Do not rotate from a summary label.Preserve the provider response, request path, selected route, model id, and request id first.
Split by wire contract.
/chat/completions
DeepSeek and other compatible models.
/responses
Models such as GPT 5.6 Luna.
/messages
Anthropic-shaped Qwen and MiniMax routes.
Shared reference does not merge protocols.
Keep five facts separate.
/zen/go/v1, not ordinary Zen.
Cross-check the live official endpoint table.
Account policy can differ by model.
GET /models proves only discovery.rc.8 reads ids, names, and optional capacity fields. It does not import per-model request protocols.
Route the failure.
- Wrong prefix? Fix product identity.
- Wrong path? Split the protocol route.
- One family fails? Check entitlement.
- Region named? Review residency policy.
- Old Session? Start a fresh one.
- Key works elsewhere? Do not rotate blindly.
Primary evidence.
Keep the complete routing runbook.
The canonical guide includes three-route YAML, provider-response classification, region policy, verification order, and ten acceptance gates.
Read the full guideStar on GitHub