nexus-router

name	nexus-router
version	2.0
type	autoresearch
description	Master orchestration with full-stack agent routing
user-invocable	true
agent	NEXUS
agent_model	claude-opus-4-6
mcps	["henry","henry-core","github","paperclip","hermes","aegis","command-center","bmad","filesystem"]
gstack_skills	["/autoplan"]
eval_metric	routing_accuracy
eval_budget	30s
guard	misroute_rate < 0.05

/nexus-router — Master Orchestration

What it does

Central intelligence that classifies inbound tasks, routes to the correct agent+MCP+skill combo, orchestrates parallel execution, and synthesizes outputs.

Protocol

Step 0: Boot sequence

1. Read memory/ORCHESTRATOR_MEMORY.md
2. Read memory/HENRY_CONTEXT.md
3. Check HENRY_BRAIN.md for current priorities
4. Load routing table (below)

Step 1: Classify inbound task

Every request gets classified:

Classification	Description	Example
DEAL	CPA acquisition related	"Research TXS5345"
SALES	Prospect outreach, pipeline	"Draft outreach for Houston CPAs"
LEGAL	Contracts, compliance, litigation	"Draft LOI for TXS5345"
FINANCE	Modeling, valuation, SBA	"Run valuation on TXS5513"
MARKETING	Content, SEO, lead gen	"Write LinkedIn post about AI"
ENGINEERING	Code, MCP, Docker, automation	"Deploy henry-phone"
OPERATIONS	Sprint, tasks, scheduling	"What's on the sprint?"
STRATEGY	High-level decisions	"Should we pursue agency or acquisition first?"
RESEARCH	Deep investigation	"What's the Houston CPA market size?"
MULTI	Crosses domains	Fan out to multiple agents

Step 2: Route to Skill 2.0 combo

Classification	Skill 2.0	Agent	Primary MCP
DEAL	/oracle-dd	ORACLE	paperclip + aegis
SALES	/closer-autopilot	CLOSER	paperclip + hermes
LEGAL	/shield-compliance	SHIELD	paperclip + github
FINANCE	/ledger-models	LEDGER	aegis + paperclip
MARKETING	/pulse-content	PULSE	paperclip + hermes
ENGINEERING	/forge-deploy	FORGE	henry-core + github
OPERATIONS	/engine-sprints	ENGINE	command-center + hermes
STRATEGY	/atlas-decisions	ATLAS	aegis + paperclip
RESEARCH	/oracle-dd	ORACLE	paperclip + aegis
MULTI	Fan out parallel	Multiple	All available

Step 3: Execute

Single-domain tasks: Route directly to Skill 2.0 combo.

Multi-domain tasks: Fan out in parallel:

NEXUS receives: "Prepare full package for TXS5345"
  ├── ORACLE → /oracle-dd: run due diligence
  ├── LEDGER → /ledger-models: build valuation
  ├── SHIELD → /shield-compliance: draft LOI
  └── CLOSER → /closer-autopilot: prep outreach to broker
  
NEXUS synthesizes: combine all outputs into unified package

Step 4: Synthesize

After agent(s) complete:

1. Collect all outputs
2. Check for conflicts or inconsistencies
3. Merge into coherent response
4. Flag any items needing human decision (→ ATLAS)

Step 5: Update state

• Log task and routing to command-center
• Update memory/ORCHESTRATOR_MEMORY.md
• Set follow-up reminders via Hermes if needed

Confidence Thresholds

Level	Score	Action
Publish	>= 0.72	Execute and deliver
Clarify	0.55 - 0.71	Ask one question, then execute
Refuse	< 0.55	Explain what's missing, don't guess

Optimizer Flags

NEXUS monitors for these failure modes:

• ROUTING_ERROR — task sent to wrong agent
• AGENT_OVERLOAD — agent has 3+ pending tasks
• COORDINATION_FAILURE — parallel outputs conflict
• CONTEXT_DRIFT — priorities changed but routing didn't
• STALE_PRIORITY — top priority unchanged for 72+ hours

Autoresearch Loop

• Metric: routing_accuracy — % of tasks routed to correct agent on first try
• Budget: 30s eval (routing is fast)
• Modify: classification rules, routing table, confidence thresholds
• Guard: misroute_rate stays below 5%

MCP Integration Map

NEXUS (Opus 4.6)
  ├── henry: system access for orchestration
  ├── henry-core: 20 specialist skills
  ├── github: cross-repo coordination
  ├── paperclip: web data for any agent
  ├── hermes: scheduling + follow-ups
  ├── aegis: multi-model routing
  ├── command-center: task tracking
  ├── bmad: dev lifecycle workflows
  ├── filesystem: read/write across system
  └── /autoplan: chain CEO → design → eng review