//! Client → daemon requests.

use core::time::Duration;

use crate::model::{ModelRef, Priority, PromptSpec, SamplingParams};

/// Opaque request identifier. Clients pick the value (typically a
/// monotonic counter or a UUID); the daemon echoes it on every
/// event for the request.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
pub struct RequestId(pub u64);

/// One client request envelope.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub struct ClientRequest {
    pub id: RequestId,
    pub kind: RequestKind,
}

/// The discriminated union of every operation the native API
/// supports. New variants are additive in semver-minor; removing
/// or changing a variant is semver-major.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub enum RequestKind {
    /// Run an inference. Token stream returned as a sequence of
    /// `ServerEvent::Tokens` events.
    Inference(InferenceRequest),
    /// Snapshot of fleet state. Returns one
    /// `ServerEvent::FleetDescription` event.
    DescribeFleet,
    /// Ask the daemon to load a model into memory. Returns an
    /// `Accepted` event followed by progress or an error.
    LoadModel(ModelRef),
    /// Ask the daemon to evict a model. Returns `Accepted` and a
    /// terminal `Tokens { finished: true }` shaped event (empty
    /// chunk) when complete; this lets clients treat completion as
    /// the same primitive across all request kinds.
    UnloadModel(ModelRef),
    /// Cancel a previously-submitted request. The request being
    /// cancelled receives a terminal event with `finished: true`
    /// and a cancellation `ErrorKind`.
    Cancel(RequestId),
}

/// Body of an `Inference` request.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub struct InferenceRequest {
    pub model: ModelRef,
    pub prompt: PromptSpec,
    pub sampling: SamplingParams,
    pub stream: bool,
    pub priority: Priority,
    /// Soft wall-clock deadline. The supervisor will not start
    /// work it believes cannot finish inside the deadline; it does
    /// not interrupt work already started. `None` means no deadline.
    pub deadline: Option<Duration>,
}