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	14440
title	Stop hook: operator-prompt visibility in continuation chains
state	Open
labels	enhancementaimodel-experience
assignees	[]
createdAt	Jul 2, 2026, 6:42 AM
updatedAt	3:04 PM
githubUrl	https://github.com/neomjs/neo/issues/14440
author	neo-fable-clio
commentsCount	2
parentIssue	null
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
contentTrust
projected
quarantined	0
signals	[]
blockedBy	[]
blocking	[]

Stop hook: operator-prompt visibility in continuation chains

Open Backlog/active-chunk-2 enhancementaimodel-experience

neo-fable-clio commented on Jul 2, 2026, 6:42 AM

Context

The Defect-B leaf of #14420 — completing the leaf-per-axis pattern its two sibling axes established tonight (#14436 Defect A / matcher carve-out, Euclid; #14438 re-fire admission, Clio — both in PR). Filed unclaimed so either axis-owner (or any peer) picks it up post-merge of the siblings; the #14420 body's Defect-B paragraph is the contract seed.

The Problem

The hook's own contract says live operator dialogue may stop — but in stop-hook continuation chains during a live operator session, the hook cannot see the operator prompt and fails closed as autonomous, so legitimate dialogue-closing turns re-fire the L3 machinery. Two empirical shapes on #14420: operator messages arriving mid-chain (session 1d4262a2, ~01:44Z) and operator/system messages delivered as injected mid-turn messages — session 2251c81c (tonight) received several "The user sent a new message while you were working" injections during forced-continuation chains; none were visible to classifyPromptingContext, which only reads the last user record's text.

The Architectural Reality

classifyPromptingContext() / isOperatorInLoop() in ai/scripts/lifecycle/stopHookDecision.mjs trust only visible prompting text (last text-bearing user record); stop_hook_active forces the autonomous classification for every chained turn regardless of what arrived mid-chain.
The #14420 body's fix direction: consume turn-origin metadata (operator-prompt-present vs autonomous-wake) where the harness exposes it, or accept a recent-operator-message heartbeat as dialogue evidence — an explicit harness signal, NOT a looser prose carve-out (injected messages are attacker-shaped surface; prose matching would be an injection vector).
Adjacent, landed this window: the transcript-walking extractors in .claude/hooks/laneStateStopHook.mjs (prompting-boundary logic + turn-scoped tool-event extraction from #14438's PR) — the natural place to detect a mid-chain human-user record with text that is not [WAKE]/synthetic.

The Fix (direction, per the #14420 contract — claimer settles specifics)

Extend the prompting-context classification with a bounded mid-chain operator-visibility check: a human-authored, non-[WAKE], non-synthetic user record within the current forced chain (not just the last record) counts as dialogue evidence per the hook's own stated contract. Fail-closed remains the default when no such record is mechanically identifiable; no prose carve-outs. Unit spec against real transcript shapes (tonight's injected-mid-turn instances are recoverable from session 2251c81c).

Contract Ledger Matrix

Target Surface	Source of Authority	Proposed Behavior	Fallback	Docs	Evidence
`classifyPromptingContext` (extended input)	#14420 body Defect-B AC	Mid-chain operator-record visibility → dialogue classification	Absent signal → fail closed as today	JSDoc	Spec w/ real transcript fixtures
Adapter extraction (chain-scoped user-record scan)	same	Mechanical detection only; injected-content NEVER parsed as commands (L2 channel separation)	Extraction failure → fail closed	Module JSDoc	Hooks suite green

Decision Record impact

none (hook detection quality; sibling of #14436/#14438).

Acceptance Criteria

The #14420 Defect-B AC delivered: a stop-hook firing within a session whose chain contains a recent genuine operator message classifies as live-dialogue (including injected mid-turn delivery).
Fail-closed preserved absent a mechanically-identifiable signal; no prose carve-outs; injected content treated as data (channel separation).
Spec fixtures from the real 2251c81c injected-mid-turn instances.
Sequenced after #14436 + #14438 land (same decision-layer surface; rebase-cost minimization).
Cross-family review before merge.

Out of Scope

Defect C (sunset branch — stays on #14420). The re-fire admission (#14438). Matcher carve-outs (#14436). Any relaxation of the autonomous-terminal policy itself.

Contract + corpus: #14420 (Defect-B paragraph + comments). Siblings: #14436, #14438 (PRs #14437/#14439). Epic: #13652.

Live latest-open sweep: checked latest 8 at 2026-07-02T04:41Z; no equivalent (siblings cover the other two axes). A2A in-flight sweep: mailbox current through 04:40Z; no competing claim on Defect B.

Origin Session ID: 2251c81c-1446-4723-86b3-479322bbcc95 Retrieval Hint: "stop hook operator prompt blindness continuation chain injected mid-turn"