Crate af_llm 

af-llm — unified async LLM access for the Agent Factory platform.

Provider-neutral model transport and request types. Provides:

• LlmClient — async chat completions against any OpenAI-compatible endpoint, with a hard per-request timeout and a shared CircuitBreaker.
• Strongly-typed request/response models (CompletionRequest, CompletionResponse, ChatMessage, Tool, …) — the Pydantic equivalent, validated at the wire boundary.
• parse_json — pull structured output out of fenced LLM text into any serde type.

Example

use af_llm::{LlmClient, LlmConfig, CompletionRequest, ChatMessage};

let client = LlmClient::new(LlmConfig::new("http://localhost:4000/v1", ""))?;
let req = CompletionRequest::new(
    "deepseek/deepseek-chat",
    vec![ChatMessage::user("Say hi in one word.")],
);
let resp = client.complete_stream_single_attempt(&req, |_, _| {}).await?;
println!("{:?}", resp.first_content());

Re-exports

pub use circuit_breaker::CircuitBreaker;

pub use circuit_breaker::State as CircuitState;

pub use circuit_breaker::Status as CircuitStatus;

pub use client::LlmClient;

pub use client::LlmConfig;

pub use client::DEFAULT_TIMEOUT;

pub use error::LlmError;

pub use error::Result;

pub use json::parse_json;

pub use json::strip_code_fence;

pub use stream::StreamDelta;

pub use types::AssistantBlock;

pub use types::ChatMessage;

pub use types::Choice;

pub use types::CompletionRequest;

pub use types::CompletionResponse;

pub use types::FinishReason;

pub use types::FunctionCall;

pub use types::FunctionDef;

pub use types::InputImage;

pub use types::ReasoningEffort;

pub use types::Role;

pub use types::StreamOptions;

pub use types::Tool;

pub use types::ToolCall;

pub use types::ToolChoice;

pub use types::Usage;

Modules

circuit_breaker

In-process circuit breaker for LLM calls.

client

Async LLM client — port of agent_core/llm/client.py::llm_completion.

error

Error type for the LLM client.

images

Governed binary resource resolution at the provider boundary. Image references remain durable; signed projections exist only for one request.

json

Helpers for coaxing structured JSON out of free-form LLM text.

stream

OpenAI-compatible SSE parsing for streaming chat completions.

types

Strongly-typed OpenAI-compatible chat-completion request/response shapes.