Outages, rate limits, latency spikes and price changes: the specific failure modes that had already hit similar integrations.

Started with the request and response layer both providers would sit behind, before writing any routing logic.

Health-based provider selection came first; failover without losing context came once routing was proven stable.

Token and spend logging went in alongside the routing logic, not after someone asked for a cost report.

Killed the primary provider mid-conversation repeatedly until failover held up without manual intervention.

An automatic switch still needs someone to understand it at 2am, including what’s safe to ignore.