Struct RunPolicy 

pub struct RunPolicy {
    pub limits: RunLimits,
    pub unknown_tool: UnknownToolPolicy,
    pub invalid_args: InvalidArgsPolicy,
    pub retry: RetryPolicy,
    pub fallback: Option<FallbackPolicy>,
    pub default_response_format: Option<ResponseFormat>,
    pub capture: PayloadCapture,
    pub cache: CachePolicy,
    pub error_on_empty_response: bool,
    pub truncated_empty_retries: u32,
}

Fields

limits: RunLimits

Hard run limits enforced fail-closed by the agent loop.

unknown_tool: UnknownToolPolicy

How the loop reacts to a model call for an unregistered tool.

invalid_args: InvalidArgsPolicy

How the loop reacts when a registered tool’s arguments fail schema validation.

retry: RetryPolicy

Retry policy applied to each model call.

fallback: Option<FallbackPolicy>

Optional ordered model fallback chain.

default_response_format: Option<ResponseFormat>

Response format attached to every model request when set.

capture: PayloadCapture

Whether the loop captures model/tool payloads onto completion events.

Defaults to PayloadCapture::default (payload-free), preserving the privacy-preserving behavior where events carry only ids and usage.

cache: CachePolicy

Default caching policy for the run.

The loop consults CachePolicy::response_cache_enabled only when a ResponseCache is attached to the harness and the per-call crate::harness::model::ModelRequest::cache_policy does not override it. A request-level cache_policy always wins over this default.

error_on_empty_response: bool

When true, an empty provider completion in the finalization branch (no text, no tool calls, and no structured output) fails the run with crate::error::TinyAgentsError::EmptyResponse instead of terminating with a blank final answer.

Defaults to false to preserve the historical behavior for callers who rely on empty finals; opt in to turn a silent blank success into a typed error the caller can re-prompt on.

truncated_empty_retries: u32

Number of automatic retries when a model call returns a truncated empty completion — finish_reason == "length" with no visible text, no tool calls, and no structured output.

This is the failure mode of local reasoning models (for example qwen3 via Ollama) that intermittently spend the entire token budget on the hidden reasoning channel and emit nothing usable. Because such a response is useless to every caller and the failure is stochastic, retrying — with a doubled token budget when the request set one (capped at 4x the original), or a plain retry when it did not — is strictly better than surfacing a blank success.

The retry runs before Self::error_on_empty_response; only once these retries are exhausted does that guard (if enabled) apply.

Defaults to 1 (one retry, two attempts total). Set to 0 to disable for exact-replay callers that must not re-issue a call.

Trait Implementations

impl Clone for RunPolicy

fn clone(&self) -> RunPolicy

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RunPolicy

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for RunPolicy

fn default() -> Self

Returns the “default value” for a type.

impl PartialEq for RunPolicy

fn eq(&self, other: &RunPolicy) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for RunPolicy