Crate oino_ai 

oino-ai: the LLM data model for Oino.

Every type serialises to a fixed JSON shape (field names, tags, optionality), so sessions, RPC records and WASM-boundary payloads stay interchangeable across processes and tools that speak the same format. The crate has no async runtime and builds for wasm32-wasip2; guests and host share these types (D5).

API map

Module	Contents
primitives	scalars: TimestampMs, ThinkingLevel (+ clamping), StopReason, Usage
content	content blocks: TextContent, ThinkingContent, ImageContent, ToolCall, UserContent, AssistantContent, UserContentField
message	Message, AgentMessage, SystemMessage, AssistantMessage, …
model	Model, ModelKey, ThinkingLevelMap, costs and limits
stream	AssistantMessageEvent, ProviderEvent, PartialAssembler
transcript	system-message replay, transcript::diff_sections, collapsing
transform	transform::transform_messages
json	JsonObject, JsonValue, json::json_semantic_eq

Unknown JSON fields are ignored on read; optional fields are omitted on write exactly where the wire format omits them; integral f64 costs serialise without a fraction like JSON.stringify.

Re-exports

pub use content::AssistantContent;

pub use content::ImageContent;

pub use content::TextContent;

pub use content::ThinkingContent;

pub use content::ToolCall;

pub use content::TypedText;

pub use content::UserContent;

pub use content::UserContentField;

pub use content::UserTextField;

pub use json::JsonObject;

pub use json::JsonValue;

pub use message::AgentMessage;

pub use message::AssistantDiagnostic;

pub use message::AssistantMessage;

pub use message::BashExecutionMessage;

pub use message::BranchSummaryMessage;

pub use message::CompactionSummaryMessage;

pub use message::ConstrainedSampling;

pub use message::ConstrainedSamplingConfig;

pub use message::CustomMessage;

pub use message::DeferredHandle;

pub use message::DiagnosticError;

pub use message::FalseLiteral;

pub use message::Message;

pub use message::NestedStatus;

pub use message::NestedToolCallRecord;

pub use message::NestedToolCalls;

pub use message::Sections;

pub use message::StrictMode;

pub use message::SystemMessage;

pub use message::ToolDecl;

pub use message::ToolRef;

pub use message::ToolResultMessage;

pub use message::UserMessage;

pub use model::AnyModel;

pub use model::ClassifierModel;

pub use model::ImageInputLimits;

pub use model::ImageModel;

pub use model::ImageResize;

pub use model::InputLimits;

pub use model::MapEntry;

pub use model::Model;

pub use model::ModelCost;

pub use model::ModelCostTier;

pub use model::ModelKey;

pub use model::ModelKindChat;

pub use model::PromptCache;

pub use model::ThinkingLevelMap;

pub use primitives::CacheRetention;

pub use primitives::InputModality;

pub use primitives::StopReason;

pub use primitives::ThinkingLevel;

pub use primitives::TimestampMs;

pub use primitives::UnknownThinkingLevel;

pub use primitives::Usage;

pub use primitives::UsageCost;

pub use stream::AssistantMessageEvent;

pub use stream::Context;

pub use stream::DoneReason;

pub use stream::ErrorReason;

pub use stream::PartialAssembler;

pub use stream::ProviderEvent;

pub use stream::StreamOptions;

pub use stream::ThinkingBudgets;

pub use stream::Transport;

Modules

content

Content blocks (TextContent, ThinkingContent, ImageContent, ToolCall).

json

JSON aliases and serde helpers that keep the wire format identical to a JavaScript host.

message

The LLM Message and the coding-agent AgentMessage union.

model

The Model and related catalogue types.

primitives

Scalar types: timestamps, thinking levels, stop reasons, usage.

stream

Streaming events, the partial-free provider form crossing the WIT boundary (ProviderEvent) and request types (Context, StreamOptions).

transcript

System-message replay and the diffSystemPromptSections patch helper.

transform

Provider-side transcript transformation: drop foreign reasoning and signatures, downgrade images for text-only models, synthesize missing tool results, and hold mid-tool-call system messages until the results are in.