@runtypelabs/persona 4.24.0 → 4.26.0

package/dist/theme-editor-preview.d.cts

@@ -6720,6 +6720,22 @@ type AgentWidgetStatusIndicatorConfig = {
     /** Status text while a reconnect attempt is in flight. */
     resumingText?: string;
 };
+/** Built-in `voiceRecognition.prewarm` triggers. See {@link AgentWidgetVoiceRecognitionConfig.prewarm}. */
+type AgentWidgetVoicePrewarmTrigger = 'hover';
+/**
+ * Custom `voiceRecognition.prewarm` trigger. Runs while the composer has a mic
+ * button, and again when the button is rebuilt or `update()` swaps in a
+ * different function. Call `warm()` whenever the visitor shows intent to talk
+ * (it is safe to call repeatedly). Return a cleanup to run on replacement.
+ */
+type AgentWidgetVoicePrewarmHook = (context: {
+    /** Fire-and-forget voice warm-up; latched and error-swallowing. */
+    warm: () => void;
+    /** The composer's current mic button. */
+    micButton: HTMLButtonElement;
+    /** The widget's mount element. */
+    mount: HTMLElement;
+}) => void | (() => void);
 type AgentWidgetVoiceRecognitionConfig = {
     /**
      * Enable the mic button and voice input. The mic still only renders when the
@@ -6785,6 +6801,29 @@ type AgentWidgetVoiceRecognitionConfig = {
      * Fires once per turn after the reply completes.
      */
     onMetrics?: (metrics: VoiceMetrics) => void;
+    /**
+     * Warm the voice path when the visitor shows intent to talk, before the mic
+     * click, so the call starts faster.
+     *
+     * - `'hover'`: when the pointer enters the mic button, it receives focus, or
+     *   (on touch) the first pointer-down on it.
+     * - `false`: never.
+     * - A function: custom trigger. Called with `{ warm, micButton, mount }`
+     *   while the mic button exists (and again when the button is rebuilt or
+     *   `update()` swaps in a different function); call `warm()` whenever you
+     *   decide, e.g. when the panel opens. May return a cleanup that runs on
+     *   destroy, rebuild, or replacement. Replaces the built-in trigger.
+     *
+     * What `warm()` does depends on the provider: `runtype` fires its prewarm
+     * (see `provider.runtype.prewarmMode`), a `custom` provider's optional
+     * {@link VoiceProvider.prewarm} runs, and the browser (Web Speech) path
+     * starts the client-token session init (as `sessionInit` does), since speech
+     * recognition itself cannot start before the click. Warming is
+     * fire-and-forget, latched, and never surfaces an error or UI state.
+     *
+     * @default 'hover' in client token mode, otherwise `false`
+     */
+    prewarm?: AgentWidgetVoicePrewarmTrigger | AgentWidgetVoicePrewarmHook | false;
     provider?: {
         type: 'browser' | 'runtype' | 'custom';
         browser?: {
@@ -6806,6 +6845,62 @@ type AgentWidgetVoiceRecognitionConfig = {
              * `@runtypelabs/persona/voice-worklet-player` for a jitter-buffered engine.
              */
             createPlaybackEngine?: () => VoicePlaybackEngine | Promise<VoicePlaybackEngine>;
+            /**
+             * How `prewarm()` warms the voice path (see `voiceRecognition.prewarm`).
+             *
+             * - `'request'` (default): a fire-and-forget
+             *   `POST {host}/v1/client/agents/{agentId}/voice/prewarm`.
+             * - `'attach'`: open the voice WebSocket ahead of the click without
+             *   starting a call; the click then starts the call on that socket. It can
+             *   create server work before the visitor clicks, so it is opt-in.
+             * @default 'request'
+             */
+            prewarmMode?: 'request' | 'attach';
+            /**
+             * `'attach'` mode: how long (ms) an attached socket waits for the click
+             * before it closes. The server clamps it to 30000..600000.
+             * @default 30000
+             */
+            attachIdleMs?: number;
+            /**
+             * Full-duplex (speech-to-speech, e.g. GPT-Live) calls: run the turns the
+             * voice model hands off to the agent through this widget's normal chat
+             * pipeline. The spoken request becomes the user message, and the answer
+             * renders as a regular assistant message (streaming, tool and approval
+             * UI, WebMCP page tools, persistence) that the voice model then reads
+             * aloud. Only takes effect when the server supports it (it confirms in
+             * `session_config`); otherwise the agent turn runs on the server and its
+             * spoken reply is transcribed, as before. Set `false` to always keep the
+             * agent turn on the server.
+             * @default true
+             */
+            clientDelegation?: boolean;
+            /**
+             * Full-duplex calls: extra context for the voice model (for example the
+             * page the visitor is on), sent after the chat history as of call start.
+             * The frame goes out once, when the visitor finishes their first
+             * utterance (or at the first agent hand-off). A function is
+             * called at call start and may be async. This text is capped at 4000 characters and the whole frame
+             * (history included) at 8000. Sent only to servers that accept call
+             * context (they announce it in `session_config`).
+             */
+            callContext?: string | (() => string | Promise<string>);
+            /**
+             * Full-duplex calls with client delegation: how long a tool approval
+             * that a spoken request raised may wait before it is declined as
+             * expired (and the voice model says so). Server-side approval gates
+             * keep their own timeout too. Hanging up never declines it: the
+             * approval card in the chat stays usable. Capped at 540000 (9 minutes),
+             * under the server's 10-minute delegation deadline.
+             * @default 300000
+             */
+            approvalTimeoutMs?: number;
+            /**
+             * Speech-to-speech calls: the one-line notice shown in the composer
+             * status area when the call starts. `false` hides it.
+             * @default "You're talking to an AI assistant. Voice is processed by OpenAI."
+             */
+            disclosureText?: string | false;
             /** @deprecated No-op on the realtime path: the server's STT owns turn-taking. */
             pauseDuration?: number;
             /** @deprecated No-op on the realtime path: the server's STT owns turn-taking. */
@@ -7044,6 +7139,13 @@ interface VoicePlaybackEngine {
     onFinished(callback: () => void): void;
     /** Release all audio resources. */
     destroy(): Promise<void> | void;
+    /**
+     * Optional. Speech-to-speech calls stream reply audio with no end-of-stream
+     * marker, so an engine that holds audio below a prebuffer waterline must, while
+     * continuous, release that held tail on its own (e.g. after a short input gap)
+     * instead of waiting for {@link markStreamEnd}. Engines without a prebuffer can omit it.
+     */
+    setContinuousMode?(enabled: boolean): void;
 }
 /**
  * A {@link VoicePlaybackEngine} that also supports pause/resume. Returned by
@@ -7087,14 +7189,23 @@ interface VoiceProvider {
      * Register a callback for incremental transcript updates during a voice turn.
      * `isFinal=false` is a live interim update (user partials, or assistant deltas
      * on providers that stream them); `isFinal=true` finalizes that role's text.
-     * On the realtime `runtype` path, interim updates fire for the `user` only and
-     * the `assistant` arrives as a single final. Providers with overlapping turns
-     * must include the same turnId on user and assistant finals; without IDs,
-     * providers must discard cancelled output before emitting the next user final.
-     */
-    onTranscript?(callback: (role: 'user' | 'assistant', text: string, isFinal: boolean, metadata?: {
-        turnId?: string;
-    }) => void): void;
+     * On the classic realtime `runtype` path, interim updates fire for the `user`
+     * only and the `assistant` arrives as a single final; without a turnId the
+     * session assumes strictly alternating turns, and providers must discard
+     * cancelled output before emitting the next user final.
+     *
+     * When `metadata.turnId` is present (full-duplex / speech-to-speech, e.g.
+     * GPT-Live), the session reconciles by `(turnId, role)`: each pair owns one
+     * bubble that later updates replace in place, turns may overlap or arrive in
+     * any order, and no empty assistant placeholder is injected. User and
+     * assistant may share a turnId (a late user bubble then renders above its
+     * reply) or use distinct ones. `metadata.startMs` (call-relative ms when the
+     * utterance began) orders a new bubble among the call's other keyed bubbles;
+     * without it bubbles order by first arrival. After a stop, keyed providers
+     * must drop the cancelled reply's output until the server acknowledges the
+     * cancel: the session renders every reply it receives.
+     */
+    onTranscript?(callback: (role: 'user' | 'assistant', text: string, isFinal: boolean, metadata?: VoiceTranscriptMetadata) => void): void;
     /** Register a callback for per-turn latency metrics (realtime path). */
     onMetrics?(callback: (metrics: VoiceMetrics) => void): void;
     /**
@@ -7115,6 +7226,116 @@ interface VoiceProvider {
     deactivateBargeIn?(): Promise<void>;
     /** Stop playback / cancel in-flight request without starting recording */
     stopPlayback?(): void;
+    /**
+     * Warm the provider ahead of `startListening()` (the widget calls it on mic
+     * intent, see `voiceRecognition.prewarm`). Must be fire-and-forget: never
+     * throw, emit errors or status, or acquire the microphone. May be called
+     * repeatedly; throttle as needed.
+     */
+    prewarm?(): void;
+    /**
+     * Called by the session once after construction with the chat-side surface
+     * a full-duplex provider uses to hand agent turns to the chat pipeline
+     * ({@link VoiceSessionBridge}). Providers without client delegation omit it.
+     */
+    setSessionBridge?(bridge: VoiceSessionBridge): void;
+    /** The AI-disclosure notice for the live call, if one should show (speech-to-speech). */
+    getDisclosure?(): string | null;
+}
+/** Metadata on a {@link VoiceProvider.onTranscript} update. */
+type VoiceTranscriptMetadata = {
+    /** Groups updates that belong to one utterance bubble (full duplex). */
+    turnId?: string;
+    /** Call-relative ms when the utterance began; orders late bubbles. */
+    startMs?: number;
+    /** Call-relative ms of the latest audio this update covers. */
+    endMs?: number;
+    /**
+     * The bubble only captions speech (see `AgentWidgetMessage.voiceCaption`).
+     * The chat pipeline owns the conversation, as with client delegation.
+     */
+    caption?: boolean;
+};
+/** An agent turn a full-duplex voice model handed to the widget. */
+type VoiceDelegationRequest = {
+    /** The voice model's delegation id (`delegation_started.delegationId`). */
+    delegationId: string;
+    /** What the visitor asked, as the voice model heard it. */
+    userText: string;
+    /**
+     * The user transcript utterances `userText` joins, oldest first (the last
+     * is the one it came from). May be empty.
+     */
+    userUtteranceIds: string[];
+    /** The voice model's view of the conversation it delegated from. */
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+};
+/**
+ * How a delegated turn ended, or that it waits on an approval (contract
+ * Amendment 5). `pending_approval` is the only non-terminal status. Open:
+ * a receiver treats an unknown terminal status as `failed`.
+ * - `completed`: the work ran.
+ * - `denied`: the visitor said no (Deny tap, or a spoken decline).
+ * - `timeout`: the approval lapsed (`approvalTimeoutMs`, or the server's own timeout).
+ * - `cancelled`: replaced by a newer request, or abandoned by the system.
+ * - `failed`: an error.
+ */
+type VoiceDelegationStatus = 'completed' | 'pending_approval' | 'denied' | 'timeout' | 'cancelled' | 'failed' | (string & {});
+/** The terminal result of a turn that parked on approvals. */
+type VoiceDelegationFollowUp = {
+    status: VoiceDelegationStatus;
+    /** What to read aloud (Markdown allowed). May be empty: nothing more to say. */
+    text: string;
+};
+/** The delegated turn's answer, sent back for the voice model to read aloud. */
+type VoiceDelegationResult = {
+    /** `pending_approval` when `followUp` is set; otherwise terminal. */
+    status: VoiceDelegationStatus;
+    /** The assistant text (Markdown allowed). */
+    text: string;
+    /**
+     * Set when the turn parked on approvals (`text` then asks for the decision).
+     * Resolves with the terminal result once the visitor decides (or the
+     * approval is replaced, declined by voice, or unanswered for
+     * `approvalTimeoutMs`) and the resumed turn finishes; `null` only when
+     * `signal` aborts first (hang-up, or the server cancelled the delegation).
+     */
+    followUp?: (options: {
+        signal: AbortSignal;
+        approvalTimeoutMs?: number;
+        /** `false`: run only the approval bookkeeping (expiry, supersede); nothing will be read back. @default true */
+        readBack?: boolean;
+        /** The resumed turn stopped on another approval: ask for it too (`text` is its approval script). */
+        onUpdate?: (text: string) => void;
+    }) => Promise<VoiceDelegationFollowUp | null>;
+};
+/**
+ * Chat-side surface the session hands a full-duplex provider: the visible
+ * history (for the call-start context) and a way to run a delegated turn
+ * through the normal chat pipeline.
+ */
+interface VoiceSessionBridge {
+    /** Visible user/assistant messages, oldest first. */
+    getHistory(): Array<{
+        role: 'user' | 'assistant';
+        content: string;
+    }>;
+    /**
+     * Submit the turn as the visitor's message (reusing its transcript bubble)
+     * and resolve with the assistant's final answer once the turn settles.
+     * Waits for any chat turn already in flight first. Never rejects.
+     */
+    runDelegatedTurn(request: VoiceDelegationRequest): Promise<VoiceDelegationResult>;
+    /**
+     * The server dropped this delegation's result (cancelled it, or refused a
+     * late result): its chat answer is no longer spoken, so browser TTS may read it.
+     * `expired`: the server's deadline passed, so its pending approvals are
+     * declined too, as the approval TTL would.
+     */
+    dropDelegation?(delegationId: string, expired?: boolean): void;
 }
 /**
  * Extra context for an approval decision, surfaced to `onDecision` and to the
@@ -9722,6 +9943,11 @@ type AgentWidgetToolCall = {
     startedAt?: number;
     completedAt?: number;
     durationMs?: number;
+    /**
+     * The call paused for an approval that was granted, and the agent re-ran it
+     * as a new tool call: that later call carries the result, this one has none.
+     */
+    superseded?: boolean;
 };
 /**
  * Represents a tool approval request in the chat conversation.
@@ -9743,6 +9969,8 @@ type AgentWidgetApproval = {
      */
     reason?: string;
     parameters?: unknown;
+    /** The paused tool call this approval gates (`approval_start.toolCallId`). */
+    toolCallId?: string;
     resolvedAt?: number;
 };
 type AgentWidgetMessageVariant = "assistant" | "reasoning" | "tool" | "approval";
@@ -9842,6 +10070,13 @@ type AgentWidgetMessage = {
      * }
      */
     voiceProcessing?: boolean;
+    /**
+     * Display-only caption of full-duplex voice speech, such as the voice
+     * model's own words, or something the visitor said that wasn't submitted
+     * as a chat turn. It renders in the transcript but is never sent to the
+     * model as conversation.
+     */
+    voiceCaption?: boolean;
     /**
      * Raw structured payload for this message (e.g., JSON action response).
      * Populated automatically when structured parsers run.
@@ -9957,6 +10192,8 @@ type InjectMessageOptions = {
      * Consumers can detect this in `messageTransform` to render custom UI.
      */
     voiceProcessing?: boolean;
+    /** See {@link AgentWidgetMessage.voiceCaption}. */
+    voiceCaption?: boolean;
     /**
      * Raw structured payload (typically a JSON string) representing the
      * full directive that produced this message: e.g. `{ "text": "...",

package/dist/theme-editor-preview.d.ts

@@ -6720,6 +6720,22 @@ type AgentWidgetStatusIndicatorConfig = {
     /** Status text while a reconnect attempt is in flight. */
     resumingText?: string;
 };
+/** Built-in `voiceRecognition.prewarm` triggers. See {@link AgentWidgetVoiceRecognitionConfig.prewarm}. */
+type AgentWidgetVoicePrewarmTrigger = 'hover';
+/**
+ * Custom `voiceRecognition.prewarm` trigger. Runs while the composer has a mic
+ * button, and again when the button is rebuilt or `update()` swaps in a
+ * different function. Call `warm()` whenever the visitor shows intent to talk
+ * (it is safe to call repeatedly). Return a cleanup to run on replacement.
+ */
+type AgentWidgetVoicePrewarmHook = (context: {
+    /** Fire-and-forget voice warm-up; latched and error-swallowing. */
+    warm: () => void;
+    /** The composer's current mic button. */
+    micButton: HTMLButtonElement;
+    /** The widget's mount element. */
+    mount: HTMLElement;
+}) => void | (() => void);
 type AgentWidgetVoiceRecognitionConfig = {
     /**
      * Enable the mic button and voice input. The mic still only renders when the
@@ -6785,6 +6801,29 @@ type AgentWidgetVoiceRecognitionConfig = {
      * Fires once per turn after the reply completes.
      */
     onMetrics?: (metrics: VoiceMetrics) => void;
+    /**
+     * Warm the voice path when the visitor shows intent to talk, before the mic
+     * click, so the call starts faster.
+     *
+     * - `'hover'`: when the pointer enters the mic button, it receives focus, or
+     *   (on touch) the first pointer-down on it.
+     * - `false`: never.
+     * - A function: custom trigger. Called with `{ warm, micButton, mount }`
+     *   while the mic button exists (and again when the button is rebuilt or
+     *   `update()` swaps in a different function); call `warm()` whenever you
+     *   decide, e.g. when the panel opens. May return a cleanup that runs on
+     *   destroy, rebuild, or replacement. Replaces the built-in trigger.
+     *
+     * What `warm()` does depends on the provider: `runtype` fires its prewarm
+     * (see `provider.runtype.prewarmMode`), a `custom` provider's optional
+     * {@link VoiceProvider.prewarm} runs, and the browser (Web Speech) path
+     * starts the client-token session init (as `sessionInit` does), since speech
+     * recognition itself cannot start before the click. Warming is
+     * fire-and-forget, latched, and never surfaces an error or UI state.
+     *
+     * @default 'hover' in client token mode, otherwise `false`
+     */
+    prewarm?: AgentWidgetVoicePrewarmTrigger | AgentWidgetVoicePrewarmHook | false;
     provider?: {
         type: 'browser' | 'runtype' | 'custom';
         browser?: {
@@ -6806,6 +6845,62 @@ type AgentWidgetVoiceRecognitionConfig = {
              * `@runtypelabs/persona/voice-worklet-player` for a jitter-buffered engine.
              */
             createPlaybackEngine?: () => VoicePlaybackEngine | Promise<VoicePlaybackEngine>;
+            /**
+             * How `prewarm()` warms the voice path (see `voiceRecognition.prewarm`).
+             *
+             * - `'request'` (default): a fire-and-forget
+             *   `POST {host}/v1/client/agents/{agentId}/voice/prewarm`.
+             * - `'attach'`: open the voice WebSocket ahead of the click without
+             *   starting a call; the click then starts the call on that socket. It can
+             *   create server work before the visitor clicks, so it is opt-in.
+             * @default 'request'
+             */
+            prewarmMode?: 'request' | 'attach';
+            /**
+             * `'attach'` mode: how long (ms) an attached socket waits for the click
+             * before it closes. The server clamps it to 30000..600000.
+             * @default 30000
+             */
+            attachIdleMs?: number;
+            /**
+             * Full-duplex (speech-to-speech, e.g. GPT-Live) calls: run the turns the
+             * voice model hands off to the agent through this widget's normal chat
+             * pipeline. The spoken request becomes the user message, and the answer
+             * renders as a regular assistant message (streaming, tool and approval
+             * UI, WebMCP page tools, persistence) that the voice model then reads
+             * aloud. Only takes effect when the server supports it (it confirms in
+             * `session_config`); otherwise the agent turn runs on the server and its
+             * spoken reply is transcribed, as before. Set `false` to always keep the
+             * agent turn on the server.
+             * @default true
+             */
+            clientDelegation?: boolean;
+            /**
+             * Full-duplex calls: extra context for the voice model (for example the
+             * page the visitor is on), sent after the chat history as of call start.
+             * The frame goes out once, when the visitor finishes their first
+             * utterance (or at the first agent hand-off). A function is
+             * called at call start and may be async. This text is capped at 4000 characters and the whole frame
+             * (history included) at 8000. Sent only to servers that accept call
+             * context (they announce it in `session_config`).
+             */
+            callContext?: string | (() => string | Promise<string>);
+            /**
+             * Full-duplex calls with client delegation: how long a tool approval
+             * that a spoken request raised may wait before it is declined as
+             * expired (and the voice model says so). Server-side approval gates
+             * keep their own timeout too. Hanging up never declines it: the
+             * approval card in the chat stays usable. Capped at 540000 (9 minutes),
+             * under the server's 10-minute delegation deadline.
+             * @default 300000
+             */
+            approvalTimeoutMs?: number;
+            /**
+             * Speech-to-speech calls: the one-line notice shown in the composer
+             * status area when the call starts. `false` hides it.
+             * @default "You're talking to an AI assistant. Voice is processed by OpenAI."
+             */
+            disclosureText?: string | false;
             /** @deprecated No-op on the realtime path: the server's STT owns turn-taking. */
             pauseDuration?: number;
             /** @deprecated No-op on the realtime path: the server's STT owns turn-taking. */
@@ -7044,6 +7139,13 @@ interface VoicePlaybackEngine {
     onFinished(callback: () => void): void;
     /** Release all audio resources. */
     destroy(): Promise<void> | void;
+    /**
+     * Optional. Speech-to-speech calls stream reply audio with no end-of-stream
+     * marker, so an engine that holds audio below a prebuffer waterline must, while
+     * continuous, release that held tail on its own (e.g. after a short input gap)
+     * instead of waiting for {@link markStreamEnd}. Engines without a prebuffer can omit it.
+     */
+    setContinuousMode?(enabled: boolean): void;
 }
 /**
  * A {@link VoicePlaybackEngine} that also supports pause/resume. Returned by
@@ -7087,14 +7189,23 @@ interface VoiceProvider {
      * Register a callback for incremental transcript updates during a voice turn.
      * `isFinal=false` is a live interim update (user partials, or assistant deltas
      * on providers that stream them); `isFinal=true` finalizes that role's text.
-     * On the realtime `runtype` path, interim updates fire for the `user` only and
-     * the `assistant` arrives as a single final. Providers with overlapping turns
-     * must include the same turnId on user and assistant finals; without IDs,
-     * providers must discard cancelled output before emitting the next user final.
-     */
-    onTranscript?(callback: (role: 'user' | 'assistant', text: string, isFinal: boolean, metadata?: {
-        turnId?: string;
-    }) => void): void;
+     * On the classic realtime `runtype` path, interim updates fire for the `user`
+     * only and the `assistant` arrives as a single final; without a turnId the
+     * session assumes strictly alternating turns, and providers must discard
+     * cancelled output before emitting the next user final.
+     *
+     * When `metadata.turnId` is present (full-duplex / speech-to-speech, e.g.
+     * GPT-Live), the session reconciles by `(turnId, role)`: each pair owns one
+     * bubble that later updates replace in place, turns may overlap or arrive in
+     * any order, and no empty assistant placeholder is injected. User and
+     * assistant may share a turnId (a late user bubble then renders above its
+     * reply) or use distinct ones. `metadata.startMs` (call-relative ms when the
+     * utterance began) orders a new bubble among the call's other keyed bubbles;
+     * without it bubbles order by first arrival. After a stop, keyed providers
+     * must drop the cancelled reply's output until the server acknowledges the
+     * cancel: the session renders every reply it receives.
+     */
+    onTranscript?(callback: (role: 'user' | 'assistant', text: string, isFinal: boolean, metadata?: VoiceTranscriptMetadata) => void): void;
     /** Register a callback for per-turn latency metrics (realtime path). */
     onMetrics?(callback: (metrics: VoiceMetrics) => void): void;
     /**
@@ -7115,6 +7226,116 @@ interface VoiceProvider {
     deactivateBargeIn?(): Promise<void>;
     /** Stop playback / cancel in-flight request without starting recording */
     stopPlayback?(): void;
+    /**
+     * Warm the provider ahead of `startListening()` (the widget calls it on mic
+     * intent, see `voiceRecognition.prewarm`). Must be fire-and-forget: never
+     * throw, emit errors or status, or acquire the microphone. May be called
+     * repeatedly; throttle as needed.
+     */
+    prewarm?(): void;
+    /**
+     * Called by the session once after construction with the chat-side surface
+     * a full-duplex provider uses to hand agent turns to the chat pipeline
+     * ({@link VoiceSessionBridge}). Providers without client delegation omit it.
+     */
+    setSessionBridge?(bridge: VoiceSessionBridge): void;
+    /** The AI-disclosure notice for the live call, if one should show (speech-to-speech). */
+    getDisclosure?(): string | null;
+}
+/** Metadata on a {@link VoiceProvider.onTranscript} update. */
+type VoiceTranscriptMetadata = {
+    /** Groups updates that belong to one utterance bubble (full duplex). */
+    turnId?: string;
+    /** Call-relative ms when the utterance began; orders late bubbles. */
+    startMs?: number;
+    /** Call-relative ms of the latest audio this update covers. */
+    endMs?: number;
+    /**
+     * The bubble only captions speech (see `AgentWidgetMessage.voiceCaption`).
+     * The chat pipeline owns the conversation, as with client delegation.
+     */
+    caption?: boolean;
+};
+/** An agent turn a full-duplex voice model handed to the widget. */
+type VoiceDelegationRequest = {
+    /** The voice model's delegation id (`delegation_started.delegationId`). */
+    delegationId: string;
+    /** What the visitor asked, as the voice model heard it. */
+    userText: string;
+    /**
+     * The user transcript utterances `userText` joins, oldest first (the last
+     * is the one it came from). May be empty.
+     */
+    userUtteranceIds: string[];
+    /** The voice model's view of the conversation it delegated from. */
+    messages: Array<{
+        role: string;
+        content: unknown;
+    }>;
+};
+/**
+ * How a delegated turn ended, or that it waits on an approval (contract
+ * Amendment 5). `pending_approval` is the only non-terminal status. Open:
+ * a receiver treats an unknown terminal status as `failed`.
+ * - `completed`: the work ran.
+ * - `denied`: the visitor said no (Deny tap, or a spoken decline).
+ * - `timeout`: the approval lapsed (`approvalTimeoutMs`, or the server's own timeout).
+ * - `cancelled`: replaced by a newer request, or abandoned by the system.
+ * - `failed`: an error.
+ */
+type VoiceDelegationStatus = 'completed' | 'pending_approval' | 'denied' | 'timeout' | 'cancelled' | 'failed' | (string & {});
+/** The terminal result of a turn that parked on approvals. */
+type VoiceDelegationFollowUp = {
+    status: VoiceDelegationStatus;
+    /** What to read aloud (Markdown allowed). May be empty: nothing more to say. */
+    text: string;
+};
+/** The delegated turn's answer, sent back for the voice model to read aloud. */
+type VoiceDelegationResult = {
+    /** `pending_approval` when `followUp` is set; otherwise terminal. */
+    status: VoiceDelegationStatus;
+    /** The assistant text (Markdown allowed). */
+    text: string;
+    /**
+     * Set when the turn parked on approvals (`text` then asks for the decision).
+     * Resolves with the terminal result once the visitor decides (or the
+     * approval is replaced, declined by voice, or unanswered for
+     * `approvalTimeoutMs`) and the resumed turn finishes; `null` only when
+     * `signal` aborts first (hang-up, or the server cancelled the delegation).
+     */
+    followUp?: (options: {
+        signal: AbortSignal;
+        approvalTimeoutMs?: number;
+        /** `false`: run only the approval bookkeeping (expiry, supersede); nothing will be read back. @default true */
+        readBack?: boolean;
+        /** The resumed turn stopped on another approval: ask for it too (`text` is its approval script). */
+        onUpdate?: (text: string) => void;
+    }) => Promise<VoiceDelegationFollowUp | null>;
+};
+/**
+ * Chat-side surface the session hands a full-duplex provider: the visible
+ * history (for the call-start context) and a way to run a delegated turn
+ * through the normal chat pipeline.
+ */
+interface VoiceSessionBridge {
+    /** Visible user/assistant messages, oldest first. */
+    getHistory(): Array<{
+        role: 'user' | 'assistant';
+        content: string;
+    }>;
+    /**
+     * Submit the turn as the visitor's message (reusing its transcript bubble)
+     * and resolve with the assistant's final answer once the turn settles.
+     * Waits for any chat turn already in flight first. Never rejects.
+     */
+    runDelegatedTurn(request: VoiceDelegationRequest): Promise<VoiceDelegationResult>;
+    /**
+     * The server dropped this delegation's result (cancelled it, or refused a
+     * late result): its chat answer is no longer spoken, so browser TTS may read it.
+     * `expired`: the server's deadline passed, so its pending approvals are
+     * declined too, as the approval TTL would.
+     */
+    dropDelegation?(delegationId: string, expired?: boolean): void;
 }
 /**
  * Extra context for an approval decision, surfaced to `onDecision` and to the
@@ -9722,6 +9943,11 @@ type AgentWidgetToolCall = {
     startedAt?: number;
     completedAt?: number;
     durationMs?: number;
+    /**
+     * The call paused for an approval that was granted, and the agent re-ran it
+     * as a new tool call: that later call carries the result, this one has none.
+     */
+    superseded?: boolean;
 };
 /**
  * Represents a tool approval request in the chat conversation.
@@ -9743,6 +9969,8 @@ type AgentWidgetApproval = {
      */
     reason?: string;
     parameters?: unknown;
+    /** The paused tool call this approval gates (`approval_start.toolCallId`). */
+    toolCallId?: string;
     resolvedAt?: number;
 };
 type AgentWidgetMessageVariant = "assistant" | "reasoning" | "tool" | "approval";
@@ -9842,6 +10070,13 @@ type AgentWidgetMessage = {
      * }
      */
     voiceProcessing?: boolean;
+    /**
+     * Display-only caption of full-duplex voice speech, such as the voice
+     * model's own words, or something the visitor said that wasn't submitted
+     * as a chat turn. It renders in the transcript but is never sent to the
+     * model as conversation.
+     */
+    voiceCaption?: boolean;
     /**
      * Raw structured payload for this message (e.g., JSON action response).
      * Populated automatically when structured parsers run.
@@ -9957,6 +10192,8 @@ type InjectMessageOptions = {
      * Consumers can detect this in `messageTransform` to render custom UI.
      */
     voiceProcessing?: boolean;
+    /** See {@link AgentWidgetMessage.voiceCaption}. */
+    voiceCaption?: boolean;
     /**
      * Raw structured payload (typically a JSON string) representing the
      * full directive that produced this message: e.g. `{ "text": "...",