Create first grid through Neural Link in a live app

Context

Release classification: post-release Agent Harness product line; attach to Agent Harness Project 13, not the v13 release board.

Parent: #13349 Grandparent: #13012 Supersedes the closed negative-ROI leaf #13353 / PR #13354.

Operator correction on 2026-06-15 rejected #13353/#13354 at the premise level: the H2 proof is not "render a static constrained blueprint that resembles a grid." The intended proof is that an agent uses Neural Link inside a live Neo app to create a new grid, with whitebox E2E evidence. The static-blueprint path made the grid contract brittle and tried to whitelist a surface that should remain normal Neo grid configuration.

ADR 0020 frames conversational app creation as chat -> live blueprint emission into peer apps, with the canonical "build me a Neo grid" result becoming a live, mutable pane/window. learn/agentos/NeuralLink.md identifies Neural Link as the runtime bridge for inspecting and mutating live component trees. learn/guides/testing/WhiteboxE2E.md and .agents/skills/whitebox-e2e/references/whitebox-e2e-protocol.md establish the evidence shape: Playwright drives the live app, Neural Link validates App Worker truth.

The Problem

The first H2 implementation attempt proved the wrong thing. It created a deterministic app-local blueprint and rendered it directly in the widget child app. That can be made green, reviewed, and visually plausible, but it does not demonstrate the product primitive: an agent creating a Neo grid inside a live application through the same runtime possession surface the harness is built around.

Keeping the static-blueprint chain as the first proof would teach future leaves the wrong seam. It would encourage a brittle allow-list of grid config fields, DOM-shaped visual proof, and isolated app code instead of NL-mediated live object creation and worker-state verification.

The Architectural Reality

• Neural Link already exposes a constrained create_component tool for adding components to a live container by ntype/className; this is the right live-runtime proof seam.
• Existing whitebox E2E examples demonstrate NL live creation and grid introspection separately: test/playwright/e2e/NeuralLinkCreateComponent.spec.mjs, test/playwright/e2e/NeuralLinkCreateComponentMultiWindow.spec.mjs, test/playwright/e2e/GridBigDataNL.spec.mjs, and test/playwright/component/list/Chip.spec.mjs for component-config setup patterns.
• The proof should create a real Neo grid using ordinary grid configuration, not invent a mirror schema that attempts to whitelist all possible grid configs.
• This remains single-writer. Topological Locking / multi-writer Scenario-C only enters if this leaf adds another writer or concurrent NL mutation path.
• Electron shell/build-root remains #13033 and is not part of this ticket.

The Fix

Implement the corrected first H2 proof as a whitebox E2E / live-app slice:

1. Boot a live Neo app with useAiClient through the E2E fixture.
2. Resolve a target live container through Neural Link, not by hardcoding a static render path.
3. Use Neural Link to create a grid-container (or current canonical grid component) in that container with normal Neo grid configuration: columns, model/store fields, and sample rows.
4. Verify via Neural Link that the created grid is present in the component tree and backed by live store/state, not just visible DOM.
5. Add a minimal DOM/visual assertion only as secondary evidence that the live component rendered.

Contract Ledger Matrix

Decision Record impact

Aligned-with ADR 0020. No ADR amendment expected; this corrects the first H2 leaf proof to match the Neural Link / co-habitation premise.

Acceptance Criteria

• A whitebox E2E test boots a live Neo app with Neural Link enabled and connects via the neuralLink fixture.
• The test creates a new grid component inside a live container through Neural Link (create_component or the fixture equivalent), not by static app render or direct DOM/HTML generation.
• The grid uses normal Neo grid configuration with columns, model/store fields, and sample row data; the implementation does not attempt to whitelist all grid config possibilities.
• Neural Link verification proves the created grid exists as a live component and exposes expected worker-side state/store data.
• A secondary rendered-DOM assertion proves the live component appears in the page.
• The PR explicitly documents why #13353/#13354 were superseded and does not revive the static constrained-blueprint path.
• The PR uses the explicit whitebox E2E command/config; do not run default npx playwright test.
• The implementation does not modify Electron shell/build-root scope and does not claim #13033.

Out of Scope

• Full natural-language chat parsing or prompt streaming.
• Persistence/write-through to Git, Memory Core, or project files.
• Multi-writer locking or Extended-NL coordination beyond the single writer used by the test.
• Electron packaging, BrowserWindow management, or Agent OS process topology; that remains #13033.
• A broad grid-builder abstraction or mirror schema for every possible grid config.
• Public performance/endurance claims.
• Partner-private or client-specific UI.

Avoided Traps / Rejected Shapes

• Static first-widget blueprint: rejected by #13354 operator correction; it proves the wrong seam.
• Whitelisting grid configs: rejected. The grid surface is too broad; use ordinary Neo config through a bounded NL action.
• Black-box DOM-only test: rejected. H2 needs worker-state proof via Neural Link.
• Arbitrary code execution: rejected. This leaf uses the existing constrained NL component-creation surface, not generated JavaScript evaluation.
• Duplicate Electron work: rejected. Shell/build-root remains #13033.

Related

• Parent: #13349
• Superseded leaf: #13353
• Rejected PR: #13354
• Grandparent: #13012
• Electron sibling lane: #13033
• Neural Link undo/write-guard siblings: #13221, #13286, #13306

Duplicate / Claim Sweep

Live latest-open sweep: checked the latest 20 open GitHub issues at 2026-06-15T16:03:35Z. Relevant live hits were #13349 (parent), #13355/#13357/#13361 (downstream H2 leaves from the now-rejected static-blueprint path), #13330 (unrelated WriteGuard cleanup), #13306/#13286/#13221 (NL undo/write-guard lanes), #13056 (Extended-NL coordination), and #13033 (Electron shell sibling); no corrected NL-created-grid H2 proof ticket was present.

Targeted GitHub search immediately before filing for "Neural Link" "grid" "whitebox" surfaced #9820, an older grid-mutability R&D issue, but no current Agent Harness H2 leaf for creating the first grid through NL inside a live app.

A2A in-flight sweep: native list_messages / add_message tools are not exposed in this Codex Desktop surface (tool_search returned no A2A tools), so a live mailbox claim sweep could not be executed. This limitation is recorded explicitly rather than silently omitted.

Semantic KB sweep: ask_knowledge_base(query='Neural Link whitebox e2e create grid live app Agent Harness first widget', type='all') pointed to ADR 0020, learn/agentos/NeuralLink.md, learn/guides/testing/WhiteboxE2E.md, and existing grid/NL examples; it did not identify an equivalent active H2 ticket.

Exact repo/content sweep: rg over active/archived issues, discussions, AgentOS docs/apps, and Playwright e2e/component tests found existing NL create-component and grid-inspection examples plus #13012/ADR 0020 references, but no current child issue for the corrected NL-created-grid proof.

Origin Session ID: unavailable in this Codex Desktop surface.

Retrieval Hint: "Agent Harness H2 first grid Neural Link create_component whitebox e2e live app supersedes #13353 #13354"