Retrospective coverage for the October 16, 2026 issue.
A model can return an error and still increase a completion counter. As of the October 16, 2026 issue date, the pinned llm-d-router v1.0.0 session-state producer documentation explicitly counts a complete model error response as a natural completion. For someone building a multi-turn agent service, that distinction determines whether a dashboard describes responses ending or useful work getting done.
The producer tracks history for identities supplied by the agent-identity plugin and exposes session state to scheduling plugins. A turn is counted only after scheduling selects at least one target; the request then enters InFlightRequests. When its response lifecycle ends, it leaves that count. CompletedRequests has a narrower condition: the response must end naturally. These are lifecycle measurements, not judgments about the answer or the task.
Consider a hypothetical maintenance assistant assembling a repair recommendation. One request returns a complete model error; a second generates a potentially useful answer, but the client disconnects. Under the documented rules, the first contributes to CompletedRequests, while the disconnected request does not. Completion can therefore count an error while excluding a potentially useful response. Neither event establishes that a usable recommendation reached the technician; delivery and task success remain separate questions.
The same boundary limits cost interpretation. Token totals include usage reported by naturally completed responses; client disconnects, evictions and internal errors do not contribute because usage may be incomplete. A natural completion without usage data adds zero tokens. State is also local to one EPP replica, with idle eviction configurable. Those limits make this state unsuitable, by itself, as a durable, service-wide usage ledger.
The stronger product choice is to keep routing history separate from task outcomes. A workflow could associate these lifecycle signals with its own record of a recommendation delivered and accepted, rather than relabeling CompletedRequests as successful jobs. That additional record requires application design; the documented producer does not supply it. The bounded opportunity is clearer diagnosis: distinguish a response that ended normally from work that actually reached its intended result.
