Enum Message 

pub enum Message {
    System {
        content: String,
    },
    User {
        content: String,
    },
    UserMultimodal {
        content: Vec<ContentBlock>,
    },
    Assistant {
        content: String,
        tool_calls: Vec<ToolCall>,
        thinking: Vec<ThinkingBlock>,
    },
    ToolResult {
        tool_use_id: String,
        content: String,
        provenance: Provenance,
    },
    ProviderOutputItems {
        protocol: String,
        items: Vec<Value>,
    },
}

A message in a multi-turn conversation.

The System variant exists so callers can express a first-class system prompt inside messages: Vec<Message> without threading it through the legacy context: Option<String> field on the request. Protocol handlers and local chat templates that have a native system-role slot (OpenAI, Anthropic, Gemini, Gemma 4, Qwen) emit it in the right place; ones that don’t can fold it into the first user turn.

Variants

System

A system prompt. Appears once, at the start of the conversation.

Fields

content: String

User

A user message (text only).

Fields

content: String

UserMultimodal

A user message with multimodal content (text + images + video + audio).

Fields

content: Vec<ContentBlock>

Assistant

An assistant response, possibly with tool calls.

Fields

content: String

tool_calls: Vec<ToolCall>

thinking: Vec<ThinkingBlock>

Extended-thinking blocks produced on this turn, preserved so they can be replayed VERBATIM on the next turn. Anthropic requires prior thinking blocks — including their opaque signature and empty-text blocks — be sent back unchanged, positioned before the tool_use blocks, or the same-model turn 400s (“thinking must be preserved”). Empty for providers/models without thinking. #[serde(default, skip_serializing_if)] keeps the wire + FFI backward-compatible: old JSON without the field deserializes, and turns without thinking add no bytes.

ToolResult

The result of executing a tool call.

Fields

tool_use_id: String

content: String

provenance: Provenance

Where content came from, relative to the runtime’s trust boundary.

#[serde(default)] keeps the wire and oplog formats backward compatible: transcripts recorded before this field existed deserialize as Provenance::Internal, which is what they were.

ProviderOutputItems

Provider-specific output items that need to round-trip verbatim across turns. The OpenAI Responses API returns reasoning blobs, encrypted_content, web-search results, etc. as opaque structured items; the next request must include them in the same form to preserve provider-side state.

protocol identifies the provider format that produced the items (currently "openai-responses"). Builder paths that don’t recognize the protocol drop the variant — there is no portable rendering across providers.

Fields

protocol: String

items: Vec<Value>

Trait Implementations

impl Clone for Message

fn clone(&self) -> Message

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Message

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Message

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl PartialEq for Message

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

Equality operator ==.

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

Inequality operator !=.

impl Serialize for Message

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

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for Message