Routing templates

Save a complete template, then bind it to a key, team or organization. Resolution selects the first whole template in key → team → organization → platform order; it does not merge fields.

Example

{
  "model_routes": [{
    "model": "chat-main", "strategy": "traffic-split",
    "allocations": [
      {"deployment_id": "deployment-a", "weight": 3},
      {"deployment_id": "deployment-b", "weight": 7}
    ]
  }],
  "retry_policy": {
    "max_attempts": 2, "timeout_seconds": 60,
    "failure_threshold": 3, "cooldown_seconds": 60
  },
  "routing_groups": [{
    "group_name": "chat-pool", "models": ["chat-main", "chat-fast"],
    "routing_strategy": "cost-based-routing"
  }],
  "fallbacks": [{"chat-pool": ["chat-backup"]}],
  "context_window_fallbacks": [{"chat-main": ["chat-long"]}],
  "content_policy_fallbacks": []
}

Contract

Field	Behavior
model_routes	Exact public model rules; strategy required
allocations omitted	Inherit model default weights
allocations: []	All compatible deployments use weight 1
explicit weight 0	Exclude the deployment
max_attempts	Attempts per deployment, including the first
timeout_seconds	Timeout per attempt
failure_threshold: 0	Disable passive cooldown
cooldown_seconds: 0	Use 60 seconds

All four retry_policy fields are required. Reject unknown fields, invalid references, duplicate models, negative/nonfinite weights and fallback cycles. Streaming output prevents switching providers. Ambiguous task creation must not be repeated.

Source and verification

internal/router/settings.go

go test ./internal/router -count=1

Test prerequisites · Next: routing-groups