//! Model identifiers, prompt shape, sampling.
//!
//! Deliberately small: the daemon's model registry is canonical;
//! clients refer to models by name and don't carry weights over the
//! wire. The format is chosen by the operator at load time, not by
//! the client.

/// Reference to a model the daemon has (or can) load.
#[derive(Clone, Debug, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
pub struct ModelRef {
    /// Operator-chosen name. Stable across reloads of the same
    /// weights.
    pub name: String,
    /// Optional revision tag — lets a client pin a specific weights
    /// version if the operator has multiple. `None` means "whatever
    /// is current."
    pub revision: Option<String>,
}

/// Prompt input to an inference request.
///
/// Token IDs are not on the wire at this layer — clients send text
/// or chat-message structure; the daemon owns the tokenizer for the
/// referenced model. Sending pre-tokenized input would couple
/// clients to the daemon's tokenizer choice, which is exactly the
/// kind of coupling the v1 freeze is meant to prevent.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub enum PromptSpec {
    /// Single text completion.
    Text(String),
    /// Chat-style message list. The daemon applies the model's
    /// chat template.
    Chat(Vec<ChatMessage>),
}

/// One message in a chat-style prompt.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub struct ChatMessage {
    pub role: ChatRole,
    pub content: String,
}

/// Conversation role. Mirrors the OpenAI-style triad — the J1 shim
/// translates one-to-one without information loss.
#[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum ChatRole {
    System,
    User,
    Assistant,
}

/// Sampling parameters. Conservative defaults; the daemon may
/// further constrain values it can't honor for the selected model.
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
pub struct SamplingParams {
    pub max_tokens: u32,
    pub temperature: f32,
    pub top_p: f32,
    /// Stop strings. Generation halts on the first match.
    pub stop: Vec<String>,
    /// Seed for reproducible sampling. `None` means daemon-chosen.
    pub seed: Option<u64>,
}

impl Default for SamplingParams {
    fn default() -> Self {
        Self {
            max_tokens: 512,
            temperature: 0.7,
            top_p: 0.95,
            stop: Vec::new(),
            seed: None,
        }
    }
}

/// Scheduling priority class. The supervisor's fair-share policy
/// uses these to weight admission and preemption between
/// concurrent clients (Thread C, A2).
#[derive(
    Clone,
    Copy,
    Debug,
    Default,
    PartialEq,
    Eq,
    PartialOrd,
    Ord,
    serde::Serialize,
    serde::Deserialize,
)]
pub enum Priority {
    /// Background work; preemptible. Batch jobs, embeddings runs.
    Background,
    /// Default. Interactive but not latency-critical.
    #[default]
    Normal,
    /// Latency-critical interactive sessions. Preempts background
    /// work but not other interactive sessions.
    Interactive,
}