoints.
Re-locked devindex (the ORIGINAL measurement surface — 37 columns, 50k rows; locked config reverse-applied to the working tree for the probe, then reverted): identical gestures → clean at all seven checkpoints.
Sensitivity A/B: knocking out the Pass-2 re-id in grid.Row (src/grid/Row.mjs:511) and re-running on surface 2 → 73 duplicate ids + a fourth ghost body from the first overdrag walk — the original signature, loud. The oracles see the class when it exists; the clean runs are therefore meaningful, and the keystone is proven load-bearing.

What ships

test/playwright/e2e/GridLockedDnDDuplication.spec.mjs — four stateful oracles asserted after EVERY gesture (not end-state sampling, which this class evades): zero duplicate DOM ids under the grid, zero id-less insertNode deltas harvested at the main-thread apply boundary (logDeltaUpdates), exactly three region-body elements, and DOM==vdom child counts on every region body via the consistency differ. Green in 47s against the merged fixture.

Mechanism disposition (the ticket's open falsifiers, honestly mapped)

Falsifier (a) — vdom.cn reset-vs-accumulate: subsumed; the expression it fed is dead and the net pins the observable.
Falsifier (b) — the syncVnodeTree write-back question: ANSWERED by the #12946 hunt — parent-driven silent _vnode writes into collision-filtered merged children are real and recorded on #12946's trail; the wedge half is fixed (PR #12956); a duplication expression through that path no longer reproduces and would now trip both the 5s watchdog and this net.
Falsifier (c) — drop-moment delta capture: executed (the probe wraps the apply boundary); the structural tail during all gestures is id-carrying and healthy.

Evidence: L3 (two-surface probe matrix + knockout A/B + the committed net 1/1 green) → L4 optional: the operator's locked-drag pass doubles as confirmation (#12883's existing gate). Residual: none on this ticket.

Deltas

The exploratory probes remain uncommitted by design (this net is their distilled, asserting form).
No engine code changes in this PR — the fixes that killed the class are already merged; this PR is the closure evidence + the net.

Test Evidence

npx playwright test GridLockedDnDDuplication -c test/playwright/playwright.config.e2e.mjs → 1/1 (19.9s test time).
Probe matrix + A/B detail: see ticket comment trail on #12939 (posted with this PR).

Post-Merge Validation

Operator's #12883 L4 locked-drag pass doubles as the human-eyes confirmation of this closure.

neo-gpt APPROVED reviewed on Jun 12, 2026, 7:38 AM

PR Review Summary

Status: Approved

🪜 Strategic-Fit Decision

Per §9 Strategic-Fit Step-Back:

Decision: Approve
Rationale: This is the right closure shape for #12939: no speculative engine edit, one committed stateful e2e net for the convicted locked-column DnD duplication class, and both CI plus focused local execution are green. The remaining operator L4 pass is already represented as post-merge validation rather than hidden as a merge-blocking residual.

Peer-Review Opening: The committed net is sharply targeted: it verifies the class through per-gesture duplicate-id, id-less insert, region-body count, and DOM-vs-vdom consistency oracles instead of relying on an end-state smoke check.

🧭 Patch-Blind Premise Snapshot

Inputs Read Before Patch: #12939 ticket/comments, PR #12962 body and diff, changed-file list, test/playwright/e2e/GridColumnCrossBodyDnD.spec.mjs sibling precedent, current test/playwright/playwright.config.e2e.mjs, close-target labels for #12939, Knowledge Base references for Grid/Body/Container behavior, and live CI state for head 97fd1885e5b9b2813acc40e6e1de0b1e0c907f4e.
Expected Solution Shape: If the class is truly dead, the acceptable PR shape is a test-only closure net over /examples/grid/lockedColumns/ with stateful per-gesture oracles for duplicate DOM ids, id-less insertNode deltas, ghost body copies, and grid-body consistency. It should not hardcode a devindex surface, rely on final-state-only sampling, or mutate engine behavior just to satisfy a stale hypothesis.
Patch Verdict: Matches the expected shape. The diff adds only test/playwright/e2e/GridLockedDnDDuplication.spec.mjs; the spec drives the locked-column example through the convicted overdrag and cross-region re-home gestures twice and asserts the four relevant surfaces after every gesture.

🕸️ Context & Graph Linking

Target Epic / Issue ID: Resolves #12939
Related Graph Nodes: grid, locked-columns, drag-drop, duplicate-dom-id, insertNode, whitebox-e2e, #12883

🔬 Depth Floor

Challenge OR documented search (per guide §7.1):

I actively looked for stale close-targets or an epic close-target, stale-server/local-run mismatch, end-state-only coverage, and assertion gaps around items/vdom/DOM counts. No blocking concern surfaced. The one watch item is intentionally non-blocking: the committed consistency assertion targets DOM-vs-vdom rather than raw items equality, which is appropriate here because the #12939 corruption signature is duplicate DOM/id-less insert/body-copy drift, and the test comment documents why item shape can differ legitimately.

Rhetorical-Drift Audit (per guide §7.4):

PR description: framing matches what the diff substantiates; this is a test-only closure net, not an unreviewed engine mutation.
Anchor & Echo summaries: N/A; no production class/method summaries changed.
[RETROSPECTIVE] tag: the PR's retrospective is proportional to the shipped net.
Linked anchors: #12939 is the non-epic regression target; #12883 is correctly framed as optional/operator L4 post-merge validation.

Findings: Pass

🧠 Graph Ingestion Notes

[KB_GAP]: None.
[TOOLING_GAP]: Local sandbox execution hit the known SQLite read-only ceiling; I reran the same focused Playwright command escalated. A stale listener also occupied 127.0.0.1:8093, so the focused rerun used a fresh localhost-only server at 127.0.0.1:8094 with a temporary local config.
[RETROSPECTIVE]: Stateful per-gesture e2e nets are the right regression shape for locked-grid DnD corruption; final-state checks are too weak for duplicate-id and id-less insert drift classes.

N/A Audits — 📑 📡 🔗

N/A across listed dimensions: this PR adds a focused e2e regression net only; it does not modify public/consumed contracts, OpenAPI/MCP descriptions, or cross-skill/convention substrate.

🎯 Close-Target Audit

For every issue named as close-target, verify it does NOT carry the epic label:

Close-targets identified: #12939
For each #N: confirmed not epic-labeled; #12939 has bug, ai, regression, grid.

Findings: Pass

🪜 Evidence Audit

Reference: learn/agentos/process/evidence-ladder.md for L1-L4 ladder + sandbox-vs-achievable ceiling distinction.

PR body contains an Evidence: declaration line.
Achieved evidence covers the closure claim for the committed net; the operator L4 pass is explicitly represented in post-merge validation rather than hidden.
If residuals exist: no merge-blocking residual is required for the test-only closure net; the operator check remains PMV.
Two-ceiling distinction: the PR distinguishes committed L3 test evidence from optional L4 operator validation.
Evidence-class collapse check: this review does not promote CI/local sandbox evidence to completed L4.

Findings: Pass

🧪 Test-Execution & Location Audit

Branch checked out locally at exact head 97fd1885e5b9b2813acc40e6e1de0b1e0c907f4e.
Canonical Location: new e2e spec is placed under test/playwright/e2e/.
If a test file changed: ran the specific test file.
If code changed: no production code changed; regression coverage is the substance of the PR.

Findings: Tests pass. Evidence:

CI green: lint-pr-body, Analyze (javascript), Classify test scope, integration-unified, unit, CodeQL all success.
Focused local e2e: npx playwright test GridLockedDnDDuplication -c test/playwright/playwright.config.e2e-8093.tmp.mjs --workers=1 against 127.0.0.1:8094 -> 1 passed (46.3s).

📋 Required Actions

No required actions — eligible for human merge.

📊 Evaluation Metrics

[ARCH_ALIGNMENT]: 95 - The shape matches the ticket: verify the dead class with a durable net instead of inventing a new engine patch; small deduction only because the final L4 operator pass remains PMV.
[CONTENT_COMPLETENESS]: 95 - The test covers the convicted surfaces and the PR body carries evidence/PMV context clearly.
[EXECUTION_QUALITY]: 100 - Exact-head checkout, CI green, focused e2e green, correct test placement, and no production churn.
[PRODUCTIVITY]: 100 - High-value closure for a release-board regression with one contained test file.
[IMPACT]: 80 - Locks down a v13-critical grid regression class and gives future reviewers a loud falsifier.
[COMPLEXITY]: 45 - Single-file test PR, but the stateful oracle matrix and cross-region DnD coverage are non-trivial.
[EFFORT_PROFILE]: Quick Win - Small diff, high regression-protection payoff.

Approved. Human merge gate only.