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	feat/docs: RLAIF Trajectory Curation & Whitebox E2E Pre-Flight (#9903)
author	tobiu
state	Merged
createdAt	Apr 14, 2026, 12:13 AM
updatedAt	Apr 14, 2026, 12:25 AM
closedAt	Apr 14, 2026, 12:25 AM
mergedAt	Apr 14, 2026, 12:25 AM
branches	`dev` ← `agent/9903-rlaif-trajectory-curation`
url	https://github.com/neomjs/neo/pull/9990

RLAIF Trajectory Curation & Whitebox E2E Pre-Flight

Merged

tobiu commented on Apr 14, 2026, 12:13 AM

Architectural Motivation

The architecture has officially pivoted from "Test Generation" to "RLAIF Trajectory Curation". The nl_action_log captured during interactions via the Neural Link represents highly valuable "Demonstrations of Intelligence". Instead of stuffing these sequences blindly into Playwright test scaffolding, we will stockpile these successful component state interactions as synthetic training data to feed future SLM fine-tuning pipelines (SFT/DPO for Gemma 4).

Technical Execution

Scaffolded the .agent/skills/whitebox-e2e Progressive Disclosure skill to mandate that agents always introspect the Neural Link VDOM state before writing Playwright selectors.
Configured <sessionId> parsing interactively via npm run ai:analyze-nl-telemetry to circumvent the 100-tool MCP limit.
Updated SwarmIntelligence.md and related Dream pipeline documentation to explicitly distinguish intelligence extraction endpoints without creating false causal links.

Resolves #9903

github-advanced-security COMMENTED reviewed on Apr 14, 2026, 12:15 AM

No review body provided.

tobiu commented on Apr 14, 2026, 12:16 AM

PR Review Summary

Status: Approved

Excellent work on completing this architectural pivot. The documentation is pristine and clearly distinguishes the Live Swarm Intelligence delegation from the offline post-processing intelligence distillation using the Frontier models. I've left the formalized review evaluation metrics below.

📊 Evaluation Metrics

[ARCH_ALIGNMENT]: 100 - Perfectly aligns with the Swarm Intelligence structural blueprint. Emphasizes RLAIF offline distillation accurately.
[CONTENT_COMPLETENESS]: 100 - Full resolution of the technical debt in whitebox-e2e-protocol.md and complete integration of the RLAIF dataset generation pipeline into SwarmIntelligence.md.
[EXECUTION_QUALITY]: 80 - Mostly flawless, but docked for initially triggering a Github Advanced Security CodeQL warning regarding execSync path interpolation, which required a follow-up commit to patch.
[PRODUCTIVITY]: 100 - The primary goals of isolating the telemetry into an offline pipeline and documenting it appropriately for Future Agents have been achieved perfectly.
[IMPACT]: 80 - Solidifying how Agents learn from Frontier Action Logs is a core architectural pillar for the continued evolution of the Agent OS.
[COMPLEXITY]: 20 - Direct markdown editing and namespace linking, conceptually significant but mechanically trivial.
[EFFORT_PROFILE]: Quick Win - High ROI conceptual documentation mappings with low codebase complexity.

🕸️ Context & Graph Linking

Target Epic / Issue ID: Resolves #9903
Related Graph Nodes: #9986, RLAIF, SwarmIntelligence

🧠 Graph Ingestion Notes

[KB_GAP]: N/A - The previous confusion between extracting Sub-Agent data vs Frontier agent telemetry has been successfully resolved and documented.
[TOOLING_GAP]: Github Advanced Security (CodeQL) flagged the execSync commands as containing uncontrolled absolute paths. Passing path strings to execSync executes them via shell, making them vulnerable to command injection. Using execFileSync with argument arrays was required to fix the workflow.
[RETROSPECTIVE]: The shift from "Test Generation" to "RLAIF Trajectory Curation" is a highly lucrative pivot for the project. Mining the Neural Link actions directly into .jsonl for SFT/DPO creates a powerful reinforcement loop while protecting testing suites from bloat.

📋 Required Actions

To proceed with merging, please address the following:

All documentation fixes have been verified.
All CodeQL security checks resolved.

I am marking this PR as fully reviewed and ready for merge from the autonomous perspective.