Enum StreamEvent 

pub enum StreamEvent {
    TextDelta(String),
    ToolCallStart {
        name: String,
        index: usize,
        id: Option<String>,
    },
    ToolCallDelta {
        index: usize,
        arguments_delta: String,
    },
    Usage {
        input_tokens: u64,
        output_tokens: u64,
        cache_read_input_tokens: u64,
        cache_creation_input_tokens: u64,
    },
    StopReason(String),
    ProviderOutputItem(Value),
    Error(String),
    Done {
        text: String,
        tool_calls: Vec<ToolCall>,
    },
}

Events emitted during a streaming inference response.

Serialize/Deserialize are derived so the whole event stream can cross the on-device inference-worker IPC boundary (car-releases#74) losslessly — including the StopReason variant the daemon’s hand-rolled WS-runner JSON mapping omits. The representation is internal (same crate version on both ends of the pipe), so the default externally-tagged enum shape is fine.

Variants

TextDelta(String)

Partial text token from the model.

ToolCallStart

A tool call is starting (name known, arguments pending).

Fields

name: String

index: usize

id: Option<String>

ToolCallDelta

Partial tool call arguments (JSON fragment).

Fields

index: usize

arguments_delta: String

Usage

Provider-reported cumulative token usage observed mid-stream.

Anthropic emits this twice: once in message_start with the finalized input_tokens (plus a stub output_tokens: 1), and again in message_delta at end of stream with the real output_tokens. Consumers should prefer per-field monotonicity (see StreamAccumulator) rather than overwriting blindly.

Fields

input_tokens: u64

Uncached prompt tokens. For providers whose streamed usage reports the cached subset inside the prompt total (OpenAI), the parser subtracts it here so this stays the uncached portion — matching the non-streaming convention.

output_tokens: u64

cache_read_input_tokens: u64

Prompt-cache read (hit) tokens reported mid-stream (Anthropic message_start.usage.cache_read_input_tokens, OpenAI prompt_tokens_details.cached_tokens). 0 when uncached / absent.

cache_creation_input_tokens: u64

Prompt-cache write tokens reported mid-stream (Anthropic cache_creation_input_tokens). OpenAI has no write bucket. 0 when absent.

StopReason(String)

Provider-reported termination reason, surfaced mid-stream as the raw provider string (OpenAI finish_reason, Anthropic delta.stop_reason). A value of "length"/"max_tokens" signals the output was cut off at the token cap. Captured by StreamAccumulator and returned from StreamAccumulator::finish_with_usage.

ProviderOutputItem(Value)

Opaque provider output item that must be replayed verbatim on a later turn. The OpenAI Responses API emits completed reasoning items through response.output_item.done; CAR does not interpret or rewrite their id, status, summary, or encrypted_content.

Error(String)

Provider or transport failure observed after the HTTP stream began. This is terminal and must never be interpreted as a normal stop/success. Messages are sanitized at the provider adapter before entering the event.

Done

Stream is complete. Contains the final aggregated result.

Fields

text: String

tool_calls: Vec<ToolCall>

Trait Implementations

impl Clone for StreamEvent

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for StreamEvent

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for StreamEvent

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for StreamEvent

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.