@burdenoff/website-sdk 2026.1105.2 → 2026.1106.1

package/dist/avatar-mic/index.d.mts

@@ -1,62 +1,3 @@
-/** What the worklet posts for each frame. */
-interface CaptureFrameMessage {
-    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
-    pcm: ArrayBuffer;
-    /** Microphone level over the frame, dBFS. */
-    rmsDb: number;
-    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
-    refDb: number;
-    /** Zero crossings per sample. */
-    zcr: number;
-    /** AudioContext time at the end of the frame, seconds. */
-    ctxTime: number;
-    seq: number;
-}
-
-/**
- * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
- * (`capture-worklet.ts`), into 16 kHz PCM frames.
- *
- * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
- * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
- * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
- * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
- * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
- * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
- * parity tests.
- *
- * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
- * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
- * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
- *
- * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
- * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
- */
-
-interface MicCaptureHandle {
-    context: AudioContext;
-    micSource: MediaStreamAudioSourceNode;
-    track: MediaStreamTrack;
-    /** What the browser reports it applied. */
-    settings: MediaTrackSettings;
-    /** Stop everything and release the microphone. Safe to call more than once. */
-    stop(): void;
-}
-
-/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
-interface SocketLike {
-    readonly readyState: number;
-    send(data: string | ArrayBufferLike | ArrayBufferView): void;
-    close(code?: number): void;
-    listen(handlers: SocketHandlers): void;
-}
-interface SocketHandlers {
-    open(): void;
-    message(data: unknown): void;
-    close(code: number): void;
-    error(): void;
-}
-
 /**
  * BOFF-9797 (Stage C). The shapes the Avatar tab's mic shares with the live avatar's view.
  *
@@ -79,7 +20,8 @@ interface AvatarMicOffer {
  * Why a press did not give text:
  *
  * - `denied`: the microphone was refused (the visitor, the browser or the page's permissions);
- * - `unsupported`: this browser cannot capture here, or has no microphone to use;
+ * - `unsupported`: this browser cannot capture here, or has no microphone to use (with the browser's recognizer: its
+ *   speech service is off or unreachable here, or it has no language to offer);
  * - `unavailable`: the server will not take a stream from this page now: no offer, an offer it refuses twice, its mic
  *   notice is another version than this SDK's, or the website avatar is paused;
  * - `limit`: a limit on voice questions is reached (the visitor's, the network's, or every visitor's together);
@@ -88,6 +30,19 @@ interface AvatarMicOffer {
  *   time).
  */
 type AvatarMicFailure = "denied" | "unsupported" | "unavailable" | "limit" | "busy" | "failed";
+/**
+ * Why a press of the voice conversation ended with text, or with none (Stage C2, a press with an
+ * {@link AvatarMicEndpoint} only):
+ *
+ * - `pause`: the visitor stopped speaking for the pause, after a question: the one way a question is sent by itself;
+ * - `tap`: the visitor stopped it;
+ * - `limit`: the question reached the longest one, counted from its first word;
+ * - `silence`: nothing that counts as speech was heard in the longest wait, or before the recognizer ended;
+ * - `failed`: the recognizer failed, or could no longer go on, after or before words.
+ *
+ * Only `pause` may be sent; every other one's words go into the box for the visitor, never sent.
+ */
+type AvatarMicDoneBy = "tap" | "pause" | "limit" | "silence" | "failed";
 /**
  * One press of the mic, as the view shows it. `epoch` names the press: a new press has a new one, and what an older one
  * reports is never shown for a newer one.
@@ -98,6 +53,11 @@ type AvatarMicFailure = "denied" | "unsupported" | "unavailable" | "limit" | "bu
  * - `finishing`: the recording has stopped (a tap, or the limit: `limit`), and the last of the text is on its way;
  * - `done`: the text, final (empty when no words were heard);
  * - `failed`: no text, and why (with any final text heard before it failed, which the view does not use).
+ *
+ * A press with an {@link AvatarMicEndpoint} (the voice conversation, Stage C2) says more, and only it: `listening`
+ * says whether what is heard counts as speech (`genuine`) and whether the visitor has paused (`pausing`); `done` says
+ * why it ended (`by`) and whether its words are the avatar's own voice (`echo`). A press without one reports exactly
+ * the shapes it did before.
  */
 type AvatarMicState = {
     name: "idle";
@@ -109,6 +69,10 @@ type AvatarMicState = {
     epoch: number;
     startedAt: number;
     text: string;
+    /** With an endpoint: the words heard are speech, not silence, sounds or a stall (`isGenuineSpeech`). */
+    genuine?: boolean;
+    /** With an endpoint: no new word for a moment, after speech ("Still listening — take your time"). */
+    pausing?: boolean;
 } | {
     name: "finishing";
     epoch: number;
@@ -119,6 +83,10 @@ type AvatarMicState = {
     epoch: number;
     text: string;
     limit: boolean;
+    /** With an endpoint: why the press ended. */
+    by?: AvatarMicDoneBy;
+    /** With an endpoint: the words are the avatar's own answer, heard back through the microphone. */
+    echo?: boolean;
 } | {
     name: "failed";
     epoch: number;
@@ -151,9 +119,159 @@ interface AvatarMic {
     /** The press as it stands. */
     state(): AvatarMicState;
 }
+/**
+ * The mic through the browser's own speech recognition (Stage C, `mic.mode: "browser"` in availability): the browser
+ * turns the voice into text, and nothing of the press reaches the server. Same press, same states, same failures as
+ * the stream's ({@link AvatarMic}, {@link AvatarMicState}); it never reports `unavailable`, `limit` or `busy`.
+ */
+interface BrowserAvatarMicOptions {
+    /** The recognizer's language, a BCP-47 tag (`en-IN`). */
+    lang: string;
+    /** Tried once, in the same press, when the browser does not recognise `lang` (`en-US`). */
+    fallbackLang: string;
+    /**
+     * The longest press, from the moment the browser starts listening: thirty seconds. With an `endpoint`, the longest
+     * wait for a question to start, and the longest question, from its first word.
+     */
+    maxPressMs: number;
+    /** Every change of the press's state. */
+    onState: (state: AvatarMicState) => void;
+    /**
+     * The voice conversation (Stage C2): every press ends itself once the visitor has paused after a question (see
+     * {@link AvatarMicEndpoint}). Without it, a press is exactly the editable one: tap to start, tap to stop.
+     */
+    endpoint?: AvatarMicEndpoint;
+}
+/**
+ * How a press of the voice conversation ends itself (Stage C2): `done` with `by: "pause"` once what is heard is speech
+ * and no word of it has changed for `pauseMs` (twice that for a sentence left open), never sooner than
+ * `BROWSER_PAUSE_AFTER_START_MS` after its recognizer (re)started, and never between two recognizers.
+ */
+interface AvatarMicEndpoint {
+    /**
+     * The pause that ends a question, in milliseconds: the server's `mic.pauseMs` (2000 to 2500, read strictly by the
+     * view's API). Read within those bounds whatever is passed.
+     */
+    pauseMs: number;
+    /**
+     * The words of the answer the avatar has just read (its captions), for the echo check, read as the press ends; null
+     * when there is none, or when the check is not for this press (the view's hook checks only the presses the
+     * conversation starts by itself).
+     */
+    echoOf: () => string | null;
+}
 /** What the mic's chunk (`dist/avatar-mic/`) exports. */
 interface AvatarMicModule {
     createAvatarMic: (options: AvatarMicOptions) => AvatarMic;
+    createBrowserAvatarMic: (options: BrowserAvatarMicOptions) => AvatarMic;
+}
+
+/** One alternative of a recognition result: what the browser heard. */
+interface RecognitionAlternativeLike {
+    readonly transcript: string;
+}
+/** One recognition result: final, or still being recognised. */
+interface RecognitionResultLike {
+    readonly isFinal: boolean;
+    readonly length: number;
+    readonly [index: number]: RecognitionAlternativeLike | undefined;
+}
+/** The whole result list of a recognizer, as each `result` event carries it. */
+interface RecognitionResultListLike {
+    readonly length: number;
+    readonly [index: number]: RecognitionResultLike | undefined;
+}
+/** What this module uses of a browser's `SpeechRecognition` (declared here: TypeScript's DOM library has none). */
+interface RecognitionLike {
+    lang: string;
+    continuous: boolean;
+    interimResults: boolean;
+    maxAlternatives: number;
+    onstart: (() => void) | null;
+    onresult: ((event: {
+        readonly results: RecognitionResultListLike;
+    }) => void) | null;
+    onerror: ((event: {
+        readonly error: string;
+    }) => void) | null;
+    onend: (() => void) | null;
+    start(): void;
+    stop(): void;
+    abort(): void;
+}
+/** What a press needs from the browser. The browser's own by default; a test hands in fakes. */
+interface BrowserAvatarMicSeams {
+    /** A new recognizer, or null where the browser has none. */
+    createRecognition: () => RecognitionLike | null;
+    /** A monotonic clock, in milliseconds, for the cooldowns. */
+    now: () => number;
+    /** {@link primeBrowserMicrophone}: called inside the tap, where the microphone is not known to be granted. */
+    primeMicrophone: () => Promise<void> | null;
+    /** {@link watchBrowserMicPermission}: called once, as the mic is made. */
+    watchMicPermission: (onState: (state: PermissionState) => void) => void;
+    /** {@link hasCoarsePointer}: read at each tap. */
+    touch: () => boolean;
+}
+/** The Avatar tab's mic through the browser's own recognizer. See the file's comment for what a press does. */
+declare function createBrowserAvatarMic(options: BrowserAvatarMicOptions, seams?: Partial<BrowserAvatarMicSeams>): AvatarMic;
+
+/** What the worklet posts for each frame. */
+interface CaptureFrameMessage {
+    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
+    pcm: ArrayBuffer;
+    /** Microphone level over the frame, dBFS. */
+    rmsDb: number;
+    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
+    refDb: number;
+    /** Zero crossings per sample. */
+    zcr: number;
+    /** AudioContext time at the end of the frame, seconds. */
+    ctxTime: number;
+    seq: number;
+}
+
+/**
+ * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
+ * (`capture-worklet.ts`), into 16 kHz PCM frames.
+ *
+ * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
+ * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
+ * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
+ * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
+ * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
+ * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
+ * parity tests.
+ *
+ * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
+ * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
+ * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
+ *
+ * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
+ * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
+ */
+
+interface MicCaptureHandle {
+    context: AudioContext;
+    micSource: MediaStreamAudioSourceNode;
+    track: MediaStreamTrack;
+    /** What the browser reports it applied. */
+    settings: MediaTrackSettings;
+    /** Stop everything and release the microphone. Safe to call more than once. */
+    stop(): void;
+}
+
+/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
+interface SocketLike {
+    readonly readyState: number;
+    send(data: string | ArrayBufferLike | ArrayBufferView): void;
+    close(code?: number): void;
+    listen(handlers: SocketHandlers): void;
+}
+interface SocketHandlers {
+    open(): void;
+    message(data: unknown): void;
+    close(code: number): void;
+    error(): void;
 }
 
 /**
@@ -176,7 +294,8 @@ interface AvatarMicModule {
  *
  * **Released on every way out.** The microphone's track is stopped the moment the recording stops, and also when the
  * press fails, is cancelled (`cancel()`: the view going, the page hidden, the Chat tab) or superseded (a new press while
- * one is still finishing: `cancel` for the old one, so it stops being paid for, and its text is never reported). A
+ * one is still finishing, a tap the view lets through for this stream mic only: `cancel` for the old one, so it stops
+ * being paid for, and its text is never reported). A
  * microphone that opens after its press has gone is closed as it opens. The stream is closed whenever no press needs
  * it: one connection serves a press and the press that supersedes it, and no other.
  */
@@ -190,4 +309,4 @@ interface AvatarMicSeams {
 /** The mic of one Avatar tab. See the file's comment for what a press does. */
 declare function createAvatarMic(options: AvatarMicOptions, seams?: Partial<AvatarMicSeams>): AvatarMic;
 
-export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, createAvatarMic };
+export { type AvatarMic, type AvatarMicDoneBy, type AvatarMicEndpoint, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, type BrowserAvatarMicOptions, createAvatarMic, createBrowserAvatarMic };

package/dist/avatar-mic/index.d.ts

@@ -1,62 +1,3 @@
-/** What the worklet posts for each frame. */
-interface CaptureFrameMessage {
-    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
-    pcm: ArrayBuffer;
-    /** Microphone level over the frame, dBFS. */
-    rmsDb: number;
-    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
-    refDb: number;
-    /** Zero crossings per sample. */
-    zcr: number;
-    /** AudioContext time at the end of the frame, seconds. */
-    ctxTime: number;
-    seq: number;
-}
-
-/**
- * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
- * (`capture-worklet.ts`), into 16 kHz PCM frames.
- *
- * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
- * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
- * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
- * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
- * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
- * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
- * parity tests.
- *
- * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
- * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
- * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
- *
- * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
- * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
- */
-
-interface MicCaptureHandle {
-    context: AudioContext;
-    micSource: MediaStreamAudioSourceNode;
-    track: MediaStreamTrack;
-    /** What the browser reports it applied. */
-    settings: MediaTrackSettings;
-    /** Stop everything and release the microphone. Safe to call more than once. */
-    stop(): void;
-}
-
-/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
-interface SocketLike {
-    readonly readyState: number;
-    send(data: string | ArrayBufferLike | ArrayBufferView): void;
-    close(code?: number): void;
-    listen(handlers: SocketHandlers): void;
-}
-interface SocketHandlers {
-    open(): void;
-    message(data: unknown): void;
-    close(code: number): void;
-    error(): void;
-}
-
 /**
  * BOFF-9797 (Stage C). The shapes the Avatar tab's mic shares with the live avatar's view.
  *
@@ -79,7 +20,8 @@ interface AvatarMicOffer {
  * Why a press did not give text:
  *
  * - `denied`: the microphone was refused (the visitor, the browser or the page's permissions);
- * - `unsupported`: this browser cannot capture here, or has no microphone to use;
+ * - `unsupported`: this browser cannot capture here, or has no microphone to use (with the browser's recognizer: its
+ *   speech service is off or unreachable here, or it has no language to offer);
  * - `unavailable`: the server will not take a stream from this page now: no offer, an offer it refuses twice, its mic
  *   notice is another version than this SDK's, or the website avatar is paused;
  * - `limit`: a limit on voice questions is reached (the visitor's, the network's, or every visitor's together);
@@ -88,6 +30,19 @@ interface AvatarMicOffer {
  *   time).
  */
 type AvatarMicFailure = "denied" | "unsupported" | "unavailable" | "limit" | "busy" | "failed";
+/**
+ * Why a press of the voice conversation ended with text, or with none (Stage C2, a press with an
+ * {@link AvatarMicEndpoint} only):
+ *
+ * - `pause`: the visitor stopped speaking for the pause, after a question: the one way a question is sent by itself;
+ * - `tap`: the visitor stopped it;
+ * - `limit`: the question reached the longest one, counted from its first word;
+ * - `silence`: nothing that counts as speech was heard in the longest wait, or before the recognizer ended;
+ * - `failed`: the recognizer failed, or could no longer go on, after or before words.
+ *
+ * Only `pause` may be sent; every other one's words go into the box for the visitor, never sent.
+ */
+type AvatarMicDoneBy = "tap" | "pause" | "limit" | "silence" | "failed";
 /**
  * One press of the mic, as the view shows it. `epoch` names the press: a new press has a new one, and what an older one
  * reports is never shown for a newer one.
@@ -98,6 +53,11 @@ type AvatarMicFailure = "denied" | "unsupported" | "unavailable" | "limit" | "bu
  * - `finishing`: the recording has stopped (a tap, or the limit: `limit`), and the last of the text is on its way;
  * - `done`: the text, final (empty when no words were heard);
  * - `failed`: no text, and why (with any final text heard before it failed, which the view does not use).
+ *
+ * A press with an {@link AvatarMicEndpoint} (the voice conversation, Stage C2) says more, and only it: `listening`
+ * says whether what is heard counts as speech (`genuine`) and whether the visitor has paused (`pausing`); `done` says
+ * why it ended (`by`) and whether its words are the avatar's own voice (`echo`). A press without one reports exactly
+ * the shapes it did before.
  */
 type AvatarMicState = {
     name: "idle";
@@ -109,6 +69,10 @@ type AvatarMicState = {
     epoch: number;
     startedAt: number;
     text: string;
+    /** With an endpoint: the words heard are speech, not silence, sounds or a stall (`isGenuineSpeech`). */
+    genuine?: boolean;
+    /** With an endpoint: no new word for a moment, after speech ("Still listening — take your time"). */
+    pausing?: boolean;
 } | {
     name: "finishing";
     epoch: number;
@@ -119,6 +83,10 @@ type AvatarMicState = {
     epoch: number;
     text: string;
     limit: boolean;
+    /** With an endpoint: why the press ended. */
+    by?: AvatarMicDoneBy;
+    /** With an endpoint: the words are the avatar's own answer, heard back through the microphone. */
+    echo?: boolean;
 } | {
     name: "failed";
     epoch: number;
@@ -151,9 +119,159 @@ interface AvatarMic {
     /** The press as it stands. */
     state(): AvatarMicState;
 }
+/**
+ * The mic through the browser's own speech recognition (Stage C, `mic.mode: "browser"` in availability): the browser
+ * turns the voice into text, and nothing of the press reaches the server. Same press, same states, same failures as
+ * the stream's ({@link AvatarMic}, {@link AvatarMicState}); it never reports `unavailable`, `limit` or `busy`.
+ */
+interface BrowserAvatarMicOptions {
+    /** The recognizer's language, a BCP-47 tag (`en-IN`). */
+    lang: string;
+    /** Tried once, in the same press, when the browser does not recognise `lang` (`en-US`). */
+    fallbackLang: string;
+    /**
+     * The longest press, from the moment the browser starts listening: thirty seconds. With an `endpoint`, the longest
+     * wait for a question to start, and the longest question, from its first word.
+     */
+    maxPressMs: number;
+    /** Every change of the press's state. */
+    onState: (state: AvatarMicState) => void;
+    /**
+     * The voice conversation (Stage C2): every press ends itself once the visitor has paused after a question (see
+     * {@link AvatarMicEndpoint}). Without it, a press is exactly the editable one: tap to start, tap to stop.
+     */
+    endpoint?: AvatarMicEndpoint;
+}
+/**
+ * How a press of the voice conversation ends itself (Stage C2): `done` with `by: "pause"` once what is heard is speech
+ * and no word of it has changed for `pauseMs` (twice that for a sentence left open), never sooner than
+ * `BROWSER_PAUSE_AFTER_START_MS` after its recognizer (re)started, and never between two recognizers.
+ */
+interface AvatarMicEndpoint {
+    /**
+     * The pause that ends a question, in milliseconds: the server's `mic.pauseMs` (2000 to 2500, read strictly by the
+     * view's API). Read within those bounds whatever is passed.
+     */
+    pauseMs: number;
+    /**
+     * The words of the answer the avatar has just read (its captions), for the echo check, read as the press ends; null
+     * when there is none, or when the check is not for this press (the view's hook checks only the presses the
+     * conversation starts by itself).
+     */
+    echoOf: () => string | null;
+}
 /** What the mic's chunk (`dist/avatar-mic/`) exports. */
 interface AvatarMicModule {
     createAvatarMic: (options: AvatarMicOptions) => AvatarMic;
+    createBrowserAvatarMic: (options: BrowserAvatarMicOptions) => AvatarMic;
+}
+
+/** One alternative of a recognition result: what the browser heard. */
+interface RecognitionAlternativeLike {
+    readonly transcript: string;
+}
+/** One recognition result: final, or still being recognised. */
+interface RecognitionResultLike {
+    readonly isFinal: boolean;
+    readonly length: number;
+    readonly [index: number]: RecognitionAlternativeLike | undefined;
+}
+/** The whole result list of a recognizer, as each `result` event carries it. */
+interface RecognitionResultListLike {
+    readonly length: number;
+    readonly [index: number]: RecognitionResultLike | undefined;
+}
+/** What this module uses of a browser's `SpeechRecognition` (declared here: TypeScript's DOM library has none). */
+interface RecognitionLike {
+    lang: string;
+    continuous: boolean;
+    interimResults: boolean;
+    maxAlternatives: number;
+    onstart: (() => void) | null;
+    onresult: ((event: {
+        readonly results: RecognitionResultListLike;
+    }) => void) | null;
+    onerror: ((event: {
+        readonly error: string;
+    }) => void) | null;
+    onend: (() => void) | null;
+    start(): void;
+    stop(): void;
+    abort(): void;
+}
+/** What a press needs from the browser. The browser's own by default; a test hands in fakes. */
+interface BrowserAvatarMicSeams {
+    /** A new recognizer, or null where the browser has none. */
+    createRecognition: () => RecognitionLike | null;
+    /** A monotonic clock, in milliseconds, for the cooldowns. */
+    now: () => number;
+    /** {@link primeBrowserMicrophone}: called inside the tap, where the microphone is not known to be granted. */
+    primeMicrophone: () => Promise<void> | null;
+    /** {@link watchBrowserMicPermission}: called once, as the mic is made. */
+    watchMicPermission: (onState: (state: PermissionState) => void) => void;
+    /** {@link hasCoarsePointer}: read at each tap. */
+    touch: () => boolean;
+}
+/** The Avatar tab's mic through the browser's own recognizer. See the file's comment for what a press does. */
+declare function createBrowserAvatarMic(options: BrowserAvatarMicOptions, seams?: Partial<BrowserAvatarMicSeams>): AvatarMic;
+
+/** What the worklet posts for each frame. */
+interface CaptureFrameMessage {
+    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
+    pcm: ArrayBuffer;
+    /** Microphone level over the frame, dBFS. */
+    rmsDb: number;
+    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
+    refDb: number;
+    /** Zero crossings per sample. */
+    zcr: number;
+    /** AudioContext time at the end of the frame, seconds. */
+    ctxTime: number;
+    seq: number;
+}
+
+/**
+ * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
+ * (`capture-worklet.ts`), into 16 kHz PCM frames.
+ *
+ * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
+ * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
+ * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
+ * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
+ * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
+ * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
+ * parity tests.
+ *
+ * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
+ * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
+ * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
+ *
+ * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
+ * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
+ */
+
+interface MicCaptureHandle {
+    context: AudioContext;
+    micSource: MediaStreamAudioSourceNode;
+    track: MediaStreamTrack;
+    /** What the browser reports it applied. */
+    settings: MediaTrackSettings;
+    /** Stop everything and release the microphone. Safe to call more than once. */
+    stop(): void;
+}
+
+/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
+interface SocketLike {
+    readonly readyState: number;
+    send(data: string | ArrayBufferLike | ArrayBufferView): void;
+    close(code?: number): void;
+    listen(handlers: SocketHandlers): void;
+}
+interface SocketHandlers {
+    open(): void;
+    message(data: unknown): void;
+    close(code: number): void;
+    error(): void;
 }
 
 /**
@@ -176,7 +294,8 @@ interface AvatarMicModule {
  *
  * **Released on every way out.** The microphone's track is stopped the moment the recording stops, and also when the
  * press fails, is cancelled (`cancel()`: the view going, the page hidden, the Chat tab) or superseded (a new press while
- * one is still finishing: `cancel` for the old one, so it stops being paid for, and its text is never reported). A
+ * one is still finishing, a tap the view lets through for this stream mic only: `cancel` for the old one, so it stops
+ * being paid for, and its text is never reported). A
  * microphone that opens after its press has gone is closed as it opens. The stream is closed whenever no press needs
  * it: one connection serves a press and the press that supersedes it, and no other.
  */
@@ -190,4 +309,4 @@ interface AvatarMicSeams {
 /** The mic of one Avatar tab. See the file's comment for what a press does. */
 declare function createAvatarMic(options: AvatarMicOptions, seams?: Partial<AvatarMicSeams>): AvatarMic;
 
-export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, createAvatarMic };
+export { type AvatarMic, type AvatarMicDoneBy, type AvatarMicEndpoint, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, type BrowserAvatarMicOptions, createAvatarMic, createBrowserAvatarMic };