mu-core 0.22.0 → 0.23.0

package/esm/index.d.ts

@@ -1,4 +1,4 @@
-export type { ContentPart, Message, Provider, Role, StreamEvent, Tool, Usage } from './types.js';
+export type { ContentPart, Message, ModelModalities, Provider, Role, StreamEvent, Tool, Usage } from './types.js';
 export { audio, image, text } from './types.js';
 export type { Agent, AgentConfig, AgentResult, Input, LoopEvent, RunOptions } from './agent.js';
 export { createAgent, run } from './agent.js';

package/esm/types.d.ts

@@ -49,6 +49,11 @@ export type StreamEvent = ContentPart | {
     type: 'reasoning';
     text: string;
 };
+/** Non-text input modalities a model accepts. */
+export interface ModelModalities {
+    vision: boolean;
+    audio: boolean;
+}
 export interface Provider {
     stream(req: {
         model: string;
@@ -56,4 +61,10 @@ export interface Provider {
         tools: Tool[];
         signal?: AbortSignal;
     }): AsyncIterable<StreamEvent>;
+    /**
+     * Probe a model's input modalities. MAY load the model (e.g. a llama.cpp `/props`
+     * fetch) — call it on model selection, not on every turn. Optional: providers that
+     * can't introspect modalities simply omit it.
+     */
+    capabilities?(model: string): Promise<ModelModalities | undefined>;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mu-core",
-  "version": "0.22.0",
+  "version": "0.23.0",
   "description": "Standalone multimodal agentic loop: content, messages, tools, provider interface, createAgent",
   "license": "MIT",
   "main": "./script/index.js",

package/script/index.d.ts

@@ -1,4 +1,4 @@
-export type { ContentPart, Message, Provider, Role, StreamEvent, Tool, Usage } from './types.js';
+export type { ContentPart, Message, ModelModalities, Provider, Role, StreamEvent, Tool, Usage } from './types.js';
 export { audio, image, text } from './types.js';
 export type { Agent, AgentConfig, AgentResult, Input, LoopEvent, RunOptions } from './agent.js';
 export { createAgent, run } from './agent.js';

package/script/types.d.ts

@@ -49,6 +49,11 @@ export type StreamEvent = ContentPart | {
     type: 'reasoning';
     text: string;
 };
+/** Non-text input modalities a model accepts. */
+export interface ModelModalities {
+    vision: boolean;
+    audio: boolean;
+}
 export interface Provider {
     stream(req: {
         model: string;
@@ -56,4 +61,10 @@ export interface Provider {
         tools: Tool[];
         signal?: AbortSignal;
     }): AsyncIterable<StreamEvent>;
+    /**
+     * Probe a model's input modalities. MAY load the model (e.g. a llama.cpp `/props`
+     * fetch) — call it on model selection, not on every turn. Optional: providers that
+     * can't introspect modalities simply omit it.
+     */
+    capabilities?(model: string): Promise<ModelModalities | undefined>;
 }