Skip to content

Duplicate effects

Duplicate-effect recognition is not cache.

Cache reuses a tool response. Duplicate-effect recognition skips an effect that the platform already recorded as done for the same company, area, agent, tool and semantic fingerprint.

A Curator declares which argument paths identify the effect. FuseOne hashes only those values, with the platform’s scope prefix around them. There is no fallback to “hash all arguments” because correlation IDs, timestamps and irrelevant fields would defeat the mechanism.

The model does not choose the tenant scope. Company, area, agent and tool are prefixes controlled by the platform.

The engine reads duplicate state before the Gate, lets the Gate decide, then reserves the fingerprint only after the proposed effect is executable. A successful tool return confirms the fingerprint. A failed tool return releases the reservation.

If another run is already executing the same effect, the waiting run retries with a stable dedupe_in_flight reason instead of pretending a human needs to intervene.