@cerebriline/core 4.100.271 → 4.100.273

package/dist/extensions/tools/media-endpoint.d.ts

@@ -9,7 +9,7 @@
  * Three kinds of server answer these routes, and they say different amounts:
  *  - **opencoti** names what is loaded in `features` on `/props`
  *    (`images_generate_v1`, `images_edit_v1`, `audio_transcriptions_v1`,
- *    `audio_speech_v1`, `videos_generate_v1`), and its `/v1/models` rows carry
+ *    `audio_speech_v1`, `videos_generate_v1`, `audio_generations_v1`), and its `/v1/models` rows carry
  *    `capabilities`. One process can serve the LLM and every media kind.
  *  - **xOllama** answers `/api/xollama`, and its `/v1/models` rows carry
  *    `input_modalities` / `output_modalities` for models with media.
@@ -33,7 +33,7 @@
  * audio/speech`, so the whole route gets pasted. Nothing is lost by taking it
  * back off: a route under a route is never what anyone meant.
  */
-export type MediaKind = "image_generation" | "image_edit" | "transcription" | "speech" | "video";
+export type MediaKind = "image_generation" | "image_edit" | "transcription" | "speech" | "video" | "sound_effects";
 export declare const MEDIA_KINDS: readonly MediaKind[];
 export type MediaServer = "opencoti" | "xollama" | "openai";
 export interface MediaEndpoint {
@@ -79,7 +79,8 @@ export declare function mediaServerOrigin(baseUrl: string): string;
  *
  * Capabilities are exact and win. Modalities are xOllama's listing and need
  * reading: an edit model takes an image and returns one, a transcription model
- * takes audio and returns text.
+ * takes audio and returns text. Sound effects are never read from them: speech
+ * returns audio too, and only the capability tells the two apart.
  */
 export declare function mediaKindsOfModelRow(row: unknown): MediaKind[] | undefined;
 export interface MediaProbeOptions {
@@ -137,13 +138,21 @@ export type ResolvedMediaEndpoint = {
 export declare function resolveMediaEndpoint(input: ResolveMediaEndpointInput): Promise<ResolvedMediaEndpoint>;
 /** `Retry-After` in milliseconds: seconds or a date, bounded, 5 s by default. */
 export declare function retryAfterMs(header: string | null, now?: number): number;
+/**
+ * How long xOllama may hold a request before it starts or answers `503`: its
+ * `mediaQueueWait`. The wait happens inside one attempt, so a tool's limit is
+ * this plus the time the work itself may take. Without it a queued request is
+ * abandoned, and reported as stopped, before the server has answered at all.
+ */
+export declare const MEDIA_QUEUE_WAIT_MS: number;
 export interface MediaRequestOptions {
     fetchImpl?: typeof fetch;
     /** The caller's cancellation. The only thing that ends the waiting. */
     signal?: AbortSignal;
     /**
-     * Milliseconds one attempt may take. Waiting on a busy engine is not
-     * counted: the limit is on the work, not on the queue in front of it.
+     * Milliseconds one attempt may take. Waiting between attempts on a busy
+     * engine is not counted, but a server's own queue is: see
+     * `MEDIA_QUEUE_WAIT_MS`.
      */
     attemptTimeoutMs?: number;
     /**

package/dist/extensions/tools/media-tools.d.ts

@@ -1,8 +1,8 @@
 /**
  * The media tools as one set, built from one configuration.
  *
- * Five tools -- `generate_image`, `edit_image`, `transcribe_audio`,
- * `synthesize_speech`, `generate_video` -- share one rule for where a request
+ * Six tools -- `generate_image`, `edit_image`, `transcribe_audio`,
+ * `synthesize_speech`, `generate_sound_effect`, `generate_video` -- share one rule for where a request
  * goes (see `media-endpoint.ts`), and two hosts offer them: the extension from
  * its settings tabs and the CLI from a file. If each host wrote its own
  * "which tools, pointed where", the two would drift the first time one of them
@@ -14,7 +14,7 @@
  * soon as it names somewhere to go, whether or not that server answers now.
  */
 import type { AgentTool } from "@cerebriline/shared";
-import { type SpeechEndpoint, type TranscriptionEndpoint } from "./audio-tools";
+import { type SoundEffectEndpoint, type SpeechEndpoint, type TranscriptionEndpoint } from "./audio-tools";
 import { type ImageGenerationEndpoint } from "./image-generation";
 import { type MediaEndpointProbe, type MediaSessionProvider, type ResolvedMediaEndpoint } from "./media-endpoint";
 import type { RenderDriver } from "./render-image";
@@ -49,6 +49,10 @@ export interface MediaToolsConfig {
         voice?: string;
         format?: string;
     };
+    /** `generate_sound_effect`. */
+    soundEffects?: MediaSectionSettings & {
+        format?: string;
+    };
     /** `generate_video`. */
     video?: MediaSectionSettings & {
         size?: string;
@@ -56,13 +60,14 @@ export interface MediaToolsConfig {
         format?: string;
     };
 }
-export declare const MEDIA_TOOL_NAMES: readonly ["generate_image", "edit_image", "transcribe_audio", "synthesize_speech", "generate_video"];
+export declare const MEDIA_TOOL_NAMES: readonly ["generate_image", "edit_image", "transcribe_audio", "synthesize_speech", "generate_sound_effect", "generate_video"];
 export type MediaToolName = (typeof MEDIA_TOOL_NAMES)[number];
 interface MediaToolEndpoints {
     generate_image: ImageGenerationEndpoint;
     edit_image: ImageGenerationEndpoint;
     transcribe_audio: TranscriptionEndpoint;
     synthesize_speech: SpeechEndpoint;
+    generate_sound_effect: SoundEffectEndpoint;
     generate_video: VideoGenerationEndpoint;
 }
 export type ResolvedMediaTool<T extends MediaToolName = MediaToolName> = (Omit<Exclude<ResolvedMediaEndpoint, {
@@ -131,6 +136,7 @@ export declare function createMediaTools(options: CreateMediaToolsOptions): Prom
  *   "image": { "model": "flux2-klein", "size": "1024x1024" },
  *   "transcription": true,
  *   "speech": { "voice": "alloy" },
+ *   "soundEffects": { "model": "mannix/stable-audio:3-small-sfx-q8_0" },
  *   "video": { "baseUrl": "http://other:8090", "model": "wan2.1" }
  * }
  * ```