RuntimePolicy

The single struct that bounds every dimension of the runtime.

RuntimePolicy is the operational configuration consumed by AgentRuntime::new. It holds limits, budgets, timeouts, and admission-control parameters. It is immutable after construction; changing a policy means constructing a new runtime.

The full file is src/runtime/policy.rs.

Definition

pub struct RuntimePolicy {
    pub max_iterations: usize,            // default 16
    pub max_input_tokens: usize,          // default 128_000
    pub max_total_tokens: usize,          // default 256_000
    pub provider_timeout: Duration,       // default 60s
    pub tool_timeout: Duration,           // default 30s per tool
    pub session_gate: SessionGateConfig,
    pub input_admission: InputAdmissionConfig,
    pub router: RouterPolicy,
    pub event_buffer: usize,              // default 1024
}

Defaults

Field	Default	Effect
max_iterations	16	The Turn FSM fails after this many iterations per run.
max_input_tokens	128_000	ContextPipeline rejects inputs larger than this.
max_total_tokens	256_000	The cumulative token budget for a single run.
provider_timeout	60 s	Wall-clock cap for a single model call.
tool_timeout	30 s	Wall-clock cap for a single tool execution.
event_buffer	1024	Broadcast channel buffer for AgentEvent.

SessionGateConfig, InputAdmissionConfig, and RouterPolicy have their own default impls.

Construction

use std::time::Duration;
use behest::runtime::policy::RuntimePolicy;

let policy = RuntimePolicy {
    max_iterations: 32,
    max_total_tokens: 1_000_000,
    provider_timeout: Duration::from_secs(120),
    ..Default::default()
};

The ..Default::default() pattern is the recommended idiom; it lets you override only the fields you care about.

See also

• AgentRuntime — the consumer.
• Turn FSM — the CheckingPolicy state reads max_iterations and max_total_tokens.
• ModelRouter — reads provider_timeout.
• CompactionService — triggered when max_input_tokens is exceeded.