@datalayer/agent-runtimes 1.5.17 → 1.5.19

package/lib/specs/protocols.js

@@ -128,12 +128,12 @@ export const TAP_PROTOCOL_0_0_1 = {
     id: 'tap',
     version: '0.0.1',
     name: 'Trusted Agent Protocol',
-    description: "An agent-recognition profile over Web Bot Auth for merchants: a signature tagged agent-browser-auth to browse or agent-payer-auth to pay. Built as a draft from Visa's public Merchant Specifications (2026-10-13) in agent-runtimes; planned for applications until Visa confirms it.",
+    description: "An agent-recognition profile over Web Bot Auth for merchants: a signature tagged agent-browser-auth to browse or agent-payer-auth to pay. Still a draft profile, built from Visa's public Merchant Specifications (2026-10-13) until Visa confirms it. In applications at test merchants only (decided 2026-10-14): the MCP gateway signs each UCP or ACP purchase with agent-payer-auth under the application's own key, and the test merchants verify it; browsing stays Web Bot Auth.",
     publisher: 'Visa',
     layer: 'identity',
     specUrl: 'https://developer.visa.com/capabilities/trusted-agent-protocol/trusted-agent-protocol-specifications',
     specVersion: 'draft (Merchant Specifications, read 2026-10-13)',
-    status: 'planned',
+    status: 'shipped',
     connects: 'Merchants that recognise an agent by its signature',
     packages: {
         python: ['agent-runtimes'],

package/lib/specs/scenes.js

@@ -17,7 +17,7 @@ export const CROP_MONITORING_SCENE_0_0_1 = {
     cast: [
         {
             member: 'crop-monitoring',
-            app: 'crop-monitoring:0.0.1',
+            app: 'crop-monitoring:0.0.2',
             ref: '',
             server: '',
             role: 'initiator',
@@ -330,7 +330,7 @@ export const DISASTER_ASSESSMENT_SCENE_0_0_1 = {
     cast: [
         {
             member: 'event-response',
-            app: 'event-response:0.0.2',
+            app: 'event-response:0.0.3',
             ref: '',
             server: '',
             role: 'initiator',
@@ -354,7 +354,7 @@ export const DISASTER_ASSESSMENT_SCENE_0_0_1 = {
         },
         {
             member: 'disaster-assessment',
-            app: 'disaster-assessment:0.0.1',
+            app: 'disaster-assessment:0.0.2',
             ref: '',
             server: '',
             role: 'contributor',
@@ -368,7 +368,7 @@ export const DISASTER_ASSESSMENT_SCENE_0_0_1 = {
         },
         {
             member: 'change-detection',
-            app: 'change-detection:0.0.1',
+            app: 'change-detection:0.0.2',
             ref: '',
             server: '',
             role: 'contributor',
@@ -755,7 +755,7 @@ export const MONTH_END_CLOSE_SCENE_0_0_1 = {
     cast: [
         {
             member: 'month-end-close',
-            app: 'month-end-close:0.0.1',
+            app: 'month-end-close:0.0.2',
             ref: '',
             server: '',
             role: 'initiator',
@@ -1912,7 +1912,7 @@ export const SALES_AND_ACCOUNTING_SCENE_0_0_1 = {
     cast: [
         {
             member: 'sales',
-            app: 'sales:0.0.2',
+            app: 'sales:0.0.3',
             ref: '',
             server: '',
             role: 'initiator',
@@ -1932,7 +1932,7 @@ export const SALES_AND_ACCOUNTING_SCENE_0_0_1 = {
         },
         {
             member: 'accounting',
-            app: 'accounting:0.0.1',
+            app: 'accounting:0.0.2',
             ref: '',
             server: '',
             role: 'contributor',

package/lib/specs/teams/teams.js

@@ -727,7 +727,7 @@ export const CROP_MONITORING_TEAM_SPEC_0_0_1 = {
     supervisor: {
         name: 'Crop Monitoring',
         ref: '',
-        app: 'crop-monitoring:0.0.1',
+        app: 'crop-monitoring:0.0.2',
         model: '',
         goal: ``,
         instructions: ``,
@@ -766,7 +766,7 @@ export const CROP_MONITORING_TEAM_SPEC_0_0_1 = {
             id: 'crop-monitoring',
             name: '',
             ref: '',
-            app: 'crop-monitoring:0.0.1',
+            app: 'crop-monitoring:0.0.2',
             runsIn: 'runtime',
             role: 'initiator',
             goal: `Take the person's field and period, search the imagery that covers them, and report the vigour, the growth and the fields to watch.`,
@@ -804,7 +804,7 @@ export const DISASTER_ASSESSMENT_TEAM_SPEC_0_0_1 = {
     supervisor: {
         name: 'Event response',
         ref: '',
-        app: 'event-response:0.0.2',
+        app: 'event-response:0.0.3',
         model: '',
         goal: ``,
         instructions: ``,
@@ -843,7 +843,7 @@ export const DISASTER_ASSESSMENT_TEAM_SPEC_0_0_1 = {
             id: 'event-response',
             name: '',
             ref: '',
-            app: 'event-response:0.0.2',
+            app: 'event-response:0.0.3',
             runsIn: 'browser',
             talksTo: [
                 { member: 'disaster-assessment', over: 'a2a' },
@@ -863,7 +863,7 @@ export const DISASTER_ASSESSMENT_TEAM_SPEC_0_0_1 = {
             id: 'disaster-assessment',
             name: '',
             ref: '',
-            app: 'disaster-assessment:0.0.1',
+            app: 'disaster-assessment:0.0.2',
             runsIn: 'runtime',
             role: 'contributor',
             goal: `Estimate the area the event affected and the extent of the damage from the imagery before and after it, over A2A.`,
@@ -879,7 +879,7 @@ export const DISASTER_ASSESSMENT_TEAM_SPEC_0_0_1 = {
             id: 'change-detection',
             name: '',
             ref: '',
-            app: 'change-detection:0.0.1',
+            app: 'change-detection:0.0.2',
             runsIn: 'runtime',
             role: 'contributor',
             goal: `Find what changed on the ground between the two dates from the imagery, over A2A.`,
@@ -1133,7 +1133,7 @@ export const MONTH_END_CLOSE_TEAM_SPEC_0_0_1 = {
     supervisor: {
         name: 'Month-end Close',
         ref: '',
-        app: 'month-end-close:0.0.1',
+        app: 'month-end-close:0.0.2',
         model: '',
         goal: ``,
         instructions: ``,
@@ -1169,7 +1169,7 @@ export const MONTH_END_CLOSE_TEAM_SPEC_0_0_1 = {
             id: 'month-end-close',
             name: '',
             ref: '',
-            app: 'month-end-close:0.0.1',
+            app: 'month-end-close:0.0.2',
             runsIn: 'runtime',
             role: 'initiator',
             goal: `Take the person's request, read the period's books, and report where the close stands: what is done, what is to book, what is still open.`,
@@ -1735,7 +1735,7 @@ export const SALES_AND_ACCOUNTING_TEAM_SPEC_0_0_1 = {
     supervisor: {
         name: 'Sales',
         ref: '',
-        app: 'sales:0.0.2',
+        app: 'sales:0.0.3',
         model: '',
         goal: ``,
         instructions: ``,
@@ -1774,7 +1774,7 @@ export const SALES_AND_ACCOUNTING_TEAM_SPEC_0_0_1 = {
             id: 'sales',
             name: '',
             ref: '',
-            app: 'sales:0.0.2',
+            app: 'sales:0.0.3',
             runsIn: 'browser',
             talksTo: [{ member: 'accounting', over: 'a2a' }],
             role: 'initiator',
@@ -1791,7 +1791,7 @@ export const SALES_AND_ACCOUNTING_TEAM_SPEC_0_0_1 = {
             id: 'accounting',
             name: '',
             ref: '',
-            app: 'accounting:0.0.1',
+            app: 'accounting:0.0.2',
             runsIn: 'runtime',
             role: 'contributor',
             goal: `Answer report requests from the Odoo books, read only, over A2A.`,

package/lib/specs/voices.d.ts

@@ -45,6 +45,10 @@ export interface SpeechModelSpec {
     id: string;
     version: string;
     name: string;
+    /** What it is for, in a sentence or two. */
+    description: string;
+    /** Who published it: the folder Datalayer's storage keeps it under, `<provider>/<id>/<path>`. */
+    provider: string;
     task: 'stt' | 'tts' | 'vad';
     engine: string;
     dtype: string;

package/lib/specs/voices.js

@@ -72,6 +72,8 @@ export const SPEECH_MODEL_CATALOGUE = {
         id: 'kokoro-82m',
         version: '0.0.1',
         name: 'Kokoro 82M',
+        description: "Text to speech on Datalayer's servers, on CPU: hexgrad's Kokoro-82M v1.0 exported to ONNX, in full precision, with the voices file that holds its 54 voices, run by kokoro-onnx. Its phonemes come from eSpeak NG (GPL-3.0), which runs on the server and is never sent to a browser. Full precision, not the 8-bit export: on a 16-core CPU the 8-bit one speaks slower than real time (6 s for 3.9 s of speech), this one about four times faster than real time (1 s).",
+        provider: 'hexgrad',
         task: 'tts',
         engine: 'kokoro-onnx',
         dtype: 'fp32',
@@ -102,6 +104,8 @@ export const SPEECH_MODEL_CATALOGUE = {
         id: 'moonshine-base-en',
         version: '0.0.1',
         name: 'Moonshine Base (English)',
+        description: "Speech to text in the browser, in English: Useful Sensors' Moonshine Base (61M parameters), its ONNX export quantized to 8 bits, run by transformers.js. One of the English models whose weights are MIT; the legacy non-English Moonshine models are under a non-commercial licence and are not listed.",
+        provider: 'usefulsensors',
         task: 'stt',
         engine: 'transformers.js',
         dtype: 'q8',
@@ -162,6 +166,8 @@ export const SPEECH_MODEL_CATALOGUE = {
         id: 'moonshine-tiny-en',
         version: '0.0.1',
         name: 'Moonshine Tiny (English)',
+        description: "Speech to text in the browser, in English: Useful Sensors' Moonshine Tiny (27M parameters), its ONNX export quantized to 8 bits, run by transformers.js. One of the English models whose weights are MIT; the legacy non-English Moonshine models are under a non-commercial licence and are not listed.",
+        provider: 'usefulsensors',
         task: 'stt',
         engine: 'transformers.js',
         dtype: 'q8',
@@ -222,6 +228,8 @@ export const SPEECH_MODEL_CATALOGUE = {
         id: 'silero-vad',
         version: '0.0.1',
         name: 'Silero VAD',
+        description: 'Where speech starts and ends in what the microphone heard, in the browser: Silero VAD (about 2 MB), the build @ricky0123/vad-web 0.0.31 ships and runs on onnxruntime-web. Push-to-talk keeps the speech it finds and drops the rest.',
+        provider: 'snakers4',
         task: 'vad',
         engine: 'vad-web',
         dtype: 'fp32',
@@ -247,6 +255,8 @@ export const SPEECH_MODEL_CATALOGUE = {
         id: 'whisper-base',
         version: '0.0.1',
         name: 'Whisper Base',
+        description: "Speech to text in the browser for the languages Moonshine's MIT models do not speak, French first: OpenAI's Whisper base (74M parameters, 99 languages), its ONNX export quantized to 8 bits, run by transformers.js. A clip at a time.",
+        provider: 'openai',
         task: 'stt',
         engine: 'transformers.js',
         dtype: 'q8',

package/lib/types/agentspecs.d.ts

@@ -1483,6 +1483,11 @@ export interface Agentspec {
     parameters?: Record<string, any>;
     /** Subagent delegation configuration. */
     subagents?: SubAgentsConfig;
+    /**
+     * Its voice, for an agent that is not an application (VOICE.md VO-41):
+     * the fields an Appspec says under `interface.voice`; absent when unsaid.
+     */
+    voice?: AppVoiceSpec;
 }
 /**
  * Where a subagent reached over A2A lives, or how to launch it.

package/lib/types/chat.d.ts

@@ -1004,6 +1004,12 @@ export interface ChatBaseProps {
     voiceSpeaking?: boolean;
     /** Stops the agent's voice: the composer's *Stop speaking*, and the microphone opening (VO-13). */
     onStopSpeaking?: () => void;
+    /**
+     * Reads an answer aloud (VOICE.md VO-21): given, each answer the agent has
+     * finished writing has a *Read aloud* under it, which calls it with the
+     * answer's text.
+     */
+    onReadAloud?: (answer: string) => void;
     /**
      * Show the information icon in the header.
      * When clicked, fires onInformationClick.

package/lib/types/protocol.d.ts

@@ -59,6 +59,23 @@ export interface ProtocolConfig {
      * never the bytes. Set for an application's agent.
      */
     uiFiles?: string;
+    /**
+     * The application's sessions (`…/api/v1/apps/sessions`), each AG-UI thread
+     * one of them: Stop is its session's (`{appSessions}/{thread}/stop`). Set
+     * for an application's agent. The runtime's generic `/ag-ui/terminate`
+     * never knew a session's thread, and a deployment's runtime does not
+     * answer it (OwnRoutesOnly).
+     */
+    appSessions?: string;
+    /**
+     * The runtime answers only the application's own routes (OwnRoutesOnly,
+     * PAP_DATALAYER's release gates): it serves a deployment — kept always on,
+     * launched by a hosted page for it — or is the visitors' runtime. Nothing
+     * outside them is opened: not the monitoring stream (an application's
+     * session asks its approvals in the conversation, over `uiSocket`), not
+     * the sandbox's socket, nor any runtime-wide setting. Needs `appSessions`.
+     */
+    ownRoutesOnly?: boolean;
     /** Additional protocol options */
     options?: Record<string, unknown>;
 }

package/lib/utils/utils.js

@@ -195,7 +195,14 @@ export function createProtocolAdapter(config) {
     };
     switch (config.type) {
         case 'ag-ui':
-            return new AGUIAdapter(adapterConfig);
+            if (config.ownRoutesOnly && !config.appSessions) {
+                // Its runtime answers no generic Stop: an application's own or none.
+                throw new Error('A chat whose runtime answers only the application’s own routes is an application’s: it needs its sessions (appSessions).');
+            }
+            return new AGUIAdapter({
+                ...adapterConfig,
+                ...(config.appSessions ? { appSessions: config.appSessions } : {}),
+            });
         case 'a2a':
             return new A2AAdapter(adapterConfig);
         case 'vercel-ai':

package/lib/voice/ReadAloud.d.ts

@@ -0,0 +1,35 @@
+/**
+ * *Read aloud*, under an answer (VOICE.md VO-21, `on_request`): the answer,
+ * read as plain words and cut into sentences, given to the speaker — what it
+ * was saying stops first. The press is the gesture a browser plays sound
+ * after.
+ *
+ * @module voice/ReadAloud
+ */
+import type { JSX } from 'react';
+import type { SentenceSink } from './useSpokenAnswers';
+/** What *Read aloud* gives an answer to: a sink a press may unlock. */
+export interface ReadingSink extends SentenceSink {
+    unlock?: () => void;
+}
+/** An answer read from its start: what was being said stops, its sentences follow. */
+export declare function readAloud(sink: ReadingSink, answer: string): string[];
+/** A message, as far as *Read aloud* reads it. */
+export interface ReadableMessage {
+    id: string;
+    role: string;
+    live?: boolean;
+    note?: unknown;
+}
+/**
+ * Whether a message has *Read aloud* under it: an answer of the agent with
+ * words, finished — not live, and not the one the agent is writing now.
+ */
+export declare function readsAloud(message: ReadableMessage, text: string, writingId: string | undefined): boolean;
+export interface ReadAloudProps {
+    /** Reads the answer aloud. */
+    onRead: () => void;
+}
+/** The button under an answer. */
+export declare function ReadAloud({ onRead }: ReadAloudProps): JSX.Element;
+export default ReadAloud;

package/lib/voice/ReadAloud.js

@@ -0,0 +1,29 @@
+import { jsx as _jsx } from "react/jsx-runtime";
+import { IconButton, Tooltip } from '@primer/react';
+import { UnmuteIcon } from '@primer/octicons-react';
+import { SentenceCutter } from './sentences';
+/** An answer read from its start: what was being said stops, its sentences follow. */
+export function readAloud(sink, answer) {
+    sink.unlock?.();
+    sink.stop();
+    const cutter = new SentenceCutter();
+    const sentences = [...cutter.feed(answer), ...cutter.end(answer)];
+    sentences.forEach(sentence => sink.say(sentence));
+    return sentences;
+}
+/**
+ * Whether a message has *Read aloud* under it: an answer of the agent with
+ * words, finished — not live, and not the one the agent is writing now.
+ */
+export function readsAloud(message, text, writingId) {
+    return (message.role === 'assistant' &&
+        !message.live &&
+        !message.note &&
+        message.id !== writingId &&
+        text.trim().length > 0);
+}
+/** The button under an answer. */
+export function ReadAloud({ onRead }) {
+    return (_jsx(Tooltip, { text: "Read aloud", direction: "n", children: _jsx(IconButton, { "data-read-aloud": "", icon: UnmuteIcon, "aria-label": "Read aloud", size: "small", variant: "invisible", onClick: onRead }) }));
+}
+export default ReadAloud;

package/lib/voice/hearing.js

@@ -12,7 +12,7 @@
  * @module voice/hearing
  */
 import { SPEECH_MODEL_CATALOGUE, transcriberFor } from '../specs/voices';
-import { pinnedBytes, pinnedFetch, downloadSize } from './pinned';
+import { pinnedBytes, pinnedFetch, downloadSize, modelFolder } from './pinned';
 import { HEARING_RATE } from './capture';
 /** A language no device model hears. */
 export class NotHeardHere extends Error {
@@ -84,7 +84,9 @@ export class DeviceHearing {
                 }
                 const total = downloadSize(model.id);
                 const loaded = new Map();
-                const recognizer = await transformers.pipeline('automatic-speech-recognition', model.id, {
+                const recognizer = await transformers.pipeline('automatic-speech-recognition', 
+                // Read at `<origin>/<provider>/<model id>/`, as the store keeps it.
+                modelFolder(model.id), {
                     dtype: model.dtype,
                     device: 'wasm',
                     progress_callback: (info) => {
@@ -106,7 +108,7 @@ export class DeviceHearing {
     activity() {
         if (!this.vad) {
             const file = SPEECH_MODEL_CATALOGUE['silero-vad'].files[0].path;
-            const url = `${this.modelsUrl.replace(/\/+$/, '')}/silero-vad/${file}`;
+            const url = `${this.modelsUrl.replace(/\/+$/, '')}/${modelFolder('silero-vad')}/${file}`;
             this.vad = this.engines.vad().then(module => module.NonRealTimeVAD.new({
                 modelURL: url,
                 modelFetcher: path => pinnedBytes(this.modelsUrl, path),

package/lib/voice/index.d.ts

@@ -18,6 +18,8 @@ export * from './consent';
 export * from './useSpokenAnswers';
 export { VoiceInput } from './VoiceInput';
 export type { VoiceInputProps, ListeningState } from './VoiceInput';
+export { ReadAloud, readAloud, readsAloud } from './ReadAloud';
+export type { ReadAloudProps, ReadableMessage, ReadingSink } from './ReadAloud';
 /**
  * A chat's voice: what a host gives `ChatBase` and `ChatFloating` to let a
  * person talk to the agent and hear it — the Appspec's `interface.voice`,
@@ -28,8 +30,11 @@ export interface ChatVoice {
     language: string;
     /** `push_to_talk`, or `off`. Hands-free is Phase V3. */
     input: 'off' | 'push_to_talk';
-    /** When answers are heard: `always`, or `off`. */
-    output: 'off' | 'always';
+    /**
+     * When answers are heard: `always`, as they are written; `on_request`, a
+     * *Read aloud* under each answer (VO-21); or `off`.
+     */
+    output: 'off' | 'on_request' | 'always';
     /** The voice of the catalogue answers are spoken with. */
     voice: string;
     /** Datalayer's origin for the browser's models: the pinned store. */

package/lib/voice/index.js

@@ -11,3 +11,4 @@ export * from './speaker';
 export * from './consent';
 export * from './useSpokenAnswers';
 export { VoiceInput } from './VoiceInput';
+export { ReadAloud, readAloud, readsAloud } from './ReadAloud';

package/lib/voice/pinned.d.ts

@@ -1,7 +1,7 @@
 /**
  * The browser's models, from Datalayer's origin and as pinned (VOICE.md
  * VO-49, VO-03): every file a model is made of is fetched from the models'
- * origin, laid out as the pinned store is (`<model id>/<path>`), and its
+ * origin, laid out as the pinned store is (`<provider>/<model id>/<path>`), and its
  * SHA-256 checked against the catalogue before anything reads it. A file
  * that is not the one pinned is refused; nothing is asked of another origin.
  *
@@ -24,5 +24,7 @@ export declare function sha256Hex(bytes: ArrayBuffer): Promise<string>;
 export declare function pinnedFetch(base: string, fetcher?: typeof fetch): (input: string | URL, init?: RequestInit) => Promise<Response>;
 /** A model's file as bytes, checked against its pin. */
 export declare function pinnedBytes(base: string, url: string, fetcher?: typeof fetch): Promise<ArrayBuffer>;
+/** Where a model's files are, under the models' origin: `<provider>/<model id>` (VO-48). */
+export declare function modelFolder(modelId: string): string;
 /** What a first use downloads for a model, in bytes. */
 export declare function downloadSize(modelId: string): number;

package/lib/voice/pinned.js

@@ -5,7 +5,7 @@
 /**
  * The browser's models, from Datalayer's origin and as pinned (VOICE.md
  * VO-49, VO-03): every file a model is made of is fetched from the models'
- * origin, laid out as the pinned store is (`<model id>/<path>`), and its
+ * origin, laid out as the pinned store is (`<provider>/<model id>/<path>`), and its
  * SHA-256 checked against the catalogue before anything reads it. A file
  * that is not the one pinned is refused; nothing is asked of another origin.
  *
@@ -21,9 +21,15 @@ export function pinOf(base, url) {
     if (!url.startsWith(root)) {
         return undefined;
     }
-    const [model, ...rest] = url.slice(root.length).split('?')[0].split('/');
+    const [provider, model, ...rest] = url
+        .slice(root.length)
+        .split('?')[0]
+        .split('/');
     const path = rest.join('/');
-    return SPEECH_MODEL_CATALOGUE[model]?.files.find(file => file.path === path);
+    const spec = SPEECH_MODEL_CATALOGUE[model];
+    return spec?.provider === provider
+        ? spec.files.find(file => file.path === path)
+        : undefined;
 }
 /** The SHA-256 of some bytes, in hex. */
 export async function sha256Hex(bytes) {
@@ -69,6 +75,14 @@ export async function pinnedBytes(base, url, fetcher) {
     }
     return answered.arrayBuffer();
 }
+/** Where a model's files are, under the models' origin: `<provider>/<model id>` (VO-48). */
+export function modelFolder(modelId) {
+    const model = SPEECH_MODEL_CATALOGUE[modelId];
+    if (!model) {
+        throw new PinRefused(`The voice catalogue has no speech model ${modelId}.`);
+    }
+    return `${model.provider}/${model.id}`;
+}
 /** What a first use downloads for a model, in bytes. */
 export function downloadSize(modelId) {
     return (SPEECH_MODEL_CATALOGUE[modelId]?.files ?? []).reduce((total, file) => total + file.size, 0);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@datalayer/agent-runtimes",
-    "version": "1.5.17",
+    "version": "1.5.19",
     "type": "module",
     "workspaces": [
         ".",
@@ -160,7 +160,7 @@
         "@datalayer/lexical-loro": "^1.0.6",
         "@datalayer/primer-addons": "^1.0.39",
         "@datalayer/primer-rjsf": "^1.0.3",
-        "@datalayer/personal-agent-protocol": "^0.23.0",
+        "@datalayer/personal-agent-protocol": "^0.24.0",
         "@datalayer/reactor": "^1.3.0",
         "@datalayer/reactor-commands": "^1.0.4",
         "@datalayer/reactor-manager": "^1.0.4",

package/scripts/codegen/generate_agents.py

@@ -304,9 +304,33 @@ def load_yaml_specs(specs_dir: Path) -> List[tuple[str, Dict[str, Any]]]:
 
     check_model_additionals(resolved, load_model_ids(specs_dir))
     check_protocols(resolved, load_protocol_ids(specs_dir))
+    check_voices(resolved)
     return resolved
 
 
+def check_voices(specs: List[tuple[str, Dict[str, Any]]]) -> None:
+    """Refuse an agent whose `voice` is not an `interface.voice` block of the voice catalogue (VOICE.md VO-41)."""
+    from agentspecs.apps import agent_voice_problems
+
+    problems = [
+        problem
+        for _, spec in specs
+        if "voice" in spec
+        for problem in agent_voice_problems(str(spec.get("id")), spec["voice"])
+    ]
+    if problems:
+        raise SystemExit("\n".join(problems))
+
+
+def voice_of(spec: Dict[str, Any]) -> Dict[str, Any] | None:
+    """An agent's voice with every field said, as the Appspec's block reads it; None when unsaid."""
+    if "voice" not in spec:
+        return None
+    from agentspecs.apps import AppVoice
+
+    return AppVoice(**spec["voice"]).model_dump(mode="json")
+
+
 def load_protocol_ids(specs_dir: Path) -> set[str]:
     """Every id of the protocols catalogue (`agentspecs/protocols`, AGENT_PROTOCOLS PS-01)."""
     ids: set[str] = set()
@@ -525,6 +549,10 @@ from agent_runtimes.types import (
                 if model_additionals
                 else ""
             )
+            voice = voice_of(spec)
+            voice_py_line = (
+                f"    voice={_fmt_py_literal(voice)},\n" if voice is not None else ""
+            )
             inference_provider = spec.get("inference_provider")
             inference_provider_str = (
                 f'"{inference_provider}"' if inference_provider else "None"
@@ -687,7 +715,7 @@ from agent_runtimes.types import (
     tool_hooks={_fmt_py_literal(tool_hooks_val)},
     parameters={_fmt_py_literal(parameters_val)},
     subagents={subagents_str},
-)
+{voice_py_line})
 
 '''
 
@@ -1154,6 +1182,10 @@ const FRONTEND_TOOL_MAP: Record<string, any> = {
                 if model_additionals
                 else ""
             )
+            voice = voice_of(spec)
+            voice_ts_line = (
+                f"    voice: {_fmt_ts_literal(voice)},\n" if voice is not None else ""
+            )
             inference_provider = spec.get("inference_provider")
             inference_provider_ts = (
                 f"'{inference_provider}'" if inference_provider else "undefined"
@@ -1262,7 +1294,7 @@ const FRONTEND_TOOL_MAP: Record<string, any> = {
     toolHooks: {_fmt_ts_literal(tool_hooks_val)},
     parameters: {_fmt_ts_literal(parameters_val)},
     subagents: {subagents_ts},
-}};
+{voice_ts_line}}};
 
 """
 

package/scripts/codegen/generate_voices.py

@@ -45,6 +45,8 @@ def _model(model: Any) -> dict[str, Any]:
         "id": model.id,
         "version": model.version,
         "name": model.name,
+        "description": " ".join(model.description.split()),
+        "provider": model.provider,
         "task": model.task,
         "engine": model.engine,
         "dtype": model.dtype,
@@ -120,7 +122,8 @@ def generate_python_code(
             "    base = (language or '').strip()",
             "    exact = [v for v in VOICE_CATALOGUE.values() if base in v['languages']]",
             "    near = [v for v in VOICE_CATALOGUE.values() if voice_speaks(v, base.split('-')[0])]",
-            "    return (exact or near or [None])[0]",
+            "    found = exact or near",
+            "    return found[0] if found else None",
             "",
         ]
     )
@@ -191,6 +194,10 @@ def generate_typescript_code(
             "  id: string;",
             "  version: string;",
             "  name: string;",
+            "  /** What it is for, in a sentence or two. */",
+            "  description: string;",
+            "  /** Who published it: the folder Datalayer's storage keeps it under, `<provider>/<id>/<path>`. */",
+            "  provider: string;",
             "  task: 'stt' | 'tts' | 'vad';",
             "  engine: string;",
             "  dtype: string;",

package/scripts/voice/measure.py

@@ -4,7 +4,7 @@
 
 """Measure the speech models on the recorded fixtures (VOICE.md VO-05, VO-51, VO-52).
 
-    python scripts/voice/measure.py --store ~/.cache/speech-store
+    python scripts/voice/measure.py --store ~/.cache/voice-work/store
 
 Writes ``tests/voice/measured/*.json``, which ``test_voice_wer.py`` checks
 against ``tests/voice/baselines.json``:
@@ -136,7 +136,7 @@ def main() -> int:
                     print(f"{model['id']} on {len(clips)} {language} clips", flush=True)
                     heard = transcribe(
                         args.store,
-                        model["id"],
+                        f"{model['provider']}/{model['id']}",
                         language,
                         [wavs[c["id"]] for c in clips],
                         env,
@@ -168,7 +168,8 @@ def main() -> int:
             # --- text to speech, and what the device hears of it -----------------------
             from kokoro_onnx import Kokoro
 
-            kokoro_dir = args.store / "kokoro-82m"
+            kokoro_spec = voices.SPEECH_MODEL_CATALOGUE["kokoro-82m"]
+            kokoro_dir = args.store / kokoro_spec["provider"] / "kokoro-82m"
             files = {
                 item["path"]
                 for item in voices.SPEECH_MODEL_CATALOGUE["kokoro-82m"]["files"]
@@ -206,7 +207,11 @@ def main() -> int:
                 ]
                 items = [item for item in spoken if item["language"] == base]
                 heard = transcribe(
-                    args.store, model["id"], base, [item["wav"] for item in items], env
+                    args.store,
+                    f"{model['provider']}/{model['id']}",
+                    base,
+                    [item["wav"] for item in items],
+                    env,
                 )
                 for item, result in zip(items, heard["results"]):
                     round_trips.append(