@wildix/wilma-agents-client 1.0.49 → 1.0.50

package/dist-types/models/models_0.d.ts

@@ -57,6 +57,176 @@ export interface AgentAccess {
      */
     grants?: AgentGrantPrincipal[] | undefined;
 }
+/**
+ * A pronunciation rule for one phrase. Belongs to a LANGUAGE, not to the agent: 'Wildix' is
+ *     respelled differently for an Italian and a German voice, so one list shared ACROSS languages
+ *     would fight itself. That is why the rules hang off the language and nowhere else — there is no
+ *     company-wide dictionary above them to inherit from, or to disagree with.
+ * @public
+ */
+export interface VoicePronunciation {
+    /**
+     * The phrase as it appears in the text, e.g. 'Wildix'. Bounded because every rule is compiled into ONE regular expression — see the list below.
+     * @public
+     */
+    phrase: string;
+    /**
+     * Respelling (ALIAS) or phonetic transcription (IPA), e.g. 'Wil-dix'.
+     * @public
+     */
+    as: string;
+    /**
+     * Defaults to ALIAS.
+     * @public
+     */
+    encoding?: VoicePronunciationEncoding | undefined;
+}
+/**
+ * One language the assistant audio channel can recognise and speak. There is deliberately no
+ *     greeting here: an assistant conversation is already open when its audio call starts, so there
+ *     is no new conversation to welcome and no initial phrase for the runtime to choose or repeat.
+ * @public
+ */
+export interface AssistantLanguage {
+    /**
+     * BCP-47 code with locale, e.g. 'en-US', 'it-IT'. Reuses the channel-neutral language value so assistant and voice configuration accept the same codes without sharing policy structures.
+     * @public
+     */
+    code: string;
+    /**
+     * Voice this language speaks with. Reuses the channel-neutral URI value so providers and
+     *         synthesis parameters have the same representation in every channel without coupling their
+     *         language rules.
+     *
+     *         REQUIRED ONLY FOR THE 'voice' ENGINE, enforced in the service rather than here (at
+     *         `validateAssistantSpeechSettings`, which reads the engine this contract has no view of).
+     *         'gpt-live', 'gemini-live' and 'gemini-live-thinking' hold the audio and speak with the
+     *         gateway's own voice (see AssistantAudioEngine); a language row for one of them carries no
+     *         voice, and a caller configuring a gateway agent must not have to invent a synthesis URI
+     *         nobody will ever hear just to satisfy this member.
+     * @public
+     */
+    voice?: string | undefined;
+    /**
+     * Marks the language used when the caller's language matches nothing configured.
+     *
+     *         THE FULL RESOLUTION ORDER, because a client author reading only this contract cannot
+     *         otherwise find it: an exact match on the call's language, then a match on its primary
+     *         subtag (`fr` finds `fr-FR`), then the language marked here, then a language whose primary
+     *         subtag is `en`, then the first entry in the list. The same order the voice channel resolves
+     *         by — the rule belongs to choosing a language, not to a channel.
+     *
+     *         At most one language may carry it. A record with two — reachable through the API, where no
+     *         form prevents it — resolves to the first marked one in list order: refusing an
+     *         already-open assistant conversation over configuration drift is worse than serving it.
+     * @public
+     */
+    isDefault?: boolean | undefined;
+    /**
+     * The rules this language is spoken with, same shape and same per-language ownership as
+     *         VoiceLanguage.pronunciations — a respelling is about how a synthesized reply SOUNDS, not
+     *         about how a conversation starts, so the rationale that leaves `greeting` off this structure
+     *         does not apply here. Absent means no respellings for this language, distinct from the voice
+     *         channel's OWN configured rules for the same language: an agent's assistant audio does not
+     *         inherit them, so a name gets fixed once and read once, on whichever channel it is saved.
+     *
+     *         READ ONLY BY THE 'voice' ENGINE, same conditionality as `voice` above and for the same
+     *         reason: 'gpt-live', 'gemini-live' and 'gemini-live-thinking' hold the audio and speak with
+     *         the gateway's own voice, never this service's synthesis path, so a rule saved here on a
+     *         gateway agent configures a respelling nothing will ever apply.
+     * @public
+     */
+    pronunciations?: VoicePronunciation[] | undefined;
+}
+/**
+ * Speech-to-text configuration owned by the assistant channel, so changes to assistant
+ *     recognition policy cannot alter telephony calls.
+ *
+ *     NO `autoDetect` HERE, unlike the voice channel: on this channel configuring a second language
+ *     IS the request to recognise which one is being spoken — nobody lists two and then asks to be
+ *     heard in only one — so the server derives identification from the language list and clamps it
+ *     to false when the recogniser cannot identify among them, or there are too few to identify. The
+ *     voice channel keeps the switch because a dialplan can hand a call its language, which is a
+ *     choice an assistant conversation never has to make.
+ * @public
+ */
+export interface AssistantTranscriptionSettings {
+    /**
+     * Recogniser used for assistant audio. Absent resolves to the platform default, resolved by the server so exactly one component decides it.
+     * @public
+     */
+    engine?: string | undefined;
+}
+/**
+ * Recognition and synthesis policy for assistant audio. This is assistant-owned rather than a
+ *     VoiceSpeechSettings reuse so telephony and assistant audio can evolve independently. It has no
+ *     language greeting and no greet-every-language switch because an assistant conversation is
+ *     already open when its audio call starts, leaving nothing for the channel to greet.
+ * @public
+ */
+export interface AssistantSpeechSettings {
+    /**
+     * Languages the assistant audio channel can recognise and speak. The @length bound is a
+     *         STORAGE ceiling, not a promise about what a recogniser accepts at once — unlike the voice
+     *         channel, where a caller opts into identification with `autoDetect`, THIS channel derives
+     *         identification from `languages` and turns it on whenever the recogniser can identify among
+     *         more than one, so every one of the up to ten rows here can become a live candidate list. The
+     *         service does not clamp that list to a recogniser's own practical limit today (Amazon
+     *         Transcribe's documented recommendation for streaming identification is five); a caller
+     *         configuring more than a recogniser accepts should expect degraded identification or a
+     *         rejected stream, not a refusal from this contract.
+     * @public
+     */
+    languages: AssistantLanguage[];
+    /**
+     * Which recogniser listens to assistant audio. Absent uses the platform default recogniser. Whether it identifies the spoken language is derived from `languages` rather than configured — see AssistantTranscriptionSettings.
+     * @public
+     */
+    transcription?: AssistantTranscriptionSettings | undefined;
+}
+/**
+ * Switchable audio for the assistant channel. The settings object keeps enablement beside its
+ *     speech policy, matching the voice channel's switchable blocks so a disabled configuration can
+ *     be retained without serving audio.
+ * @public
+ */
+export interface AssistantAudioSettings {
+    /**
+     * Whether the assistant accepts audio calls. Required so the presence of saved speech configuration is never mistaken for enablement.
+     * @public
+     */
+    enabled: boolean;
+    /**
+     * Which runtime carries the call. Absent is 'voice'.
+     * @public
+     */
+    engine?: string | undefined;
+    /**
+     * Voice the gateway speaks with. Read only by 'gpt-live', 'gemini-live' and 'gemini-live-thinking' —
+     *         the same three engines `AssistantLanguage.voice` is silent for, and for the mirror-image reason:
+     *         that member is read only by the 'voice' engine, which is the one engine that synthesizes; this one
+     *         is read only by the gateways, which are the ones with a voice of their own to name. Absent is the
+     *         gateway's own default voice. Ignored, but kept, by the 'voice' engine — an agent switched back to
+     *         it keeps whatever was saved here, unused until it is switched to a gateway again.
+     *
+     *         A NAME, NOT A URI, unlike `AssistantLanguage.voice`. A gateway's voice is never synthesized by this
+     *         stack, so there is no provider scheme to prefix and no synthesis parameter to carry on it — only
+     *         the vendor's own prebuilt voice name ('marin', 'Kore', ...), exactly as the gateway's own API takes
+     *         it.
+     *
+     *         VALIDATED AGAINST THE SELECTED GATEWAY'S OWN CATALOG, not a shared one: OpenAI's Realtime voices
+     *         and Gemini's prebuilt voices do not overlap. Enforced in the service
+     *         (`validateAssistantLiveVoice`, wilma-common), which reads the engine this contract has no view
+     *         of — the same split `AssistantLanguage.voice`'s own REQUIRED ONLY FOR note describes.
+     * @public
+     */
+    liveVoice?: string | undefined;
+    /**
+     * Languages, voices and recognition policy for assistant audio. Absent leaves those choices to platform defaults, including when audio is enabled.
+     * @public
+     */
+    speech?: AssistantSpeechSettings | undefined;
+}
 /**
  * A starter prompt offered as a card on an empty thread. Picking it sends `prompt` as the user's first message, so the user never types it.
  * @public
@@ -116,6 +286,7 @@ export interface AssistantUiSettings {
     placeholder?: string | undefined;
 }
 /**
+ * Assistant channel settings. Presence enables the in-app assistant surface; audio remains unavailable unless its separate settings are present and enabled.
  * @public
  */
 export interface AssistantChannelSettings {
@@ -129,6 +300,11 @@ export interface AssistantChannelSettings {
      * @public
      */
     ui?: AssistantUiSettings | undefined;
+    /**
+     * Audio configuration owned by the assistant channel. Absent means the assistant has no audio at all, preserving the behaviour of every existing agent without consulting `channels.voice`.
+     * @public
+     */
+    audio?: AssistantAudioSettings | undefined;
 }
 /**
  * @public
@@ -331,30 +507,6 @@ export interface VoiceSilenceTimeoutSettings {
     seconds: number;
     action: VoiceTerminateAction;
 }
-/**
- * A pronunciation rule for one phrase. Belongs to a LANGUAGE, not to the agent: 'Wildix' is
- *     respelled differently for an Italian and a German voice, so one list shared ACROSS languages
- *     would fight itself. That is why the rules hang off the language and nowhere else — there is no
- *     company-wide dictionary above them to inherit from, or to disagree with.
- * @public
- */
-export interface VoicePronunciation {
-    /**
-     * The phrase as it appears in the text, e.g. 'Wildix'. Bounded because every rule is compiled into ONE regular expression — see the list below.
-     * @public
-     */
-    phrase: string;
-    /**
-     * Respelling (ALIAS) or phonetic transcription (IPA), e.g. 'Wil-dix'.
-     * @public
-     */
-    as: string;
-    /**
-     * Defaults to ALIAS.
-     * @public
-     */
-    encoding?: VoicePronunciationEncoding | undefined;
-}
 /**
  * One configured language: its voice, its initial phrase and its pronunciation rules.
  * @public
@@ -453,9 +605,14 @@ export interface VoiceTranscriptionSettings {
  */
 export interface VoiceSpeechSettings {
     /**
-     * Configured languages. The @length bound is a STORAGE ceiling, not the product rule: the real
-     *         cap is the recogniser's own candidate limit (five for Amazon Transcribe), enforced in the
-     *         service so a recogniser with a different limit needs no model change.
+     * Configured languages. The @length bound is a STORAGE ceiling, not a promise about what a
+     *         recogniser accepts at once: automatic detection here is opt-in (`transcription.autoDetect`),
+     *         and once a caller turns it on, the service does not clamp the resulting candidate list to a
+     *         recogniser's own practical limit today (Amazon Transcribe's documented recommendation for
+     *         streaming identification is five) — a caller who both opts in and configures more than a
+     *         recogniser accepts should expect degraded identification or a rejected stream, not a refusal
+     *         from this contract. With detection off, only the single resolved language is ever sent, so
+     *         the ceiling never becomes a candidate-list problem at all.
      * @public
      */
     languages: VoiceLanguage[];
@@ -521,6 +678,10 @@ export interface AgentChannels {
      */
     voice?: VoiceChannelSettings | undefined;
     chat?: ChatChannelSettings | undefined;
+    /**
+     * Assistant channel settings. Presence enables the in-app assistant surface; audio remains unavailable unless its separate settings are present and enabled.
+     * @public
+     */
     assistant?: AssistantChannelSettings | undefined;
 }
 /**
@@ -1031,17 +1192,17 @@ export interface AgentKnowledgeBase {
     annotations?: AgentToolAnnotation[] | undefined;
 }
 /**
- * Sandbox configured inline on the agent, without a pre-built snapshot.
+ * Sandbox launched from a pre-built snapshot managed by the wilma-sandbox service.
  * @public
  */
-export interface AgentInlineSandboxOptions {
+export interface AgentSnapshotSandboxOptions {
     /**
-     * Platform CPU, memory and workspace allocation. Defaults to SMALL; see SandboxSize.
+     * Snapshot to launch sandboxes from. Managed in the wilma-sandbox service; must be READY.
      * @public
      */
-    size?: SandboxSize | undefined;
+    snapshotId: string;
     /**
-     * Network access. Defaults to deny-all.
+     * Override the snapshot's network policy for this agent. Defaults to the snapshot's policy.
      * @public
      */
     network?: SandboxNetworkAccess | undefined;
@@ -1050,19 +1211,24 @@ export interface AgentInlineSandboxOptions {
      * @public
      */
     allowedDomains?: string[] | undefined;
+    /**
+     * Override the compute size for this agent. Defaults to the snapshot's size.
+     * @public
+     */
+    size?: SandboxSize | undefined;
 }
 /**
- * Sandbox launched from a pre-built snapshot managed by the wilma-sandbox service.
+ * Vercel Sandbox configuration.
  * @public
  */
-export interface AgentSnapshotSandboxOptions {
+export interface AgentVercelSandboxOptions {
     /**
-     * Snapshot to launch sandboxes from. Managed in the wilma-sandbox service; must be READY.
+     * Number of vCPUs to allocate (2048 MB memory each).
      * @public
      */
-    snapshotId: string;
+    vcpus?: number | undefined;
     /**
-     * Override the snapshot's network policy for this agent. Defaults to the snapshot's policy.
+     * Network access. Defaults to deny-all.
      * @public
      */
     network?: SandboxNetworkAccess | undefined;
@@ -1071,27 +1237,22 @@ export interface AgentSnapshotSandboxOptions {
      * @public
      */
     allowedDomains?: string[] | undefined;
-    /**
-     * Defaults to the snapshot's built size. If supplied, must match that size: prepared images cannot be resized at launch. Build a separate snapshot to use another size.
-     * @public
-     */
-    size?: SandboxSize | undefined;
 }
 /**
  * Embedded sandbox: shell + file tools backed by an execution environment. The only capability with no `channels` list — being a union it has nowhere to carry one — so it is offered on every channel the agent serves.
  * @public
  */
-export type AgentSandboxCapability = AgentSandboxCapability.InlineMember | AgentSandboxCapability.SnapshotMember | AgentSandboxCapability.$UnknownMember;
+export type AgentSandboxCapability = AgentSandboxCapability.SnapshotMember | AgentSandboxCapability.VercelMember | AgentSandboxCapability.$UnknownMember;
 /**
  * @public
  */
 export declare namespace AgentSandboxCapability {
     /**
-     * Ad-hoc sandbox configured inline on the agent.
+     * Ad-hoc Vercel Sandbox configured inline on the agent.
      * @public
      */
-    interface InlineMember {
-        inline: AgentInlineSandboxOptions;
+    interface VercelMember {
+        vercel: AgentVercelSandboxOptions;
         snapshot?: never;
         $unknown?: never;
     }
@@ -1100,7 +1261,7 @@ export declare namespace AgentSandboxCapability {
      * @public
      */
     interface SnapshotMember {
-        inline?: never;
+        vercel?: never;
         snapshot: AgentSnapshotSandboxOptions;
         $unknown?: never;
     }
@@ -1108,7 +1269,7 @@ export declare namespace AgentSandboxCapability {
      * @public
      */
     interface $UnknownMember {
-        inline?: never;
+        vercel?: never;
         snapshot?: never;
         $unknown: [string, any];
     }
@@ -1117,7 +1278,7 @@ export declare namespace AgentSandboxCapability {
      *
      */
     interface Visitor<T> {
-        inline: (value: AgentInlineSandboxOptions) => T;
+        vercel: (value: AgentVercelSandboxOptions) => T;
         snapshot: (value: AgentSnapshotSandboxOptions) => T;
         _: (name: string, value: any) => T;
     }
@@ -3440,6 +3601,11 @@ export interface AgentInfo {
      * @public
      */
     speechConfigured?: boolean | undefined;
+    /**
+     * Whether an end-user client should offer a microphone for this agent: `channels.assistant.audio.enabled` of the definition the agent CURRENTLY RUNS — the deployment's routed version, else the latest published version, else the draft only when nothing was ever published — never an unpublished draft edit. That is the same definition the assistant voice socket resolves and gates on, so `true` here means a call will be accepted; the console preview, which dials `agentDraft=true`, decides from its own form state instead. `ChannelType.ASSISTANT` on `channels` says the assistant surface is served; this is the one bit end-user clients need on top of it, since `channels.assistant.audio` itself carries speech policy an end user must not see. Named for the flag it carries rather than the block it is read from — `assistant` two members below is the whole `AssistantUiSettings` structure, and a name that read like another one of those would point a reader at the wrong shape. Absent is read as false.
+     * @public
+     */
+    assistantAudioEnabled?: boolean | undefined;
     /**
      * Ids of the skills attached to this agent. Always active in this agent's sessions; user skill preferences cannot disable them.
      * @public

package/dist-types/schemas/schemas_0.d.ts

@@ -36,7 +36,6 @@ export declare var AgentHandoverPipeline$: StaticStructureSchema;
 export declare var AgentHangupCapability$: StaticStructureSchema;
 export declare var AgentHangupPipeline$: StaticStructureSchema;
 export declare var AgentInfo$: StaticStructureSchema;
-export declare var AgentInlineSandboxOptions$: StaticStructureSchema;
 export declare var AgentKnowledgeBase$: StaticStructureSchema;
 export declare var AgentSettings$: StaticStructureSchema;
 export declare var AgentSnapshotSandboxOptions$: StaticStructureSchema;
@@ -60,6 +59,7 @@ export declare var AgentToolVariablesAnnotation$: StaticStructureSchema;
 export declare var AgentTransferCapability$: StaticStructureSchema;
 export declare var AgentTransferPipeline$: StaticStructureSchema;
 export declare var AgentVariable$: StaticStructureSchema;
+export declare var AgentVercelSandboxOptions$: StaticStructureSchema;
 export declare var AgentVersion$: StaticStructureSchema;
 export declare var AgentVersionChange$: StaticStructureSchema;
 export declare var AgentVersionInfo$: StaticStructureSchema;
@@ -96,8 +96,12 @@ export declare var AgentWorkflowStep$: StaticStructureSchema;
 export declare var AgentWorkflowStepPosition$: StaticStructureSchema;
 export declare var AgentWorkflowTransition$: StaticStructureSchema;
 export declare var AgentWorkflowVariablesHandler$: StaticStructureSchema;
+export declare var AssistantAudioSettings$: StaticStructureSchema;
 export declare var AssistantChannelSettings$: StaticStructureSchema;
+export declare var AssistantLanguage$: StaticStructureSchema;
 export declare var AssistantQuickCommand$: StaticStructureSchema;
+export declare var AssistantSpeechSettings$: StaticStructureSchema;
+export declare var AssistantTranscriptionSettings$: StaticStructureSchema;
 export declare var AssistantUiSettings$: StaticStructureSchema;
 export declare var AssistantWelcomeSettings$: StaticStructureSchema;
 export declare var ChatChannelSettings$: StaticStructureSchema;

package/package.json

@@ -1,7 +1,7 @@
 {
     "name": "@wildix/wilma-agents-client",
     "description": "@wildix/wilma-agents-client client",
-    "version": "1.0.49",
+    "version": "1.0.50",
     "scripts": {
         "build": "concurrently 'yarn:build:cjs' 'yarn:build:es' 'yarn:build:types'",
         "build:cjs": "tsc -p tsconfig.cjs.json",
@@ -56,4 +56,4 @@
     "browser": {
         "./dist-es/runtimeConfig": "./dist-es/runtimeConfig.browser"
     }
-}
+}