@juspay/neurolink/dist/proxy/routingEvidence.js

The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002

/** Schema-v1 routing evidence values shared by emitters and offline readers. */
export const PROXY_ACCOUNT_ROUTING_STRATEGIES = [
    "round-robin",
    "fill-first",
];
export const PROXY_ACCOUNT_ROUTING_MODES = [
    "quota",
    "primary",
    "round_robin",
    "single_account",
];
export const PROXY_ACCOUNT_ROUTING_REASONS = [
    "single_account",
    "round_robin",
    "configured_primary",
    "insertion_order",
    "availability",
    "cooldown_recovery",
    "quota_evidence",
    // Retained for backwards-compatible analysis of pre-fix logs. New routing
    // decisions must never select a production request for quota discovery.
    "quota_probe",
    "session_headroom",
    "scoped_headroom",
    "session_reset",
    "weekly_reset",
    "weekly_utilization",
    "scoped_utilization",
    "headroom",
    "session_affinity",
    "preferred_primary",
    "spill_inflight",
    "routing_policy_error",
];
export const PROXY_ACCOUNT_TYPES = ["oauth", "api_key"];
export const ACCOUNT_COOLING_REASONS = [
    "weekly",
    "session",
    "unified",
    "transient",
    "auth",
];
/**
 * Longest a cooldown may last for each reason, measured from when it was set.
 *
 * A reason names a specific provider window, so it also bounds the wait: a
 * "session" cooldown describes a 5-hour window and can never legitimately run
 * for days. Without a per-reason ceiling a single bogus reset timestamp parks an
 * account for as long as the global 8-day clamp allows — observed in the wild as
 * a 206-hour "session" cooldown. Values carry slack so a genuine window that
 * resets slightly late is not cut short.
 */
export const MAX_COOLDOWN_MS_BY_REASON = {
    session: 5 * 60 * 60 * 1000 + 15 * 60 * 1000,
    weekly: 7 * 24 * 60 * 60 * 1000 + 12 * 60 * 60 * 1000,
    // No named window bounds this one: it fires when the 5h/7d statuses still read
    // "allowed" and the provider's own Retry-After is the only signal. Generous on
    // purpose — truncating a provider-directed wait just re-hammers the account on
    // a shorter cycle — while still refusing an absurd multi-day park.
    unified: 12 * 60 * 60 * 1000,
    transient: 15 * 60 * 1000,
    auth: 5 * 60 * 1000,
};