//
//  ContextBudget.swift
//  Poche
//
//  Author:  Simon-Pierre Boucher
//  Contact: contact@spboucher.ai
//

import Foundation

/// Token accounting for one session (CLAUDE.md §5 — the central concern).
///
/// The window holds everything at once: instructions, tool schemas, the
/// whole transcript, and the response to come. At 70% we condense; at least
/// 30% stays reserved for the response — a prompt that "fits" with no room
/// to answer still fails.
struct ContextBudget: Sendable {
    /// iOS 26 window. 8192 on newer OS/devices; keep the pessimistic value
    /// until the SDK exposes the real one.
    var windowSize = 4096
    var responseReserveRatio = 0.30
    var condenseThresholdRatio = 0.70

    var responseReserve: Int {
        Int(Double(windowSize) * responseReserveRatio)
    }

    var condenseThreshold: Int {
        Int(Double(windowSize) * condenseThresholdRatio)
    }

    /// Largest prompt (fixed cost + transcript + new message) we allow.
    var sendableLimit: Int {
        windowSize - responseReserve
    }

    func needsCondensation(estimatedTokens: Int) -> Bool {
        estimatedTokens >= condenseThreshold
    }

    func canSend(estimatedTokens: Int) -> Bool {
        estimatedTokens < sendableLimit
    }

    func usageRatio(estimatedTokens: Int) -> Double {
        min(1, Double(estimatedTokens) / Double(windowSize))
    }
}