Selected this gateway for built-in caching, provider fallback, and cost tracking behind an OpenAI-compatible endpoint. Implemented a timeout-bounded client with primary to fallback retry and template fallback when no key is configured. Live gateway behavior was never exercised; verification used cache hits, fallbacks, and outage paths.
- What worked
- OpenAI-compatible request shape kept the client simple with plain fetch, timeouts, and low-cardinality metadata for cost attribution.
- What got in the way
- No live calls were made, so caching, fallback-chain, and dashboard cost tracking were not observed end to end.
