@datalayer/agent-runtimes 1.5.27 → 1.5.30

package/lib/apps/apps/AppKept.js

@@ -4,6 +4,7 @@ import { contribution, definePlugin, signal } from '@datalayer/reactor';
 import { useContributions, useSignalValue } from '@datalayer/reactor/react';
 import { LoopChatTurn, LoopPromptPanel } from '../core';
 import { beforeFirstMessage } from './kept';
+import { ASSISTANT_CLEARANCE_VAR } from '../../chat/assistant/clearance';
 export const APP_KEPT_PLUGIN_NAME = '@datalayer/loop-plugin-app-kept';
 /* A signal to read when no chat contributed a turn: the hook needs one. */
 const NO_TURN = signal({ id: 0, status: 'idle' });
@@ -13,7 +14,10 @@ export function AppKept({ says }) {
     if (!beforeFirstMessage(turn)) {
         return null;
     }
-    return (_jsx("div", { role: "note", style: { padding: '4px 8px' }, children: _jsx(Text, { sx: { fontSize: 0, color: 'fg.muted' }, children: says }) }));
+    return (
+    // Clear of the floating assistant in its corner, at a phone's width too
+    // (REVIEWS Q-08).
+    _jsx("div", { role: "note", style: { padding: `4px var(${ASSISTANT_CLEARANCE_VAR}, 8px) 4px 8px` }, children: _jsx(Text, { sx: { fontSize: 0, color: 'fg.muted' }, children: says }) }));
 }
 /**
  * The plugin, for one application: what it keeps, under its prompt. Made

package/lib/apps/apps/appspec.d.ts

@@ -99,11 +99,6 @@ export declare const PAP_PROFILE: PapProfile;
  * (`pap_profile_problem`); `undefined` when it is one Datalayer speaks.
  */
 export declare function papProfileProblem(value: unknown): string | undefined;
-/**
- * Why `interface.voice.input: hands_free` is refused until VOICE.md VO-12 is
- * built (decided 14 October 2026): never taken for click-to-talk.
- */
-export declare const HANDS_FREE_NOT_BUILT = "hands_free is not built yet (VOICE.md VO-12): say push_to_talk \u2014 click the microphone or hold it, speak, and it stops when let go or once the person stops speaking";
 /** An application's voice (VOICE.md VO-41): off unless said. */
 export declare const DEFAULT_VOICE: AppVoiceSpec;
 /** What a page's output may be drawn with (LOOP P-05), and the property of each that shows its value. */

package/lib/apps/apps/appspec.js

@@ -713,11 +713,6 @@ function parseSurface(data) {
         composedAt: text(data.composed_at),
     };
 }
-/**
- * Why `interface.voice.input: hands_free` is refused until VOICE.md VO-12 is
- * built (decided 14 October 2026): never taken for click-to-talk.
- */
-export const HANDS_FREE_NOT_BUILT = 'hands_free is not built yet (VOICE.md VO-12): say push_to_talk — click the microphone or hold it, speak, and it stops when let go or once the person stops speaking';
 /** An application's voice (VOICE.md VO-41): off unless said. */
 export const DEFAULT_VOICE = {
     enabled: false,
@@ -731,7 +726,7 @@ function parseVoice(data) {
     const voice = isData(data) ? data : {};
     return {
         enabled: voice.enabled === true,
-        input: oneOf(voice.input, ['off', 'push_to_talk'], DEFAULT_VOICE.input),
+        input: oneOf(voice.input, ['off', 'push_to_talk', 'hands_free'], DEFAULT_VOICE.input),
         output: oneOf(voice.output, ['off', 'on_request', 'always'], DEFAULT_VOICE.output),
         voice: text(voice.voice),
         language: text(voice.language),

package/lib/apps/apps/checks.js

@@ -39,7 +39,7 @@ import { getTeamSpec } from '../../specs/teams';
 import { getBackendToolSpec } from '../../specs/backendTools';
 import { getTrack } from '../../specs/tracks';
 import { pluginsOffSetupNotes } from '../plugins/canvas-blocks';
-import { APP_THEME_MODES, APP_THEME_VARIANTS, HANDS_FREE_NOT_BUILT, isAssistantCharacterId, PAGE_OUTPUT_SHOWS, parseAppspec, } from './appspec';
+import { APP_THEME_MODES, APP_THEME_VARIANTS, isAssistantCharacterId, PAGE_OUTPUT_SHOWS, parseAppspec, } from './appspec';
 import { classesOf, splitRef, toolBehaviours } from './rules';
 import { COMMAND_INPUT, COMMAND_NAME, MODE_ID } from './composer';
 import { HOST_NAME, HOST_USERS, hostToolsOf } from './hostTools';
@@ -685,7 +685,7 @@ const ENUMS = {
     visibility: ['private', 'invited', 'organization', 'link', 'public'],
     mode: ['inline', 'bubble', 'panel', 'assistant'],
     balloon: ['history', 'current'],
-    voiceInput: ['off', 'push_to_talk'],
+    voiceInput: ['off', 'push_to_talk', 'hands_free'],
     voiceOutput: ['off', 'on_request', 'always'],
     voiceWhere: ['auto', 'device', 'server'],
     themeVariant: APP_THEME_VARIANTS,
@@ -843,14 +843,9 @@ export function documentShapeProblems(document) {
         if (ui.assistant !== undefined && !isAssistantCharacterId(ui.assistant)) {
             at('interface.assistant', 'names a character by the id a plugin contributes it under, lowercase words joined by a hyphen: `paperclip`, `acme-owl`');
         }
-        // Its voice (VOICE.md VO-41); hands-free refused until it is built (VO-12).
+        // Its voice (VOICE.md VO-41, VO-12).
         mapping(ui.voice, 'interface.voice', voice => {
-            if (voice.input === 'hands_free') {
-                at('interface.voice.input', HANDS_FREE_NOT_BUILT);
-            }
-            else {
-                oneOf(voice.input, ENUMS.voiceInput, 'interface.voice.input');
-            }
+            oneOf(voice.input, ENUMS.voiceInput, 'interface.voice.input');
             oneOf(voice.output, ENUMS.voiceOutput, 'interface.voice.output');
             oneOf(voice.where, ENUMS.voiceWhere, 'interface.voice.where');
         });

package/lib/apps/embed/element.js

@@ -70,7 +70,7 @@ import { parseAppspec } from '../apps/appspec';
 import { readAppspecYaml } from '../apps/yaml';
 import { AppEmbed, embedAssistantCharacter } from './AppEmbed';
 import { WINDOW_NOT_YET } from './embedSession';
-import { EMBED_HOST_VARIABLES, EMBED_OBSERVED_ATTRIBUTES, EMBED_TAG, EmbedAttributeError, adoptPropertiesSetEarly, embedLookOf, inlineHeightOf, languageOf, resumeOf, runnable, } from './embedConfig';
+import { EMBED_HOST_VARIABLES, EMBED_OBSERVED_ATTRIBUTES, EMBED_TAG, EMBED_TOKEN_NAMES_NO_DEPLOYMENT, EmbedAttributeError, adoptPropertiesSetEarly, embedLookOf, embedTokenDeployment, inlineHeightOf, languageOf, resumeOf, runnable, } from './embedConfig';
 import { embedShadowCss } from './embedTheme';
 import { DEFAULT_API, readVisitorApp } from './visitorApp';
 export { runnable };
@@ -164,9 +164,15 @@ export async function readEmbeddedApp({ api, origin, app, token, fetcher = fetch
         return { kind: 'frame', src: frameSrcOf(origin, app, token) };
     }
     const spec = runnable(parseAppspec(definition.spec));
-    return spec.kind === 'decision'
-        ? { kind: 'frame', src: frameSrcOf(origin, app, token) }
-        : { kind: 'spec', app: spec };
+    if (spec.kind === 'decision') {
+        return { kind: 'frame', src: frameSrcOf(origin, app, token) };
+    }
+    // The visit runs the deployment its token names, and no other (VOICE.md
+    // decision 16): a token naming none is refused before anything is drawn.
+    if (!embedTokenDeployment(token)) {
+        return { kind: 'said', text: EMBED_TOKEN_NAMES_NO_DEPLOYMENT };
+    }
+    return { kind: 'spec', app: spec };
 }
 /** The script that is running, when it is a script tag. */
 function currentScriptUrl() {
@@ -377,7 +383,7 @@ export function defineDatalayerAppElement(options = {}) {
                 // A public application, as a visitor without an account (D-07):
                 // an example kept warm or an address open to visitors, run on the
                 // visitors' runtime; a decision framed; a refusal said.
-                const visitor = await readVisitorApp({ api: this.api(), app });
+                const visitor = await readVisitorApp({ api: this.agents(), app });
                 if (visitor.kind === 'decision') {
                     return { kind: 'frame', src: frameSrcOf(this.origin(), visitor.app) };
                 }
@@ -402,11 +408,22 @@ export function defineDatalayerAppElement(options = {}) {
             }
             return this.spoken;
         }
-        /** Datalayer's services: the `api` attribute, else the platform's, else prod1's. */
+        /**
+         * ai-agents, where a visitor's address is read and beside which the
+         * visitors' runtime and ai-inference answer: the `api` attribute, else
+         * the platform's, else r1's. Not Spacer's address (`api()`), which lives
+         * on another plane.
+         */
+        agents() {
+            return (this.getAttribute('api') ||
+                coreStore.getState().configuration?.aiAgentsUrl ||
+                DEFAULT_API).replace(/\/+$/, '');
+        }
+        /** Spacer, where an application is read with its embed token: the `api` attribute, else the platform's, else prod1's. */
         api() {
             return (this.getAttribute('api') ||
                 coreStore.getState().configuration?.spacerUrl ||
-                DEFAULT_API).replace(/\/+$/, '');
+                'https://prod1.datalayer.run').replace(/\/+$/, '');
         }
         async load() {
             const turn = ++this.loading;

package/lib/apps/embed/embedConfig.d.ts

@@ -25,20 +25,29 @@ export declare const EMBED_STYLESHEET_PATH = "/embed/datalayer-app.css";
 export declare const EMBED_ATTRIBUTES: {
     readonly app: "The application’s id on Datalayer.";
     readonly origin: "Where Datalayer is; the origin the script was loaded from by default.";
-    readonly token: "An embed token for an application that is not public, issued to your server for each visit.";
+    readonly token: "An embed token for an application that is not public, issued to your server for each visit; it names the embedded deployment the visit runs, which nothing else on the page names.";
     readonly spec: "The address of an Appspec (YAML or JSON) to run, in place of `app`.";
     readonly mode: "inline, bubble, panel or assistant; the Appspec’s `deployment.embedded.mode` by default, then inline.";
     readonly accent: "green, rose, sky, lime, sun or violet; the Appspec’s `interface.accent` by default, else the theme’s own colours.";
     readonly theme: "light, dark or auto (the visitor’s system); auto by default.";
     readonly font: "A CSS font family for the application’s text; Inter, then the system’s sans-serif, by default.";
     readonly server: "An agent-runtimes server the application’s agent runs on; a Datalayer runtime by default.";
-    readonly api: "Datalayer’s API, where an application is read with its token; the platform’s by default.";
+    readonly api: "Datalayer’s API: where an application is read with its token (Spacer), and where a visitor’s address is read (ai-agents); the platform’s by default.";
     readonly height: "The height of an inline application, in pixels; 640 by default.";
     readonly resume: "true or false: whether a visit’s conversation is picked up again after the page reloads, kept in the page’s storage; true by default.";
     readonly language: "The language the visitor reads the application in, as BCP 47 tags it (fr, pt-BR): its translation into it, and the chat’s own words; the visitor’s browser’s by default.";
     readonly 'user-token': "A token your server signed with the deployment’s secret, naming your user (HS256: sub, name, exp at most an hour away): what an application that says user: signed opens its session with.";
 };
 export type EmbedAttribute = keyof typeof EMBED_ATTRIBUTES;
+/**
+ * The deployment an embed token names (VOICE.md decision 16): the embedded
+ * deployment its owner's server asked it for. Read without vouching for it —
+ * the services that take the token verify it — and the only place the
+ * element learns its deployment: it takes no attribute for one.
+ */
+export declare function embedTokenDeployment(token: string): string | undefined;
+/** What an embed token naming no deployment is told (VOICE.md decision 16). */
+export declare const EMBED_TOKEN_NAMES_NO_DEPLOYMENT = "datalayer-app: the embed token names no deployment. Your server asks for one naming the application\u2019s embedded deployment: POST \u2026/api/spacer/v1/apps/<id>/embed-token with {\"deployment\": \"<its uid>\"}.";
 /** Every attribute the element watches. */
 export declare const EMBED_OBSERVED_ATTRIBUTES: EmbedAttribute[];
 /**

package/lib/apps/embed/embedConfig.js

@@ -16,19 +16,43 @@ export const EMBED_STYLESHEET_PATH = '/embed/datalayer-app.css';
 export const EMBED_ATTRIBUTES = {
     app: 'The application’s id on Datalayer.',
     origin: 'Where Datalayer is; the origin the script was loaded from by default.',
-    token: 'An embed token for an application that is not public, issued to your server for each visit.',
+    token: 'An embed token for an application that is not public, issued to your server for each visit; it names the embedded deployment the visit runs, which nothing else on the page names.',
     spec: 'The address of an Appspec (YAML or JSON) to run, in place of `app`.',
     mode: 'inline, bubble, panel or assistant; the Appspec’s `deployment.embedded.mode` by default, then inline.',
     accent: 'green, rose, sky, lime, sun or violet; the Appspec’s `interface.accent` by default, else the theme’s own colours.',
     theme: 'light, dark or auto (the visitor’s system); auto by default.',
     font: 'A CSS font family for the application’s text; Inter, then the system’s sans-serif, by default.',
     server: 'An agent-runtimes server the application’s agent runs on; a Datalayer runtime by default.',
-    api: 'Datalayer’s API, where an application is read with its token; the platform’s by default.',
+    api: 'Datalayer’s API: where an application is read with its token (Spacer), and where a visitor’s address is read (ai-agents); the platform’s by default.',
     height: 'The height of an inline application, in pixels; 640 by default.',
     resume: 'true or false: whether a visit’s conversation is picked up again after the page reloads, kept in the page’s storage; true by default.',
     language: 'The language the visitor reads the application in, as BCP 47 tags it (fr, pt-BR): its translation into it, and the chat’s own words; the visitor’s browser’s by default.',
     'user-token': 'A token your server signed with the deployment’s secret, naming your user (HS256: sub, name, exp at most an hour away): what an application that says user: signed opens its session with.',
 };
+/**
+ * The deployment an embed token names (VOICE.md decision 16): the embedded
+ * deployment its owner's server asked it for. Read without vouching for it —
+ * the services that take the token verify it — and the only place the
+ * element learns its deployment: it takes no attribute for one.
+ */
+export function embedTokenDeployment(token) {
+    const payload = token.split('.')[1];
+    if (!payload) {
+        return undefined;
+    }
+    try {
+        const base64 = payload.replace(/-/g, '+').replace(/_/g, '/');
+        const claims = JSON.parse(atob(base64.padEnd(base64.length + ((4 - (base64.length % 4)) % 4), '=')));
+        return typeof claims.deployment_uid === 'string' && claims.deployment_uid
+            ? claims.deployment_uid
+            : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+/** What an embed token naming no deployment is told (VOICE.md decision 16). */
+export const EMBED_TOKEN_NAMES_NO_DEPLOYMENT = 'datalayer-app: the embed token names no deployment. Your server asks for one naming the application’s embedded deployment: POST …/api/spacer/v1/apps/<id>/embed-token with {"deployment": "<its uid>"}.';
 /** Every attribute the element watches. */
 export const EMBED_OBSERVED_ATTRIBUTES = Object.keys(EMBED_ATTRIBUTES);
 /**

package/lib/apps/embed/visitorApp.d.ts

@@ -42,8 +42,15 @@ export declare const VISITORS_PATH = "/api/runtime-visitors";
 export declare const VISITOR_EXAMPLES: readonly string[];
 /** The key of an application at its address, as a visitor's token names it. */
 export declare const AT_PREFIX = "at:";
-/** Datalayer's services, where the element names none (`api`). */
-export declare const DEFAULT_API = "https://prod1.datalayer.run";
+/**
+ * Datalayer's services for a visitor, where the element names none (`api`)
+ * and the platform's configuration says none: ai-agents, where an address is
+ * read, and beside it the visitors' runtime and ai-inference, which mints the
+ * visitor's token. Not Spacer's address, which reads an application with its
+ * embed token: the two live on different planes (ai-agents on r1, Spacer on
+ * prod1), found 15 October 2026 on datalayer.ai's embed.
+ */
+export declare const DEFAULT_API = "https://r1.datalayer.run";
 /** An application's id on Datalayer: a ULID. */
 export declare const isAppUid: (value: string) => boolean;
 /** What `app` with no token resolves to. */

package/lib/apps/embed/visitorApp.js

@@ -14,8 +14,15 @@ export const VISITOR_EXAMPLES = [
 ];
 /** The key of an application at its address, as a visitor's token names it. */
 export const AT_PREFIX = 'at:';
-/** Datalayer's services, where the element names none (`api`). */
-export const DEFAULT_API = 'https://prod1.datalayer.run';
+/**
+ * Datalayer's services for a visitor, where the element names none (`api`)
+ * and the platform's configuration says none: ai-agents, where an address is
+ * read, and beside it the visitors' runtime and ai-inference, which mints the
+ * visitor's token. Not Spacer's address, which reads an application with its
+ * embed token: the two live on different planes (ai-agents on r1, Spacer on
+ * prod1), found 15 October 2026 on datalayer.ai's embed.
+ */
+export const DEFAULT_API = 'https://r1.datalayer.run';
 /** An application's id on Datalayer: a ULID. */
 export const isAppUid = (value) => /^[0-9A-HJKMNP-TV-Z]{26}$/.test(value);
 /** An address, as ai-agents spells one. */

package/lib/apps/plugins/chat/ChatView.js

@@ -1321,9 +1321,9 @@ export default function ChatView({ workspace }) {
     /*
      * Its voice asked of the speech service as this chat speaks to its runtime
      * (VOICE.md VO-37, VO-19): the host's token, else this one — a visitor's,
-     * an embed's or the person's — and the deployment it runs.
+     * an embed's or the person's; an embed's names its deployment (decision 16).
      */
-    const askedVoice = useMemo(() => voiceAskedAs(config?.voice, iamToken || undefined, deploymentUid), [config?.voice, iamToken, deploymentUid]);
+    const askedVoice = useMemo(() => voiceAskedAs(config?.voice, iamToken || undefined), [config?.voice, iamToken]);
     /*
      * A deployment's runtime, and the visitors', answer only the application's
      * own routes (OwnRoutesOnly, PAP_DATALAYER's release gates): its chat calls
@@ -1700,7 +1700,10 @@ export default function ChatView({ workspace }) {
         footerExtras: promptActions.length > 0 || (voice && voice.input !== 'off') ? (_jsxs(_Fragment, { children: [promptActions.length > 0 ? (_jsx(ReactorSlot, { slot: LoopSlots.promptAction, props: { workspace } })) : null, voice && voice.input !== 'off' ? (_jsx(ComposerVoice, { voice: voice, input: draft, setInput: setDraft, disabled: chatDisabled, 
                     // The chat's own speaker, never a second one: the microphone
                     // opening stops it, and *Stop speaking* while it speaks.
-                    speaking: speaking, onStopSpeaking: () => controlsRef.current?.stopSpeaking?.(), onInterimChange: setVoiceInterim, 
+                    speaking: speaking, onStopSpeaking: () => controlsRef.current?.stopSpeaking?.(), 
+                    // Hands-free (VO-12, VO-13): the run in progress, stopped when
+                    // the person speaks over it.
+                    running: busy, onStopRun: () => controlsRef.current?.stop?.(), onInterimChange: setVoiceInterim, 
                     // The menu offers when answers are heard where the
                     // application lets them be (VO-21, VO-30).
                     answers: Boolean(config?.voice?.output &&

package/lib/chat/assistant/AssistantStage.js

@@ -7,6 +7,7 @@ import { Dialog, Text } from '@primer/react';
 import { assistantCharacter } from './characters';
 import { BalloonHistoryLarge, SpeechBalloon, } from './SpeechBalloon';
 import { conversationHeaderText } from './BalloonParts';
+import { ASSISTANT_CLEARANCE_VAR } from './clearance';
 import { AssistantContextMenu, LONG_PRESS_MS, opensContextMenu, } from './AssistantContextMenu';
 import { BalloonExpandContext, ExpandedVisual, focusExpanded, } from './BalloonVisual';
 import { SpriteCharacter } from './SpriteCharacter';
@@ -405,6 +406,22 @@ export function AssistantStage({ character, state, size = 88, place, stageRef, o
         telling ||
         detailing !== null ||
         stayPut);
+    // What it covers in its corner, for a line written under a prompt there
+    // (REVIEWS Q-08).
+    const cornerRight = place.left === undefined && place.top === undefined
+        ? pixels(place.right)
+        : undefined;
+    const cornerBottom = pixels(place.bottom);
+    useEffect(() => {
+        if (aside || cornerRight === undefined || cornerBottom === undefined) {
+            return;
+        }
+        const root = document.documentElement;
+        root.style.setProperty(ASSISTANT_CLEARANCE_VAR, `${cornerRight + size}px`);
+        return () => {
+            root.style.removeProperty(ASSISTANT_CLEARANCE_VAR);
+        };
+    }, [aside, cornerRight, cornerBottom, size]);
     useMouth(stageRef, mouthLevel, state === 'speaking');
     // A decision being asked, or its answer, keeps the balloon up.
     const [deciding, setDeciding] = useState(false);

package/lib/chat/assistant/clearance.d.ts

@@ -0,0 +1,10 @@
+/**
+ * What the floating assistant covers from the right edge of the window while
+ * it stands in a bottom corner, as a CSS length on the document: a line
+ * written under a prompt there (what an application keeps, R-31) leaves it
+ * free, so that the character never stands over its words at a phone's width
+ * (REVIEWS Q-08). Unset when it stands anywhere else, steps aside or is gone.
+ *
+ * @module chat/assistant/clearance
+ */
+export declare const ASSISTANT_CLEARANCE_VAR = "--assistant-clearance";

package/lib/chat/assistant/clearance.js

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2025-2026 Datalayer, Inc.
+ * Distributed under the terms of the Modified BSD License.
+ */
+/**
+ * What the floating assistant covers from the right edge of the window while
+ * it stands in a bottom corner, as a CSS length on the document: a line
+ * written under a prompt there (what an application keeps, R-31) leaves it
+ * free, so that the character never stands over its words at a phone's width
+ * (REVIEWS Q-08). Unset when it stands anywhere else, steps aside or is gone.
+ *
+ * @module chat/assistant/clearance
+ */
+export const ASSISTANT_CLEARANCE_VAR = '--assistant-clearance';

package/lib/chat/base/ChatBase.js

@@ -2968,6 +2968,8 @@ showToolApprovalBanner = true, pendingApprovals: pendingApprovalsProp, onApprove
     const ownVoice = useVoiceSpeaker(settled, !hostSpeaks);
     useSpokenAnswers(displayItems, isLoading, !hostSpeaks && settled?.output === 'always', ownVoice.speaker);
     const speakingNow = voiceSpeaking || ownVoice.state.speaking;
+    // Answers spoken over, hands-free (VOICE.md VO-13): marked *interrupted*.
+    const [interruptedIds, setInterruptedIds] = useState(() => new Set());
     const stopSpeaking = onStopSpeaking ?? (ownVoice.speaker ? ownVoice.stop : undefined);
     const readAnswerAloud = onReadAloud ?? (ownVoice.speaker ? ownVoice.read : undefined);
     const handleSend = useCallback(async (messageOverride, 
@@ -3724,10 +3726,10 @@ showToolApprovalBanner = true, pendingApprovals: pendingApprovalsProp, onApprove
             .reverse()
             .find((item) => !isToolCallMessage(item) && item.role === 'assistant')?.id
         : undefined;
-    const readAloudFooter = readAnswerAloud
+    const readAloudFooter = readAnswerAloud || interruptedIds.size
         ? (message) => {
             const text = getMessageText(message);
-            return readsAloud(message, text, writingId) ? (_jsx(ReadAloud, { onRead: () => readAnswerAloud(text) })) : null;
+            return (_jsxs(_Fragment, { children: [interruptedIds.has(message.id) ? (_jsx(Text, { sx: { fontSize: 0, color: 'fg.muted', fontStyle: 'italic' }, "data-interrupted": "", children: "interrupted" })) : null, readAnswerAloud && readsAloud(message, text, writingId) ? (_jsx(ReadAloud, { onRead: () => readAnswerAloud(text) })) : null] }));
         }
         : undefined;
     const messagesContent = children ? (children) : (_jsxs(Box, { display: "flex", flexDirection: "column", minHeight: 0, 
@@ -3745,7 +3747,14 @@ showToolApprovalBanner = true, pendingApprovals: pendingApprovalsProp, onApprove
                 // so every chat with a headless notebook gets the same transcript.
                 notebookToolSurfacesId: notebookToolSurfacesId, approvalConfig: approvalConfig, messagesEndRef: messagesEndRef, onRespond: handleRespond, showTurnFooters: showTurnFooter, renderMessageFooter: readAloudFooter, agentUsage: agentUsage, onRemoveItems: handleRemoveItems, mcpServers: filteredMcpServers, emptyContent: launching ? null : (_jsx(ChatEmptyState, { emptyState: emptyState, brandIcon: brandIcon, description: resolvedDescription, suggestions: resolvedSuggestions, submitOnSuggestionClick: submitOnSuggestionClick, onSuggestionSubmit: handleSuggestionSubmit, onSuggestionFill: handleSuggestionFill })) }), trailingContent] }));
     // The microphone, in the composer's footer (VO-10, VO-11).
-    const voiceControl = settled && settled.input !== 'off' ? (_jsx(ComposerVoice, { voice: settled, input: input, setInput: setInput, disabled: disabled || launching, speaking: speakingNow, onStopSpeaking: stopSpeaking, onInterimChange: setVoiceInterim, 
+    const voiceControl = settled && settled.input !== 'off' ? (_jsx(ComposerVoice, { voice: settled, input: input, setInput: setInput, disabled: disabled || launching, speaking: speakingNow, onStopSpeaking: stopSpeaking, 
+        // Hands-free (VO-12, VO-13): the run in progress, stopped when the
+        // person speaks over it, and the answer marked interrupted.
+        running: isLoading, onStopRun: handleStop, onInterrupted: () => {
+            if (writingId) {
+                setInterruptedIds(ids => new Set([...ids, writingId]));
+            }
+        }, onInterimChange: setVoiceInterim, 
         // The menu offers when answers are heard where the application
         // lets them be (VO-21, VO-30).
         answers: Boolean(voice?.output && voice.output !== 'off' && voice.speechUrl), onSpoken: metadata => {

package/lib/specs/appspecSchema.js

@@ -2522,8 +2522,8 @@ export const APPSPEC_SCHEMA = {
             type: 'string',
         },
         VoiceInput: {
-            description: 'How a person talks to the application (VOICE.md VO-10).\n\n`hands_free` (VO-12) is not one yet: it is refused in a sentence until it\nis built, never taken for click-to-talk.',
-            enum: ['off', 'push_to_talk'],
+            description: 'How a person talks to the application (VOICE.md VO-10, VO-12).\n\n`push_to_talk`: the microphone in the composer, clicked or held.\n`hands_free`: a conversation by voice — the composer offers *Talk with\nit*; a turn ends after a short silence and is sent, the answer is said,\nand listening goes on; speaking over the agent stops its voice (VO-13).\nHands-free implies answers said as they are written (`output: always`).',
+            enum: ['off', 'push_to_talk', 'hands_free'],
             title: 'VoiceInput',
             type: 'string',
         },

package/lib/specs/voices.js

@@ -328,6 +328,95 @@ export const SPEECH_MODEL_CATALOGUE = {
             },
         ],
     },
+    'whisper-large-v3-turbo': {
+        id: 'whisper-large-v3-turbo',
+        version: '0.0.1',
+        name: 'Whisper Large v3 Turbo',
+        description: "Speech to text on Datalayer's servers, on CPU: OpenAI's Whisper large-v3-turbo (809M parameters, 99 languages), converted to CTranslate2 in half precision and run by faster-whisper, its weights quantized to 8 bits when loaded. A recording of up to a minute at a time; the language is detected when the page does not say it.",
+        provider: 'openai',
+        task: 'stt',
+        engine: 'faster-whisper',
+        dtype: 'float16',
+        languages: ['en', 'fr', 'es', 'it', 'pt', 'de', 'nl', 'ja', 'zh', 'hi'],
+        where: ['server'],
+        streaming: false,
+        licence: {
+            weights: 'MIT',
+            code: 'MIT',
+        },
+        attribution: '',
+        upstream: 'https://huggingface.co/openai/whisper-large-v3-turbo',
+        revision: '0a363e9161cbc7ed1431c9597a8ceaf0c4f78fcf',
+        files: [
+            {
+                path: 'config.json',
+                sha256: 'b0253ea6c0d3bea6b1e19e91a02acfd3b53f4467362efcb5a3e6b16c9b3a9b7e',
+                size: 2263,
+            },
+            {
+                path: 'preprocessor_config.json',
+                sha256: '7ccc62c6f2765af1f3b46c00c9b5894426835a05021c8b9c01eecb6dfb542711',
+                size: 340,
+            },
+            {
+                path: 'tokenizer.json',
+                sha256: '297b13372ac43916285644fb9687add3cc62ee2a1adb60da3dc25cc94c1871fd',
+                size: 2710337,
+            },
+            {
+                path: 'vocabulary.json',
+                sha256: 'c69260f2ab26d659b7c398f9a2b2b48ed0df16c3b47d7326782fd9cba71690c1',
+                size: 1068114,
+            },
+            {
+                path: 'model.bin',
+                sha256: 'e76620f83d5f5b69efd3d87e3dc180c1bd21df9fbebacfd4335e5e1efcc018da',
+                size: 1617884929,
+            },
+        ],
+    },
+    'whisper-small': {
+        id: 'whisper-small',
+        version: '0.0.1',
+        name: 'Whisper Small',
+        description: "Speech to text on Datalayer's servers for short spoken turns, on CPU: OpenAI's Whisper small (244M parameters, 99 languages), converted to CTranslate2 and run by faster-whisper, its weights quantized to 8 bits when loaded. About four times faster than large-v3-turbo; on the fixtures it hears English as well (3.3% against 2.4%) and French less well (16.1% against 2.6%). Recordings longer than a turn go to large-v3-turbo.",
+        provider: 'openai',
+        task: 'stt',
+        engine: 'faster-whisper',
+        dtype: 'float16',
+        languages: ['en', 'fr', 'es', 'it', 'pt', 'de', 'nl', 'ja', 'zh', 'hi'],
+        where: ['server'],
+        streaming: false,
+        licence: {
+            weights: 'MIT',
+            code: 'MIT',
+        },
+        attribution: '',
+        upstream: 'https://huggingface.co/openai/whisper-small',
+        revision: '536b0662742c02347bc0e980a01041f333bce120',
+        files: [
+            {
+                path: 'config.json',
+                sha256: 'b55496ac7940a7ae47d2c01eab40edfd8701feec1229d9cce3b40014383fb828',
+                size: 2370,
+            },
+            {
+                path: 'tokenizer.json',
+                sha256: 'fb7b63191e9bb045082c79fd742a3106a12c99513ab30df4a0d47fa6cb6fd0ab',
+                size: 2203239,
+            },
+            {
+                path: 'vocabulary.txt',
+                sha256: '34ce3fe1c5041027b3f8d42912270993f986dbc4bb34cf27f951e34a1e453913',
+                size: 459861,
+            },
+            {
+                path: 'model.bin',
+                sha256: '3e305921506d8872816023e4c273e75d2419fb89b24da97b4fe7bce14170d671',
+                size: 483546902,
+            },
+        ],
+    },
 };
 /** Whether a voice speaks a language: `fr` and `fr-FR` are spoken by a `fr-FR` voice. */
 export function voiceSpeaks(voice, language) {

package/lib/types/agentspecs.d.ts

@@ -958,7 +958,7 @@ export interface AppPageSpec {
 export interface AppVoiceSpec {
     enabled: boolean;
     /** `hands_free` (VO-12) is refused until it is built. */
-    input: 'off' | 'push_to_talk';
+    input: 'off' | 'push_to_talk' | 'hands_free';
     output: 'off' | 'on_request' | 'always';
     voice: string;
     language: string;

package/lib/voice/ComposerVoice.d.ts

@@ -7,7 +7,9 @@
  * person to edit and send — or sent at once with *Send what I say*. While
  * the words are still being heard the host is told (`onInterimChange`), to
  * grey them until they are final (VO-11). Beside the microphone, the
- * person's voice menu (`VoiceMenu`, VO-30).
+ * person's voice menu (`VoiceMenu`, VO-30). Where the application's voice
+ * says `hands_free`, *Talk with it* too (`HandsFree`, VO-12, VO-13): a turn
+ * is sent by itself and the agent's answer may be spoken over.
  *
  * @module voice/ComposerVoice
  */
@@ -29,6 +31,12 @@ export interface ComposerVoiceProps {
     speaking?: boolean;
     /** Stops the agent's voice, also when the microphone opens (VO-13). */
     onStopSpeaking?: () => void;
+    /** The agent is answering: a run is in progress (hands-free, VO-12). */
+    running?: boolean;
+    /** Stops the run while the answer is still being written (barge-in, VO-13). */
+    onStopRun?: () => void;
+    /** The answer being written was spoken over: marked in the conversation (VO-13). */
+    onInterrupted?: () => void;
     /**
      * The words in the box are still being heard (`true`), or final
      * (`false`): the host greys them until they are (VO-11).
@@ -42,5 +50,5 @@ export interface ComposerVoiceProps {
 }
 /** Whether something rides on a run as a spoken message's metadata (VO-27). */
 export declare function spokenOf(value: unknown): SpokenMetadata | undefined;
-export declare function ComposerVoice({ voice, input, setInput, onSpoken, send, disabled, speaking, onStopSpeaking, onInterimChange, answers, }: ComposerVoiceProps): JSX.Element | null;
+export declare function ComposerVoice({ voice, input, setInput, onSpoken, send, disabled, speaking, onStopSpeaking, running, onStopRun, onInterrupted, onInterimChange, answers, }: ComposerVoiceProps): JSX.Element | null;
 export default ComposerVoice;

package/lib/voice/ComposerVoice.js

@@ -1,6 +1,8 @@
 import { jsx as _jsx, jsxs as _jsxs } from "react/jsx-runtime";
 import { useMemo, useRef } from 'react';
 import { DeviceHearing } from './hearing';
+import { HandsFree } from './HandsFree';
+import { ServerHearing, hearingPlace } from './serverHearing';
 import { VoiceInput } from './VoiceInput';
 import { VoiceMenu } from './VoiceMenu';
 /** Whether something rides on a run as a spoken message's metadata (VO-27). */
@@ -13,17 +15,45 @@ export function spokenOf(value) {
         ? said
         : undefined;
 }
-export function ComposerVoice({ voice, input, setInput, onSpoken, send, disabled = false, speaking = false, onStopSpeaking, onInterimChange, answers = false, }) {
-    const hearing = useMemo(() => voice.input !== 'off'
-        ? new DeviceHearing(voice.modelsUrl, voice.engines)
-        : undefined, [voice.input, voice.modelsUrl, voice.engines]);
+export function ComposerVoice({ voice, input, setInput, onSpoken, send, disabled = false, speaking = false, onStopSpeaking, running = false, onStopRun, onInterrupted, onInterimChange, answers = false, }) {
+    // In the browser, or on Datalayer's servers where the application, the
+    // language or the person's detection asks for it (§5, VO-14, VO-34).
+    const place = hearingPlace(voice);
+    const hearing = useMemo(() => {
+        if (voice.input === 'off') {
+            return undefined;
+        }
+        const device = new DeviceHearing(voice.modelsUrl, voice.engines);
+        return place === 'server' && voice.speechUrl
+            ? new ServerHearing(device, {
+                speechUrl: voice.speechUrl,
+                token: voice.token,
+                detect: voice.detect,
+            })
+            : device;
+    }, [
+        voice.input,
+        voice.modelsUrl,
+        voice.engines,
+        place,
+        voice.speechUrl,
+        voice.token,
+        voice.detect,
+    ]);
     // What the box held when the microphone opened: the words heard go after it.
     const base = useRef('');
     const afterBase = (words) => [base.current, words].filter(Boolean).join(' ');
     if (!hearing) {
         return null;
     }
-    return (_jsxs("span", { "data-composer-voice": "", style: { display: 'inline-flex', alignItems: 'center', gap: 2 }, children: [_jsx(VoiceMenu, { voice: voice, answers: answers, disabled: disabled }), _jsx(VoiceInput, { hearing: hearing, language: voice.language, consentKey: voice.consentKey ?? 'chat', silenceMs: voice.silenceMs, microphoneId: voice.microphoneId, disabled: disabled, speaking: speaking, onStopSpeaking: onStopSpeaking, onListen: () => {
+    return (_jsxs("span", { "data-composer-voice": "", style: { display: 'inline-flex', alignItems: 'center', gap: 2 }, children: [_jsx(VoiceMenu, { voice: voice, answers: answers, disabled: disabled }), voice.input === 'hands_free' ? (_jsx(HandsFree, { hearing: hearing, language: voice.language, consentKey: voice.consentKey ?? 'chat', microphoneId: voice.microphoneId, turnSilenceMs: voice.turnSilenceMs, disabled: disabled, running: running, speaking: speaking, onStopSpeaking: onStopSpeaking, onStopRun: onStopRun, onInterrupted: onInterrupted, onInterim: (words) => {
+                    setInput(words ? afterBase(words) : base.current);
+                    onInterimChange?.(words !== '');
+                }, send: (text, metadata) => {
+                    onInterimChange?.(false);
+                    setInput(base.current);
+                    send(text, metadata);
+                } })) : null, _jsx(VoiceInput, { hearing: hearing, language: voice.language, consentKey: voice.consentKey ?? 'chat', silenceMs: voice.silenceMs, microphoneId: voice.microphoneId, disabled: disabled, speaking: speaking, onStopSpeaking: onStopSpeaking, onListen: () => {
                     base.current = input.trim();
                     onStopSpeaking?.();
                 },