LLM agents cache recovery suggestions after API errors and reuse them in later episodes, cutting token usage and model calls. Server‑side data drift turns those caches stale; the usual fix is to re‑derive on every episode, which erases the savings. We introduce invalidation contracts, a protocol layer that attaches a version stamp and cacheability hint to each suggestion so the client can evict stale entries after drift without trial‑and‑error. The contract splits realized savings into validity (fraction of cached suggestions that stay correct after drift) and compliance (fraction the planner uses on the first try). Validity depends only on the protocol and is vendor‑independent; compliance depends on the planner model: identical request bytes yield 100% first‑try compliance on Claude Haiku 4.5 but about 11% on Claude Sonnet 5, which conservatively rejects fixes that add fields not present in the original request. We evaluate seven models, three serving paths, two domains, and roughly 9,400 episodes. Row‑level invalidation raises compliance by 0‑66.7 percentage points, 55.6‑66.7 on three models, and recovers 29‑33% of baseline token cost on four models; table‑level invalidation destroys co‑located entries and drops post‑drift first‑try rates to 0% on five models. Eviction precision is 1.00 at row granularity under the row‑level oracle of Section 4.1. The contract adds 15% to the response payload. Version‑stamp validity is deterministic by construction and yields identical results across all models and paths, with zero contract failures in the whole evaluation.
Review