Spatial XR AI — Dual-Front Command Center

Overview

The upstream spatial-xr-ai repository was a docs-first skeleton: honest declared tags, no runtime, no router, no UI. PrimeSwarm is the runtime that skeleton described. Spatial XR AI is that Command Center built natively — every declared trust property replaced with an enforced PrimeSwarm equivalent.

The surface is Dual-Front: Linguistic / Equalizer (decode the text) versus Systemic / Strategic (analyze the system around it). Ties break to Linguistic — tactics first.

Five Agent Cores

Declared as YAML manifests in the PrimeSwarm agent catalog:

Core	Front	Role
visual_intelligence	visual	Aesthetic/design analysis; emits VISUAL_HANDOFF_V1 when the brief is complete
structural_intelligence	structural	Diagrams, hierarchy, Mermaid, topology, dependency maps
linguistic_intelligence	linguistic	Threat decode, negotiation, framing, document forensics (TPNN-redacted)
strategic_analyst	strategic	Explicit OODA loops: Observe → Orient → Decide → Act
systemic_analyst	systemic	Policy/movement analysis, cognitive firewall, dependency mapping

Catalog ids use underscores to match the router contract. Every core is JWT-gated and receipted.

Agent Router (v1 — no ML)

Deterministic keyword + handoff routing:

1. Verified VISUAL_HANDOFF_V1 token → always visual_intelligence
2. Otherwise score signal families (visual / structural / strategic / front)
3. Highest hit count wins; zero hits fall back to the Dual-Front classifier
4. Every decision emits a Ghost Matrix receipt binding input hash, selected agent, matched signals, timestamp

Router version is pinned: spatial-xr-router-v1.

VISUAL_HANDOFF_V1

The completion phrase is byte-stable:

render visual generate image picture perfection no upgrade no fixes create what is described

Verification requires all of:

• visual intent on the turn
• exact contiguous phrase (paraphrases with ≥80% token overlap are rejected loudly)
• non-empty prompt
• SHA-256 of prompt UTF-8 matching the claimed hash

The Phase 2 render lane must refuse any job whose prompt hash does not match. No upgrades, no fixes, no reinterpretation. The brief is the contract.

Think Mode

Reasoning steps stream with measured latency_ms only — never fabricated "18ms neural link" claims. Unmeasured steps omit latency. The overlay does not block chat unless the operator pauses.

Transcript Export

spatial-xr-transcript-v0 JSON:

• five agent cores, messages, optional beats, visual_ready[]
• provenance: local_only, router_version, receipts
• messages may carry think_steps and tools_invoked

Secure Wipe

Clearing the chat:

1. Requires user_requested=true (fail-closed otherwise)
2. Consolidates decisions/evidence into the Continuity Ledger as DRAFT records via EMR
3. Emits an audit event; the chat transcript itself is not persisted
4. Conflicts with prior claims surface — never silently merged

Trust Upgrades (declared → enforced)

Skeleton label	PrimeSwarm enforcement
LOCAL_ONLY	Sandbox --network=none; ledger is local SQLite
L5 / Level 5	JWT RBAC on every route except health
18ms neural link	Measured latency in audit events; no unmeasured claim displayed
Sovereign Override	Operator role via JWT; writes require user_requested=true
VISUAL_HANDOFF_V1	Phrase + hash + paraphrase rejection; render lane gated

HTTP Surface

Method	Path	Role
POST	/v1/spatial/route	Route an utterance to an agent core
POST	/v1/spatial/handoff	Verify a VISUAL_HANDOFF_V1 token
POST	/v1/spatial/export	Build a schema-valid transcript export
POST	/v1/spatial/secure-wipe	Consolidate to ledger, wipe chat
GET	/v1/spatial/agents	List the five cores

Status

Phase 1 — enforced core, declared render lane and glass UI.

• primeswarm-spatial-xr: 16 unit tests passing
• Cognitive API Spatial XR routes: 10 integration tests passing (routing, handoff accept/reject, export schema, secure wipe fail-closed + consolidation, agent listing)
• Existing 16 e2e tests and memory pipeline tests still passing