Module types 

Agent message shapes.

AgentMessage is the canonical typed conversation transcript. Apps that need richer shapes either extend the Custom variant (kind-tagged JSON payload) or wrap the entire enum in their own outer enum. The loop never peeks into Custom — it’s pass-through context.

The discriminator lives on the role tag, content is typed, and the loop avoids Value walking via field-name strings.

Structs

AgentContext

Snapshot of agent state passed into the loop.

AssistantContent

Assistant-message content. Carries text, hidden reasoning blocks, and tool call requests in source order.

ImageContent

ReasoningDetailsContent

Persistent envelope for provider-native reasoning items on an assistant turn. The wire shape is Vec<Value> matching OpenRouter’s reasoning_details[] schema (the broadest typed surface across providers); as_items lifts it to typed crate::reasoning::ReasoningItems for codec operations and from_items projects typed items back. The details field stays the source of truth so persisted trajectories round-trip byte-exact, even when a future provider sends shapes the typed enum doesn’t yet recognize.

RunIdentity

Identity of one agent run.

TextContent

ToolResultContent

Tool result content. Multiple blocks support image-returning tools.

Usage

Provider-reported token accounting for one LLM call.

Enums

AgentMessage

One message in the conversation transcript.

AssistantBlock

Blocks an assistant message can carry.

StopReason

Why an assistant turn ended.

ToolResultBlock

UserBlock

UserContent

User-message content. Plain text is the common case; the block form supports images, attachments, and other multimodal inputs.