Crate nucel_agent_core 

Nucel Agent SDK — Core

Provider-agnostic traits and types for AI coding agents. Implement AgentExecutor and SessionImpl to add support for any coding-agent backend (CLI subprocess, HTTP server, in-process library, …).

Most application code should depend on the umbrella crate nucel-agent-sdk rather than this crate directly. Depend on nucel-agent-core only if you are:

• Writing a new provider and want the bare trait surface, or
• Embedding the abstraction into something that should stay agnostic to which providers are compiled in.

Minimal example

With a real provider plugged in (see crates/unified/examples/claude_basic.rs in the workspace), the shape of a session is:

use nucel_agent_core::{AgentExecutor, SpawnConfig};
use std::path::Path;
let session = executor.spawn(
    Path::new("/my/repo"),
    "What does this codebase do?",
    &SpawnConfig { budget_usd: Some(1.0), ..Default::default() },
).await?;

let resp = session.query("Now write me a one-line summary.").await?;
println!("{}", resp.content);
session.close().await?;

See also

• Workspace README
• docs/architecture.md
• CONTRIBUTING.md — how to add a new provider.

Re-exports

pub use error::AgentError;

pub use error::Result;

pub use executor::AgentCapabilities;

pub use executor::AgentExecutor;

pub use executor::AvailabilityStatus;

pub use executor::ExecutorConfig;

pub use executor::SpawnConfig;

pub use session::AgentSession;

pub use session::EventStream;

pub use session::SessionImpl;

pub use session::SessionMetadata;

pub use types::AgentCost;

pub use types::AgentResponse;

pub use types::CachePoint;

pub use types::ExecutorType;

pub use types::HookConfig;

pub use types::HookHandler;

pub use types::MessageEvent;

pub use types::PermissionMode;

pub use types::ToolCall;

pub use types::ToolResult;

Modules

error

executor

session

types