Struct StreamDelta 

pub struct StreamDelta {
    pub content: String,
    pub thinking: Option<String>,
    pub usage: Option<Usage>,
    pub model: Option<String>,
    pub stop_reason: Option<StopReason>,
    pub is_final: bool,
}

An incremental chunk from a streaming language model response.

Fields

content: String

Incremental text content (may be empty on non-content events).

thinking: Option<String>

Extended-thinking / reasoning text emitted on this chunk. Streamed alongside (or before) content on providers that expose chain-of-thought as a distinct content block:

• Bedrock Converse: populated from ContentBlockDelta::Reasoning events (gpt-oss, DeepSeek R1, Magistral, MiniMax M2, Nemotron-super on the Runtime endpoint).

None on the OpenAI Chat Completions surface (reasoning isn’t streamed per OpenAI spec — it shows up only in completion_tokens_details.reasoning_tokens accounting), and on adapters that haven’t been extended to surface reasoning yet (Anthropic thinking blocks, Mantle Responses reasoning events).

usage: Option<Usage>

Token usage statistics. Presence and timing depend on the provider:

• Anthropic: populated on message_delta events (mid-stream, sometimes on multiple events as token counts evolve). The final message_stop event carries None.
• OpenAI / Ollama: populated only on the final pre-[DONE] chunk, because we set stream_options.include_usage = true on the request. Without that flag the field would be None for every chunk.
• Default generate_stream impl (non-streaming providers wrapped into a one-item stream): populated on the single delta which is also is_final.

Consumers that need cost tracking should accumulate or take the last non-None value rather than expect usage on every chunk.

model: Option<String>

The model name — typically only present on the first chunk.

stop_reason: Option<StopReason>

Normalized reason generation stopped — populated on whichever chunk carries the wire-level stop signal:

• OpenAI Chat Completions: the final content chunk with a non-null finish_reason.
• Anthropic Messages: the message_delta event (which also carries final usage).
• Bedrock Converse: the MessageStop event (precedes the terminating Metadata chunk that carries usage).
• Mantle Responses: the response.completed event (also carries usage). response.incomplete events with recoverable reasons (max_output_tokens, content_filter) surface their stop reason here too instead of being raised as errors.

None on intermediate content deltas and on providers that don’t surface this on streams. Downstream consumers comparing against max_tokens should prefer this field when present rather than inferring truncation from output-token counts.

is_final: bool

Whether this is the final chunk in the stream.

Trait Implementations

impl Clone for StreamDelta

fn clone(&self) -> StreamDelta

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for StreamDelta

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.