Streaming and provider failover

AI Observatory aggregates streaming lifecycle data from the completed Laravel AI streamed response. It does not create one record per token.

When the response includes the required stream events, the model span records:

• Request start time.
• First token time.
• Response completion time.
• Time to first token in milliseconds.
• Final text and finish reason.
• Aggregate token usage.
• A non-recoverable stream error when exposed.

It also records stream_started, first_token_received, and response_completed instant events.

Interrupted streams#

If the client disconnects before Laravel AI dispatches AgentStreamed, the package cannot see a completion event. The trace remains running until ai-observatory:recover-stale marks it cancelled.

Provider failover#

Agent failover events include the active agent and are attached to its trace. The generic Laravel AI ProviderFailedOver event has no invocation ID, so it can be correlated only when an Observatory trace context is active.

AI Observatory records failover information exposed by official SDK events. It does not infer retries or provider switches that the SDK does not dispatch.