01 / Eligibility
Filter before scoring.
Every request resolves an organization, model alias, and immutable configuration snapshot before a provider is considered. Hard constraints run first so a cheap or fast route can never bypass a residency, capability, plan, or safety rule.
- Require the requested modalities, tools, context window, and structured-output support.
- Apply organization, country, region, provider, privacy, quota, and safety constraints.
- Exclude disabled, retired, unhealthy, or stale-price routes before scoring.
02 / Decision
Make the trade-off explicit.
Eligible routes are scored against the selected policy. Balanced routing combines health, observed latency, current cost, rate-limit headroom, and regional preference; fastest and lowest-cost policies change the weights without removing safety constraints.
{
"model": "openwaya/auto",
"openwaya": { "routing_policy": "balanced" },
"metadata": { "workload": "customer-support" }
}03 / Recovery
Fail over without losing the story.
Fallbacks are ordered, bounded, and failure-aware. Retry a different route for timeouts, rate limits, and provider unavailability; do not retry validation, authentication, balance, safety, or idempotency conflicts.
- Preserve one request ID across all provider attempts.
- Respect the request deadline and maximum attempt count.
- Record exclusions, scores, selected route, failure class, and final outcome.