ModelChunk

"delta": a piece of streamed text, shown live and not recorded on its own.

The new text since the last delta.

"part": one finished output part.

The finished part (text, tool call, reasoning or hosted tool result). Parts are recorded in the order they arrive.

"done": the last chunk of a successful send, with its stop reason and usage.

Only end_turn, stop_sequence and refusal complete a turn; tool_use dispatches; max_tokens continues; pause_turn re-requests with the paused content as-is, up to context.max_pause_continuations; context_window_exceeded ends the turn context_exhausted; other (a reason the adapter can't name) ends it error.

Per-field measured-or-unknown usage.

"rejected": the send failed before any content; always the stream's only and last chunk.

The terminal error of a send: exactly Model.send returns.errors. Provider classes are recorded as model_attempt_abandoned{provider_outcome: failed}; stale_epoch appends nothing (the writer lost its lease); content_unsupported, continuation_unsupported and transport_fence_unsupported are recorded as not_sent and end the turn with that code, never retried.

The provider's HTTP status code (for example 429 or 529), recorded with the abandoned attempt. Omitted when there was no HTTP response.

The provider's retry-after hint in milliseconds. When present the retry waits this long instead of backing off, and a hint longer than the retry policy's max_retry_after_ms ends the turn. Omitted when the provider gave none.

not_billed only when provider_outcome is not_sent or the adapter's declared billing contract (in the system prompt adapter settings) states this rejection class is not billed; a status code alone is not proof.