neuron-turn-kit 0.4.0

neuron-turn-kit

Planning and execution primitives for a single agent turn — focused on sequencing and concurrency only. These traits are intentionally narrow and do not bake in provider streaming, hooks, or operator concerns; those live at higher layers (e.g., operator implementations).

Contents:

• Concurrency and ConcurrencyDecider — classify tool calls as Shared vs Exclusive
• ToolExecutionPlanner and BarrierPlanner — sequence tool calls into batches
• SteeringSource — optional source of mid-loop steering messages
• BatchExecutor — run batches, with a simple sequential baseline executor

Example: planning with a barrier.

use neuron_turn_kit::{BarrierPlanner, Concurrency, ConcurrencyDecider, ToolExecutionPlanner};

struct SharedIfStartsWith;
impl ConcurrencyDecider for SharedIfStartsWith {
    fn concurrency(&self, tool_name: &str) -> Concurrency {
        if tool_name.starts_with("shared_") { Concurrency::Shared } else { Concurrency::Exclusive }
    }
}

let calls = vec![
    ("1".to_string(), "shared_a".to_string(), serde_json::json!({})),
    ("2".to_string(), "exclusive".to_string(), serde_json::json!({})),
    ("3".to_string(), "shared_b".to_string(), serde_json::json!({})),
];
let planner = BarrierPlanner;
let plan = planner.plan(&calls, &SharedIfStartsWith);
assert!(matches!(plan[0], neuron_turn_kit::BatchItem::Shared(_)));