ADR 0028: Keep the last successful target for the provider cache lifetime

• Status: Accepted
• Date: 2026-10-25
• Supersedes: none

Context

A return to the primary after fallback can lose a warm prompt cache on the backup.

Decision

Read the sticky entry with admission and write the successful target with settlement. Scope the entry by tenant and session. Its TTL is providers.<name>.cache_ttl. Validate the stored target hash against the current route. Breakers, cooldowns and retry rules still apply. Record target switches and an upper bound for their cache loss.

Alternatives

Always use the first target. This discards warm backup caches after recovery. Add a separate Redis read. This adds another request-path round trip.

Consequences

Concurrent requests can race to select a target. Entries are advisory. Different routes validate the saved target before use. Cache loss includes new input and is not an exact counterfactual price.