Module backend 

AgentBackend Contract

The AgentBackend trait is the control-plane boundary between Luft’s orchestration runtime and the agent execution environment. Prompt goes in, structured AgentResult comes out.

Implementing a Backend

use luft_core::contract::backend::*;
use async_trait::async_trait;

struct MyBackend;

impl MyBackend {
    fn new() -> Self { Self }
}

#[async_trait]
impl AgentBackend for MyBackend {
    fn id(&self) -> &'static str { "my-backend" }

    fn capabilities(&self) -> AgentCapabilities {
        AgentCapabilities {
            streaming: true,
            ..Default::default()
        }
    }

    async fn run(&self, task: AgentTask, ctx: RunContext)
        -> Result<AgentResult, BackendError>
    {
        // 1. Observe cancellation
        if ctx.cancel.is_cancelled() {
            return Err(BackendError::Cancelled);
        }

        // 2. Execute the agent task (your custom logic)
        let output = serde_json::json!({ "text": "hello" });

        // 3. Return structured result
        Ok(AgentResult {
            agent_id: task.agent_id,
            status: AgentStatus::Ok,
            output,
            findings: vec![],
            tokens_used: Default::default(),
            artifacts: vec![],
            logs: LogRef::default(),
            thread_id: None,
        })
    }

    fn as_any(&self) -> &dyn std::any::Any { self }
}

Structs

AgentCapabilities

AgentResult

AgentTask

Artifact

LogRef

McpEndpoint

MCP data-plane endpoint injected into an agent for structured reporting.

RunContext

Per-agent runtime context: cancellation + event sink + run association.

ToolPolicy

Tool permission policy. v0.1 translates to a backend’s acceptEdits + command allowlist.

Enums

AgentStatus

BackendError

Traits

AgentBackend

A pluggable agent backend (e.g. OpenCode via ACP). Prompt in, structured result out.