Provider-specific options using an extensible GADT. Each provider registers its own typed key without circular dependencies.

Reason why model generation finished.

Token usage statistics.

Warnings emitted by providers for unsupported features.

Errors from provider API calls.

Parse HTTP rate-limit hints into a retry delay in seconds.

Prompt messages with role-constrained content parts.

Tool definition for function calling.

Strategy for tool selection during generation.

Generation mode -- controls the shape of model output.

Response content parts returned by model generation.

Unified input options for both generate and stream calls.

Result of a non-streaming generation call.

Individual parts emitted during streaming generation.

Result of a streaming generation call.

Core abstraction for AI language models.

Factory for creating model instances from a provider.

Middleware for cross-cutting concerns (logging, caching, retries).

Server-Sent Events parser for streaming API responses.

HTTP timeout configuration.

HTTP client wrappers that apply Http_timeouts to cohttp calls.