Builds model requests and streams responses through provider adapters. ModelExecutor manages requests, Conversation builds conversation input, and reasoning handles reasoning configuration. Provider credentials arrive through AuthResolver; the caller supplies authorized network and accounting context.
Chat Completions compatible and Anthropic Messages use the pure Rust protocol core. The native adapter retains shared input reservations through physical cleanup and output reservations through consumption. Its idle deadline covers network waits; waiting for local capacity is backpressure.
Canonical response items and text remain complete across HTTP, SSE, WebSocket and durable reconstruction. Chunking and bounded channels apply backpressure without a separate whole-response quota. Disposable continuation caches may be evicted; canonical history remains the recovery source. Request token selection, provider protocol validation and owned idle/cancellation cleanup stay separate from content retention.