Exactly-Once Semantics for LLM Function Calling via Idempotency Patterns at Scale

LLM tools induce side effects (payments, emails, mutations) under retries, timeouts, and re-deliveries. Without explicit idempotency, operations duplicate. We present a practical framework for application-level exactly-once semantics based on idempotency keys and an atomic get-or-execute primitive over durable storage.

We formalize failure modes, prove correctness under at-least-once delivery and bounded retry horizons, and provide deployable database- and cache-backed patterns. We discuss key generation strategies, retention/TTL/GC, and broker EOS integration (Kafka, Pub/Sub, SQS FIFO). The design composes with typed-effects, integrates with observability, and requires no heavyweight distributed transactions.