Streaming delivery and cost

Verify streaming support for the exact model/profile and protocol.

For an enabled OpenAI-compatible normal stream, consume deltas and wait for exactly one normal [DONE]. Treat errors, interruptions and pending settlement separately; a closed socket does not imply successful completion. Parse Anthropic terminal events according to their protocol.

Disconnecting after usable partial output does not imply a full refund. Reconcile the published rate contract, trusted measurement and delivery facts. Missing final usage must not become zero cost; retain the consumption request ID for usage/bills.

Agent Run has a separate event/tool contract. Compliance long-running operations generally use provider-request/operation polling. There is no universal cross-product SSE contract.