Trait Provider

Source

pub trait Provider: Send + Sync {
    // Required methods
    fn api_schema(&self) -> &str;
    fn complete<'life0, 'life1, 'async_trait>(
        &'life0 self,
        request: &'life1 ConversationRequest,
    ) -> Pin<Box<dyn Future<Output = Result<Completion, ProviderError>> + Send + 'async_trait>>
       where 'life0: 'async_trait,
             'life1: 'async_trait,
             Self: 'async_trait;
}

Expand description

A model-sampling wire: builds a request, calls the model, normalizes the response.

One impl per wire schema — not per gateway. A Provider identifies a request/response protocol shape (Anthropic Messages, OpenAI Chat, OpenAI Responses, or the mock double); a gateway/endpoint (OpenRouter, Bedrock, a proxy, a local model) is configuration (base_url + auth + headers) pointed at one of these schemas, never a separate impl. This is Grok Build’s split (ApiBackend schema vs an un-enumerated base_url/auth) and what ADR-0007’s per-model { base_url, api_backend, extra_headers } record encodes.

Required Methods§

Source

fn api_schema(&self) -> &str

The wire-schema id this provider speaks — e.g. "anthropic", "mock".

This is the protocol shape, not a gateway. Stamped into the report’s provider field.

Source

fn complete<'life0, 'life1, 'async_trait>( &'life0 self, request: &'life1 ConversationRequest, ) -> Pin<Box<dyn Future<Output = Result<Completion, ProviderError>> + Send + 'async_trait>>
where 'life0: 'async_trait, 'life1: 'async_trait, Self: 'async_trait,

Sample one completion for request.

§Errors

Returns ProviderError; use ProviderError::retryable to classify retry-vs-terminal. The transport-tier backoff lives in the wire; the bounded loop-level resample lives in the engine.