Routing Skill -- Message Router

Purpose

Route messages between agents (Jay, Z, Rick, Leroy) based on the dependency graph and business rules. All inter-agent communication flows through EM. EM is the message broker that ensures dependencies are respected, safety gates are enforced, and messages don't get lost or arrive out of order.

Core logic

Every message that arrives at EM gets a routing decision:

1. •Route immediately: Message prerequisites are met, target agent is healthy, message is safe to deliver.
2. •Hold pending input: Message prerequisites not yet met (e.g., Z data not ready for Jay submission). Hold and retry when prerequisites available.
3. •Escalate to human: Message involves safety-critical decision (blocking submission due to duplicate, profile ban) or high-uncertainty situation. Present to human for approval.
4. •Drop silently: Message is duplicate or superseded (e.g., two status checks from same agent in same minute). Log and discard.

Routing rules

All routing rules are configured in routing_rules.json. EM reads this file and applies rules to every incoming message.

Key rules:

• •URGENT priority: Inbound leads, critical alerts, agent failures → route immediately
• •Z safety gate: If Z is DEAD/ERROR, PAUSE all submissions (no data validation possible)
• •Dependency enforcement: Rick doesn't get submission requests until Jay submitted AND Z approved
• •Deadline tracking: Messages miss SLA → escalate to human
• •Queue management: If target agent queue is too deep, batch and hold new messages

Input/Output

Input: Message envelope

{
  "from": "Jay",
  "to": "Z",
  "type": "REQUEST",
  "priority": "NORMAL",
  "timestamp": "2025-02-15T09:00:00Z",
  "reference": "SUBM-C042-001",
  "payload": {
    "consultant_id": "C-042",
    "job_posting_id": "LI-9876",
    "vendor": "TrueBlue",
    "end_client": "Microsoft",
    "confidence_score": 8.2
  },
  "context": "Jay completed research, confidence >6.5, no staleness risk"
}

Output: Routing decision

{
  "message_id": "MSG-001",
  "original_from": "Jay",
  "original_to": "Z",
  "routing_decision": "ROUTE_IMMEDIATELY",
  "target": "Z",
  "reason": "Z has fresh data (2 hours old), no dependency violations, priority is NORMAL",
  "timestamp": "2025-02-15T09:00:15Z",
  "sla": "Expected Z response within 15 minutes",
  "logged": true,
  "posted_to_channel": "#agent-feed"
}

Usage

from message_router import route_message

# Route a single message
envelope = {
    "from": "Jay",
    "to": "Z",
    "type": "REQUEST",
    "priority": "NORMAL",
    "timestamp": datetime.now().isoformat(),
    "reference": "SUBM-C042-001",
    "payload": {...},
    "context": "..."
}

decision = route_message(envelope, system_state)
# decision = {
#   "routing_decision": "ROUTE_IMMEDIATELY",
#   "target": "Z",
#   "reason": "...",
#   ...
# }

Implementation details

• •Rules engine: Load routing_rules.json, evaluate against every message
• •State checks: Before routing, check target agent state (ACTIVE/IDLE/BUSY/SLOW/ERROR/DEAD)
• •Dependency graph: Maintain in-memory graph from routing_rules.json, validate prerequisites
• •Logging: Every routing decision logged to agent-feed (JSON line)
• •Escalation: If decision is ESCALATE_TO_HUMAN, post to #alerts with full context
• •SLA tracking: Record expected response time, track if response late

Error cases

Metrics tracked

• •Messages routed per hour (should be ~100+)
• •Routing decisions by type (IMMEDIATE, HOLD, ESCALATE, DROP)
• •Average hold time (should be <30 min for NORMAL priority)
• •Escalation count (should be <5/day)
• •Message loss (should be zero)