Struct Failover 

pub struct Failover { /* private fields */ }

Tries the primary, and on a transient exhaustion tries each fallback in order — turn-local, so the next call starts from the primary again.

Two rules carry the design:

• Only errors carrying a retry::ProviderError in their chain are eligible, and only transient ones. The providers attach that marker exactly when nothing of the attempt was consumed — a mid-stream failure has already shown the user deltas, and re-issuing it would replay half an answer. Invalid/ContextOverflow fail identically everywhere, and Auth/Billing are the primary’s problem, not a routing decision.
• Each fallback answers as itself. The request’s model name is rewritten to the fallback’s own default — sending one server’s model name to another server was a real recorded bug (mecha replay -p).

Implementations

impl Failover

pub fn new( primary: Box<dyn Provider>, fallbacks: Vec<(String, Box<dyn Provider>)>, ) -> Self

Trait Implementations

impl Provider for Failover

fn id(&self) -> &str

Stable identifier used in config and --provider.

fn default_model(&self) -> &str

Model used when the caller doesn’t name one.

fn complete<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, req: &'life1 CompletionRequest, sink: Option<&'life2 StreamSink>, ) -> Pin<Box<dyn Future<Output = Result<CompletionResponse>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Run one turn. With sink, stream and emit deltas as they arrive; the accumulated response is still returned.