Enum ModelChunk 

pub enum ModelChunk {
    TextDelta {
        msg_id: String,
        delta: String,
    },
    ThinkingDelta {
        thinking_id: String,
        delta: String,
        signature: Option<String>,
    },
    ToolCallStart {
        id: String,
        name: String,
    },
    ToolCallInputDelta {
        id: String,
        delta: String,
    },
    ToolCallEnd {
        id: String,
        input: Option<Value>,
    },
    Done {
        stop_reason: String,
        usage: Option<HarnessUsage>,
    },
}

One token-level event from the model. The harness loop consumes a Stream<ModelChunk> and forwards TextDelta straight through as HarnessInternalEvent::AssistantTextChunk so callers see live generation; tool-call chunks are accumulated internally and only emitted once ToolCallEnd lands so the harness can dispatch the call with a complete serde_json::Value input.

Provider-agnostic: OpenAiCompatibleModelClient translates from OpenAI’s chat-completions SSE shape, and the future Anthropic client will project Messages-API events into the same enum. ScriptedModelClient synthesizes whichever sequence its scripted ModelResponse would have implied.

Variants

TextDelta

Fields

msg_id: String

delta: String

ThinkingDelta

Fields

thinking_id: String

delta: String

signature: Option<String>

Provider-emitted signature (Anthropic extended thinking) that MUST be round-tripped to history verbatim, otherwise some providers reject the next turn. None for OpenAI today.

ToolCallStart

First chunk of a tool call. The harness records (id, name) and starts buffering ToolCallInputDeltas under this id.

Fields

id: String

name: String

ToolCallInputDelta

Streaming JSON arguments for a previously-announced tool call. OpenAI emits these as a string that, when concatenated, is valid JSON. Harness accumulates these into a single string per id, then parses on ToolCallEnd.

Fields

id: String

delta: String

ToolCallEnd

Tool call finalised. input is the parsed JSON value if the provider sent the full object on this chunk (Anthropic), or a placeholder if the harness still needs to parse the accumulated ToolCallInputDelta buffer (OpenAI). Either way the harness treats input as authoritative when present.

Fields

id: String

input: Option<Value>

Done

Final chunk. usage carries the provider’s reported token count for this call (None if the gateway elides it).

Fields

stop_reason: String

usage: Option<HarnessUsage>

Trait Implementations

impl Clone for ModelChunk

fn clone(&self) -> ModelChunk

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ModelChunk

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

Formats the value using the given formatter.

impl PartialEq for ModelChunk

fn eq(&self, other: &ModelChunk) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for ModelChunk