AI Provider Migration & Compatibility Readiness Review

The likely issue is a contract mismatch around AI provider migrations, model identifiers, authentication, tool calling, structured outputs, streaming, token/context limits, safety settings, latency, fallback behavior, evaluation, observability, rollout, and rollback. I would first define the invariant: An AI provider migration is ready only when required application behavior, safety boundaries, tool/schema contracts, observability, cost/latency constraints, and rollback evidence are mapped and tested against the replacement provider. Then I would trace the smallest owning boundary, keep VERIFIED facts separate from LIKELY/UNKNOWN state, apply or propose the narrowest repair, and add deterministic normal, boundary, failure, retry/cancellation, and recovery tests. The workflow preserves authorization, privacy, security policy, public interfaces, user-visible behavior, and documented ownership. With repository/workspace access, findings can map to exact paths and focused local validation; with partial evidence, the result remains an explicit diagnosis/evidence map with no invented execution claims.