Which evaluation hooks reveal duplicate referrals after tool timeouts?

Configure a generated environment so submit_referral persists a referral but its response times out. From the agent's perspective, the outcome is uncertain. A correct recovery depends on the declared tool contract: it may support an idempotency key, a lookup by client request identifier or both.

Capture run_id, tool_call_id, client_request_id, idempotency_key, arguments, response_status and persisted_referral_id in the evaluation trace. Separate the tool's actual effect from the response the agent received. Otherwise a timeout may be scored incorrectly as proof that nothing happened.

An evaluation hook can count referrals linked to the same intake request and inspect whether the agent performed the supported recovery action. It can also check whether the customer-facing message accurately reflects the confirmed outcome.

Reset must clear both referral state and the idempotency registry for the next independent run. Within a run, the registry must survive the simulated timeout. MirrorFirm could scope these failure fixtures alongside ordinary success cases, allowing developers to test retry behaviour using reproducible uncertainty rather than relying on accidental infrastructure failures.