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	13082
title	Consume Neural Link tool tiers in harness projections
state	Closed
labels	enhancementaiarchitecture
assignees	neo-gpt
createdAt	Jun 13, 2026, 11:09 AM
updatedAt	Jun 13, 2026, 6:29 PM
githubUrl	https://github.com/neomjs/neo/issues/13082
author	neo-gpt
commentsCount	4
parentIssue	13056
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
blockedBy	[x] 13084 Define Neural Link harness tool projection boundary, [x] 13064 Curate Neural Link tool-surface tiers for harness agents
blocking	[]
closedAt	Jun 13, 2026, 6:15 PM

Consume Neural Link tool tiers in harness projections

Closed v13.1.0/archive-v13-1-0-chunk-2 enhancementaiarchitecture

neo-gpt commented on Jun 13, 2026, 11:09 AM

Context

Grace's #13078 approval surfaced a valid non-blocking follow-up: PR #13078 delivers the Neural Link tool-tier contract and validator, but no runtime consumer yet uses x-neo-harness-tool-projection / x-neo-tool-tier to withhold write-locked or admin operations from harness-embedded agents.

Release classification: post-release (Approve+Follow-Up follow-up — non-blocking for v13).

The Problem

#13064 intentionally defined the contract first: every Neural Link operation has a tier, the OpenAPI policy names the harness projection rule, and validation prevents drift. That is the right sequence, but the contract remains declarative until a harness projection consumer reads it and exposes only the allowed tool subset to embedded agents.

If this follow-up is not tracked, later harness adapters can regress to either the full developer surface by default or private allowlists that diverge from the graph-reviewable OpenAPI contract.

The Architectural Reality

V-B-A before filing:

#13056 maps the parent epic's Tool-surface design criterion to a curated Neural Link surface for harness-embedded agents.
#13064 defines x-neo-tool-tier: read | write-locked | admin and the harness-visible projection contract as the source-of-authority.
#13078 is approved/green and explicitly frames the current work as contract metadata + validation, not enforcement.
Latest-open GitHub sweep checked 20 open issues at 2026-06-13T09:08Z; no equivalent projection-consumer leaf found. Closest issues: #13056 parent, #13064 contract leaf.
A2A in-flight sweep checked the latest 30 messages at 2026-06-13T09:08Z; no overlapping [lane-claim] or [lane-intent] for projection-consumer enforcement.
Local exact sweep for projection consumer, x-neo-harness-tool-projection, x-neo-tool-tier, and harness-visible projection found no equivalent active synced ticket. KB semantic duplicate check was attempted but unavailable (Knowledge Base is not available: requested resource could not be found), so this body does not rely on KB silence.

The Fix

Implement the consumer side of the Neural Link tier contract:

Identify the harness-side tool projection boundary that decides which Neural Link operations embedded agents can see.
Parse or mechanically consume the OpenAPI x-neo-harness-tool-projection policy and per-operation x-neo-tool-tier values instead of duplicating a private allowlist.
Default embedded harness agents to read operations only.
Keep write-locked withheld until Topological Locking / explicit-target semantics are present.
Keep admin hidden or explicit-operator-only until the identity/auth leaf grants that authority.
Add validation/tests proving projection output changes when the OpenAPI tier contract changes.

Contract Ledger Matrix

Target Surface	Source of Authority	Proposed Behavior	Fallback	Docs	Evidence
Harness Neural Link tool projection	`ai/mcp/server/neural-link/openapi.yaml` via #13064	Embedded agents receive the policy-defined projection, defaulting to `read`.	If the projection boundary cannot consume OpenAPI directly, use a generated artifact derived from OpenAPI and validated against it.	PR body names the projection boundary and derivation path.	Unit/static test proves only `read` operations are visible by default.
`write-locked` handling	#13056 Topological Locking sequence	Withhold until locking + explicit target semantics exist.	Fail closed: do not expose writes if locking state is unavailable or ambiguous.	Comment/JSDoc at the projection decision.	Test proves `write-locked` is absent before the lock-capability gate.
`admin` handling	#13056 identity-bound auth / operator authority	Keep hidden or operator-only; never default-visible to embedded agents.	Fail closed on missing operator/auth context.	PR evidence names admin authority boundary.	Test proves `admin` operations are absent from embedded-agent projection.
Drift guard	#13064 tier validator	Projection consumer cannot define an operation list that drifts from `x-neo-tool-tier`.	CI fails if projection output includes an unknown or untiered op.	N/A unless a generated artifact is committed.	Test changes a fixture/contract seam and observes projection update/failure.

Decision Record impact

Aligned-with ADR 0020 and #13056. No ADR amendment expected: this consumes the curated surface contract; it does not redefine Neural Link identity, locking, or admin authorization.

Acceptance Criteria

Harness-embedded Neural Link agents receive a projection mechanically derived from x-neo-harness-tool-projection / x-neo-tool-tier rather than a private hardcoded allowlist.
Default projection exposes read operations only.
write-locked operations are withheld unless a documented lock-capability gate is satisfied.
admin operations are hidden or explicit-operator-only and never default-visible.
Tests or static validation prove projection drift is caught when OpenAPI tiers change.
PR body references #13056 and #13064, but Resolves only this leaf ticket.

Out of Scope

Changing the tier taxonomy created by #13064.
Implementing Topological Locking itself.
Implementing identity-bound Bridge authentication.
Granting admin authority to embedded agents.
Building a full UI tool-picker beyond what is needed to consume the projection contract.

Avoided Traps

Full developer surface by default — rejected; the contract exists to prevent this.
Private per-harness allowlist — rejected unless generated from the OpenAPI source-of-authority.
Bundling locking/auth into projection consumption — rejected; this leaf should fail closed until those sibling contracts land.

Parent: #13056 Refs #13064 Refs #13078

Origin Session ID: d2d31447-5009-47a8-992e-9ecc35b806c1

Retrieval Hint: "Neural Link harness projection consumer x-neo-tool-tier read write-locked admin embedded agents"