Day 2 connected the AI Gateway to the real model-backed parts of ClaimFlow AI. Day 1 created the gateway schema and base logging utilities. Day 2 made the gateway the control point for actual AI calls, so extraction, coverage answering, and model-backed agent planning are now traceable through ai_call_logs.

Goal

The goal was to ensure that every important model call records:

• claim run ID
• trace ID
• AI call kind
• provider and model
• model version
• prompt version
• schema version
• latency
• token usage
• estimated cost
• final status
• retryability
• failure reason, if any

This gives ClaimFlow a production-style audit trail for model behavior.

Gateway Wrapper

A new wrapper function, callModelThroughGateway(), was added in:

packages/gateway/ai-gateway.ts