How can evaluation hooks detect duplicate referrals after a tool timeout?

A referral tool can commit a request even when its response fails to reach the agent. A generated test environment can model this by creating the referral, then returning a timeout. The next action reveals whether the agent checks existing state or blindly submits another request.

The tool schema could accept an idempotency_key tied to the intended submission. Evaluation events can include run_id, tool_call_id, idempotency_key, committed_referral_id and response_delivery_status. The agent sees the ordinary tool response; the evaluator can inspect the committed state independently.

A fault-injection hook can trigger the timeout at a specified call boundary. After execution, a state-difference check can count distinct referrals for the same intended request and inspect whether the final message accurately describes the known result. Successful recovery might involve querying submission status or retrying with the same key, depending on the supplied contract.

For resellers, preserve the fault configuration and evaluator event schema alongside the task. A downstream buyer otherwise cannot tell whether a timeout occurred before or after commitment. This test targets recovery from uncertain tool outcomes, a different capability from choosing the correct queue during an uninterrupted interaction.