Run Telemetry Model ID Extraction#

Overview#

When assistant-ui reports a run to the cloud backend, it includes a model_id field derived during telemetry extraction. Model ID is never read directly from the AI SDK's native response object — it is only available if the application explicitly forwards it through messageMetadata. This is the core gap: if a route omits the messageMetadata callback (or omits the finish-step branch), model_id will be absent from all cloud run reports.

How Model ID Flows In#

The model ID must be injected by the API route via the AI SDK's messageMetadata callback in toUIMessageStreamResponse. The cloud template route shows the canonical pattern:

• On finish-step, part.response.modelId is surfaced and returned as { modelId: part.response.modelId }.
• On finish, token usage is forwarded as { usage: part.totalUsage }.

This metadata is serialized into the AI SDK UI message's metadata field and flows to the client.

Extraction Logic#

packages/cloud-ai-sdk — extractRunTelemetry.ts#

The extractRunTelemetry function (used by the cloud-ai-sdk package) processes the last assistant UIMessage and reads model_id only from message.metadata.modelId 1:

const metadata = assistant.metadata as Record<string, unknown> | undefined;
const modelId =
  typeof metadata?.modelId === "string" ? metadata.modelId : undefined;

There is no fallback to any native AI SDK response property. If messageMetadata was not configured on the route, modelId is undefined and is omitted from the returned RunTelemetryData 2.

packages/core — AssistantCloudThreadHistoryAdapter.ts#

The history adapter contains two parallel extraction paths 3:

Both paths funnel through extractModelId, which checks two locations in message metadata:

1. metadata.modelId (top-level)
2. metadata.custom.modelId (nested custom field)

Neither path reads the AI SDK's native response.modelId directly from the stream or response object — they only read what was forwarded into the message metadata by the route.

For aui/v0, the lookup is 4:

const modelId =
  msg.metadata?.modelId ??
  (typeof msg.metadata?.custom?.modelId === "string"
    ? msg.metadata.custom.modelId
    : undefined);

Final Report#

The extracted modelId (or undefined) is included in the _sendReport payload as model_id 5 and submitted via this.cloudRef.current.runs.report(...) 6.

The Gap#

The AI SDK exposes the actual model that served the request as part.response.modelId on the finish-step event. This is the provider-confirmed model ID (e.g., the resolved alias like gpt-4o-2024-08-06). The extraction layer never reads this directly — it only reads whatever the application chose to pass through messageMetadata.

Consequences:

• Routes that omit the messageMetadata callback entirely will produce run reports with no model_id.
• Routes that only handle finish (usage) but skip finish-step will also omit model_id.
• If a route forwards modelId for only some steps in a multi-step run, only the last step's metadata is used (since aggregateAiSdkV6RunSteps overwrites metadata on each assistant message it encounters) 7.

Required Route Configuration#

To populate model_id in telemetry, every API route must include the finish-step branch in messageMetadata 8:

messageMetadata: ({ part }) => {
  if (part.type === "finish-step") {
    return { modelId: part.response.modelId };
  }
  if (part.type === "finish") {
    return { usage: part.totalUsage };
  }
  return undefined;
},

Without this, the cloud dashboard will show runs with missing model information. See the cloud template route for the full working example.

Key Files#