Actor is the "org/sub" identity the run was executed and billed AS. Empty means there was no PERSON — a schedule or a service token — which is a different fact from "we do not know", and the difference is what an audit asks about.
What an operator needs to answer "what ran, for whom, and what did it do" — and, through traceId, to leave this record for the waterfall of the very same run rather than a search that hopefully lands near it. Agent is on the row because the org-wide feed lists runs across agents, and a run that cannot name its agent is an orphan in exactly the view built to make sense of many of them. Every field is omitempty: a run recorded before these columns existed reports absence rather than a zero it never measured.
CompletionTokens is the same measurement for what the model produced, on the same final completion. It is a count of TOKENS, not of turns and not of money.
ID is the run's handle, minted as "run_" + 32 hex characters. It is the key the metering ledger records this run's per-round token spend under, so it is how a bill and a run are joined.
Model is the model that actually SERVED this run, which is not always the one the agent is defined on — a failover records what answered. Normalized to our name on the way out; the stored row is left exactly as it happened, because a run is a record and rewriting it would be worse than the name it carries.
Output is what the model produced. Empty on an error run, and empty is also a legitimate answer from a run that succeeded with nothing to say — Status is what separates those.
PromptTokens is what the gateway reported for the run's FINAL completion, and only that one — a tool loop's earlier rounds are the metering ledger's account, joined by this run's id. Reading it as the run's total spend undercounts a loop.
Status is the run's outcome, and there are exactly two: "ok" when the model answered, "error" when it did not. It is written when the run ends, so no row here is in flight.
ToolCalls is how many tool dispatches the run made — a count of ACTIONS, which is a different measurement from the token counts above and from the turns a build reports. Zero is a run that answered straight from the model.
TraceID is the trace this run IS, so the record and its spans are one thing to move between: it opens the waterfall for THIS run rather than a search that lands near it. Empty when the process had no tracer, never a fabricated id.
