/**
 * Pre-gate addressing eligibility for mic triggers.
 *
 * Root cause of the false-delegation storm: the proactive gate saw only raw
 * transcript text — no speaker identity, no room context, no notion of who the
 * speech was *addressed to*. A room full of people, a video, and game audio
 * produced ~1500 mic segments the gate then rationalized as "the user asking
 * the koi", even though the wearer was never talking to it.
 *
 * This layer enforces the core principle BEFORE the expensive gate model ever
 * runs: delegate ONLY when the WEARER is ADDRESSING the koi. It is pure and
 * deterministic (all state lives in an injectable tracker), so it is unit
 * testable against the real mic-window fixture.
 *
 * Three signals, cheapest/strongest first:
 *   1. Ambient room detection — distinct Soniox speaker tags in a rolling
 *      window. A room/video (>=3 voices) means overheard speech is NOT for the
 *      koi; only an explicit address escapes.
 *   2. Wake-word address — "koi" as a VOCATIVE (utterance-initial, standalone,
 *      or comma-delimited), not a bare substring. This is the strongest signal
 *      and needs no voice-ID.
 *   3. Wearer voice-ID (best-effort, never required) — a CONFIDENT non-wearer
 *      can never invoke the owner's koi.
 *
 * Whether nameless speech may flow at all is the ENGAGEMENT mode's call
 * (infra/voice-mode): wake = name only · conversation = name opens a sliding
 * window · ambient = quiet wearer speech always reaches the relevance gate.
 */
/**
 * Rolling window for the distinct-speaker count. Long enough that a burst of
 * cross-talk still reads as "a room" a few seconds later; short enough that
 * handing the device back to a solo wearer re-baselines to quiet within a
 * minute. 45s comfortably spans a handful of conversational turns.
 */
export declare const AMBIENT_WINDOW_MS = 45000;
/**
 * Distinct Soniox speaker tags at/above which the environment is ambient (a
 * room, a video, several people talking). 3 voices in 45s is not a 1:1 chat
 * with the device — it is a scene, and overheard speech is not for the koi.
 */
export declare const AMBIENT_SPEAKER_THRESHOLD = 3;
/**
 * Sub-window + count for the "sustained continuous mixed speech" heuristic.
 * Many back-to-back segments from >=2 speakers in a short span reads as live
 * crosstalk / media dialogue even before a third distinct tag appears.
 */
export declare const SUSTAINED_WINDOW_MS = 20000;
export declare const SUSTAINED_SEGMENT_THRESHOLD = 8;
/**
 * At/above this distinct-speaker count voice-ID must NOT auto-enroll the wearer
 * (that is what poisoned the "You" profile during the 6-speaker room). Shared
 * with identify.ts so enrollment and addressing agree on "quiet".
 */
export declare const LOW_AMBIENT_MAX_SPEAKERS = 2;
/**
 * Cosine at/above which a resolved non-wearer identity counts as a CONFIDENT
 * other (mirrors voiceid's VOICEID_THRESHOLD). A confident other must never be
 * able to invoke the owner's koi.
 */
export declare const WEARER_MATCH_CONFIDENCE = 0.72;
export interface AmbientSnapshot {
    /** Distinct Soniox speaker tags in the last AMBIENT_WINDOW_MS. */
    distinctSpeakers: number;
    /** distinctSpeakers >= AMBIENT_SPEAKER_THRESHOLD. */
    ambient: boolean;
    /** Back-to-back mixed-speaker speech (see SUSTAINED_* constants). */
    sustained: boolean;
}
/**
 * Rolling count of DISTINCT Soniox `speaker` tags over recent mic segments.
 * Fed by whichever subsystem tails the audio JSONL (triggers.ts); read by both
 * addressing (classify) and voice-ID enrollment.
 */
export declare class AmbientSpeakerTracker {
    private entries;
    /** Monotonic view of "latest known time" so out-of-order arrivals + explicit
     * (test/replay) timestamps prune deterministically without touching the wall
     * clock — the tracker's window is defined purely by the timestamps it sees. */
    private clock;
    note(speaker: string, ts?: number): void;
    snapshot(now?: number): AmbientSnapshot;
    private prune;
}
/** Module singleton: one gateway process, one shared ambient view. */
export declare const ambientTracker: AmbientSpeakerTracker;
/** Feed the shared tracker with a non-echo mic voice. */
export declare function noteMicSpeaker(speaker: string | null | undefined, ts: number): void;
export declare function getAmbientSnapshot(now?: number): AmbientSnapshot;
/** Low-ambient = safe to auto-enroll the wearer (<=2 distinct recent voices). */
export declare function isLowAmbient(now?: number): boolean;
/**
 * True when `text` addresses the koi with the wake word as a VOCATIVE — the
 * strongest address signal. Position/word-boundary based, NOT a substring
 * search, so "uninstall koi and reinstall" or "talking to the koi" do NOT
 * count (the wake word is neither utterance-initial nor comma-set-off there).
 */
export declare function isWakeAddress(text: string): boolean;
export declare function containsKoiMention(text: string): boolean;
/**
 * Cheap "reads like a request/command" gate for quiet mode. Filters filler and
 * fragments; a real question mark or an imperative/question lead-in qualifies.
 * NOTE: this alone is NOT sufficient to delegate — quiet mode additionally
 * requires a positive wearer voice-ID signal (see classifyMicAddressing).
 */
export declare function looksLikeDirectRequest(text: string): boolean;
/** Minimal structural shape of a resolved voice identity (see voiceid). */
export interface ResolvedSpeakerIdentity {
    wearer: boolean;
    anon: boolean;
    score: number;
}
export interface MicSegment {
    text: string;
    speaker: string | null;
    ts: number;
    audioPath?: string;
}
export interface AddressingContext {
    distinctSpeakers: number;
    /** distinctSpeakers >= AMBIENT_SPEAKER_THRESHOLD. */
    ambient: boolean;
    /** Sustained continuous mixed-speaker speech (treated as ambient). */
    sustained?: boolean;
    /** How the koi listens (infra/voice-mode): wake | conversation | ambient. */
    engagement?: "wake" | "conversation" | "ambient";
    /** conversation mode: a recent exchange keeps the window open — speech
     *  passes without the wake word until the conversation goes quiet. */
    conversationLive?: boolean;
    /** Koi's own TTS is playing over the speakers right now — the mic is
     *  hearing the koi. A mid-sentence "koi" in such a segment is almost
     *  certainly the koi saying its own name; only a true VOCATIVE ("koi, …")
     *  may wake, so a person can still barge with "koi, stop". */
    koiSpeaking?: boolean;
    /**
     * Best-effort voice-ID resolver, keyed by clip path. Returns null when the
     * clip is still in flight / pruned / voice-ID unavailable — which is the
     * common case at classify time (resolution lands ~1-2s after the clip).
     */
    resolveIdentity?: (audioPath: string | undefined) => ResolvedSpeakerIdentity | null;
}
export type AddressingMode = "wake" | "conversation" | "ambient-flow" | "vocative-rescue" | "blocked";
export interface AddressingResult {
    eligible: boolean;
    mode: AddressingMode;
    reason: string;
}
/**
 * Decide whether a batch of mic segments is eligible to reach the proactive
 * pipeline, and how. See the module doc for the principle.
 *
 * - ambient environment (>=3 speakers or sustained crosstalk): eligible ONLY
 *   on a wake-word address by the wearer (or, when voice-ID can't confirm
 *   anyone, a wake-word address at all). NO inference is allowed in a room.
 * - quiet (<=2 speakers): a wake-word address always passes as 'wake'.
 *   Nameless speech flows only per the ENGAGEMENT mode: an open conversation
 *   window ('conversation') or ambient listening ('ambient-flow'). In wake
 *   engagement — and conversation engagement with the window closed — the
 *   name is the only door.
 * - a CONFIDENT non-wearer is never eligible.
 */
export declare function classifyMicAddressing(batch: MicSegment[], ctx: AddressingContext): AddressingResult;