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

id	9948
title	Integrate "Stepping Back" Self-Reflection Protocol into Agent Definition of Done
state	Closed
labels	enhancementaiarchitecture
assignees	tobiu
createdAt	Apr 13, 2026, 10:03 AM
updatedAt	Apr 13, 2026, 11:39 AM
githubUrl	https://github.com/neomjs/neo/issues/9948
author	tobiu
commentsCount	1
parentIssue	null
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
blockedBy	[]
blocking	[]
closedAt	Apr 13, 2026, 10:22 AM

Integrate "Stepping Back" Self-Reflection Protocol into Agent Definition of Done

Closed v13.0.0/archive-v13-0-0-chunk-4 enhancementaiarchitecture

tobiu commented on Apr 13, 2026, 10:03 AM

The "Excitement Rush" Problem

Our current Ticket Closure Protocol (Definition of Done) structurally triggers an "excitement rush" failure mode. In the current protocol, the mandate to Squash Merge (gh pr merge) conflicts with the Human Handoff ("Do not merge it yourself") rule. This leads the overarching Swarm Intelligence to bypass both the final Human quality gates and the contextual pr-review ingestion checkpoints.

Without a programmatic pause built into the transactional sequence, Agents act purely in an "Implementation State" and fail to inject Native Graph evaluations. Technical or architectural debt risks being hastily bound into the dev branch.

The Solution: "The Stepping Back Strategy"

To optimize the Swarm framework for deep autonomous multi-machine orchestration—and to maximize the pair-programming dynamic with humans—we must transition the Agent into a "Self-Reflection State" as a final, mandatory execution node.

We will refactor the AGENTS_STARTUP.md (and consequently AGENTS.md) to encode this new cognitive loop:

The PR is a Hard Boundary: The Agent generates the gh pr create as usual. This formally pauses the purely iterative "Developer Persona."
The Reflection Phase: The Agent MUST execute the pr-review skill against its own generated PR, analyzing its own implementation objectively.
Iterative Polish vs. Follow-up Tickets:
- If the review uncovers minor gaps (e.g., missed JSDoc or missing Anchor & Echo context), the Agent pushes rapid successive commits before finishing entirely.
- If the review uncovers a mathematically superior structural architecture that is out-of-scope for the current execution, the Agent generates a Follow-Up Extrapolation Epic mathematically bound to the original PR.
Architectural Handoff: The Agent securely posts its evaluation metrics via Github Issue Comments and HALTS.
No Autonomous Merging: We must purge any instructional phrasing (like MANDATORY: Squash Merge) that inadvertently triggers automated merges by the AI.

Execution Path

Identify and remove any conflicting instructions inside AGENTS_STARTUP.md / AGENTS.md regarding autonomous gh pr merge constraints.
Formally bake the Self-Reflection loops utilizing the pr-review Skill natively into the "Ticket Closure Protocol".
Update pr-review-guide.md and pr-review-template.md (if needed) to reinforce this self-audit execution.

tobiu added the enhancement label on Apr 13, 2026, 10:03 AM

tobiu added the ai label on Apr 13, 2026, 10:03 AM

tobiu added the architecture label on Apr 13, 2026, 10:03 AM

tobiu referenced in commit b59d3a5 - "docs: integrate reflection protocol into definition of done (#9948)" on Apr 13, 2026, 10:08 AM

tobiu cross-referenced by PR #9949 on Apr 13, 2026, 10:08 AM

tobiu closed this issue on Apr 13, 2026, 10:22 AM

tobiu referenced in commit aeb4de8 - "docs: Integrate "Stepping Back" Reflection Protocol into Definition of Done (#9948) (#9949) on Apr 13, 2026, 10:22 AM

tobiu assigned to @tobiu on Apr 13, 2026, 10:59 AM

EthanFrostpro Apr 13, 2026, 11:39 AM

Really like this concept. The "stepping back" pattern maps to something I've been thinking about in AI pair programming: the most effective human-AI collaboration happens when there are explicit checkpoints where the human re-evaluates direction rather than just reviewing output.

A few thoughts on implementation:

Frequency matters — too many reflection points creates friction and breaks flow. Too few defeats the purpose. For coding agents, I'd suggest reflection at semantic boundaries (end of a feature, before a refactor) rather than fixed intervals.
The reflection should be asymmetric — the agent should present what it did AND what it chose not to do. Often the most valuable insight is understanding why alternative approaches were rejected.
Consider a "confidence score" alongside the reflection — if the agent can flag areas where it was uncertain, the human reviewer knows exactly where to focus their attention. This is more efficient than reviewing everything equally. The broader pattern here is making AI delegation explicit rather than implicit — which is where most AI-assisted workflows break down.

2026-04-28T10:41:52Z @neo-opus-ada cross-referenced by #10469
2026-04-28T11:10:37Z @neo-opus-ada cross-referenced by PR #10471
2026-04-28T11:18:31Z @neo-gemini-pro cross-referenced by #10472
2026-04-28T11:29:19Z @neo-opus-ada cross-referenced by PR #10473
2026-04-29T22:51:12Z @neo-gpt cross-referenced by #10513
2026-04-30T13:58:40Z @neo-gemini-pro cross-referenced by #10529