Crate rpi_ai 

Mirrors packages/ai/src — the unified, multi-provider LLM layer.

Provider-agnostic types (Content, Message, AssistantMessage, AssistantMessageEvent, Context, Tool, Usage) live in types; model-layer configuration (Model, compat structs) lives in model. Schema validation + LLM-sloppy-JSON coercion (mirroring validation.ts) is in schema; the SPSC streaming queue mirroring event-stream.ts is in event_stream; the Provider trait seam is in provider; concrete providers live under providers (faux now, anthropic in M3).

Every module’s doc comment names the TS file it mirrors.

Re-exports

pub use error::AiError;

pub use event_stream::AssistantMessageEventStream;

pub use event_stream::AssistantMessageEventStreamProducer;

pub use model::Model;

pub use model::StreamingProtocolCompat;

pub use provider::CacheRetention;

pub use provider::ContextPatch;

pub use provider::NoopProviderHooks;

pub use provider::Provider;

pub use provider::ProviderHooks;

pub use provider::SimpleStreamOptions;

pub use provider::SimpleStreamOptionsPatch;

pub use provider::DEFAULT_LLM_API_TIMEOUT;

pub use schema::validate_tool_arguments;

pub use types::Api;

pub use types::AssistantMessage;

pub use types::AssistantMessageEvent;

pub use types::Content;

pub use types::Context;

pub use types::DeferredHandle;

pub use types::DoneReason;

pub use types::ErrorReason;

pub use types::ImageContent;

pub use types::InputModality;

pub use types::Message;

pub use types::ModelCost;

pub use types::ModelCostRates;

pub use types::ModelCostTier;

pub use types::ProviderId;

pub use types::Schema;

pub use types::StopReason;

pub use types::TextContent;

pub use types::ThinkingBudgets;

pub use types::ThinkingContent;

pub use types::ThinkingLevel;

pub use types::ThinkingLevelMap;

pub use types::Tool;

pub use types::ToolCall;

pub use types::ToolCallType;

pub use types::ToolResultMessage;

pub use types::Usage;

pub use types::UsageCost;

pub use types::UserContent;

pub use types::UserMessage;

Modules

constrained

Constrained sampling — grammar-constrained tool output formatting.

error

Mirrors the error surfaces implied by packages/ai/src — no single TS file; the variants collect every failure mode the ai crate can surface to a caller.

event_stream

Mirrors packages/ai/src/utils/event-stream.ts — the SPSC async event stream that carries AssistantMessageEvents from a provider to the agent loop, plus the result() future resolving to the final AssistantMessage.

frames

Assistant message frames — a compact, durable encoding of one assistant stream.

http

HTTP client construction: system-proxy resolution and pooled-connection idle timeout.

images

Image generation API.

json_parse

Mirrors packages/ai/src/utils/json-parse.ts — JSON repair + streaming partial-JSON parsing used by the Anthropic SSE mapper to recover tool-call arguments that arrive as malformed / incomplete JSON deltas.

model

Mirrors packages/ai/src/models.ts — Model<TApi> + the compatibility structs.

provider

Mirrors packages/ai/src/models.ts::Provider — the trait every concrete provider (faux, anthropic, …) implements. The agent loop never touches a provider’s wire format directly; it calls stream_simple and consumes the returned AssistantMessageEventStream.

providers

Concrete providers. Mirrors packages/ai/src/api/* + the in-tree faux provider (packages/agent/src/faux.ts, hoisted here so every crate’s tests can reuse it without depending on pi-agent).

schema

Mirrors packages/ai/src/utils/validation.ts — JSON-Schema validation of tool arguments plus a hand-written coerce_with_json_schema layer that tolerates LLM-sloppy JSON (stringified numbers, "true", null on optional fields, …).

types

Mirrors packages/ai/src/types.ts — the core LLM type contract every other crate consumes.