@mozaik-ai/core 3.10.3 → 3.12.0

package/dist/index.d.mts

@@ -1,4 +1,5 @@
 import OpenAI from 'openai';
+import Anthropic from '@anthropic-ai/sdk';
 
 declare abstract class ContextItem {
     abstract readonly type: string;
@@ -192,8 +193,8 @@ declare class Gpt54Nano implements GenerativeModel {
         supportFunctionCalling: boolean;
     };
     private tools;
-    private readonly effort;
     private streaming;
+    private readonly effort;
     setStreaming(streaming: boolean): void;
     getStreaming(): boolean;
     setTools(tools: Tool[]): void;
@@ -214,11 +215,11 @@ declare class Gpt54 implements GenerativeModel {
     };
     private tools;
     private streaming;
+    private readonly effort;
     setStreaming(streaming: boolean): void;
     getStreaming(): boolean;
     setTools(tools: Tool[]): void;
     getTools(): Tool[];
-    private readonly effort;
     setReasoningEffort(effort: OpenAIReasoningEffortType): void;
     getReasoningEffort(): OpenAIReasoningEffortType;
 }
@@ -234,8 +235,8 @@ declare class Gpt54Mini implements GenerativeModel {
         supportFunctionCalling: boolean;
     };
     private tools;
-    private readonly effort;
     private streaming;
+    private readonly effort;
     setStreaming(streaming: boolean): void;
     getStreaming(): boolean;
     setReasoningEffort(effort: OpenAIReasoningEffortType): void;
@@ -291,7 +292,9 @@ declare class SemanticEvent<T> {
 
 declare class AgenticEnvironment {
     protected subscribers: Participant[];
+    private name?;
     private isActive;
+    constructor(name?: string);
     subscribe(participant: Participant): void;
     unsubscribe(participant: Participant): void;
     deliverFunctionCall(source: Participant, item: FunctionCallItem): void;
@@ -300,16 +303,21 @@ declare class AgenticEnvironment {
     deliverFunctionCallOutput(source: Participant, item: FunctionCallOutputItem): void;
     deliverMessage(source: Participant, message: string): void;
     deliverSemanticEvent(source: Participant, item: SemanticEvent<unknown>): void;
+    private getListeningParticipants;
+    private deliverToSubscribers;
     start(): Promise<void>;
     stop(): void;
 }
 
+type ParticipantClassConstructor = new (...args: any[]) => Participant;
 declare abstract class Participant {
     private environments;
+    protected listens: ParticipantClassConstructor[];
     join(environment: AgenticEnvironment): void;
     leave(environment: AgenticEnvironment): void;
     protected isJoinedTo(environment: AgenticEnvironment): boolean;
     getEnvironments(): AgenticEnvironment[];
+    getListeners(): ParticipantClassConstructor[];
     abstract onParticipantJoined(participant: Participant): Promise<void> | void;
     abstract onParticipantLeft(participant: Participant): Promise<void> | void;
     abstract onJoined(): Promise<void> | void;
@@ -385,12 +393,43 @@ declare class BaseHuman extends Human {
     onExternalEvent(source: Participant, item: SemanticEvent<unknown>): void;
 }
 
+declare class InferenceRequest {
+    readonly model: GenerativeModel & ReasoningEffort<string> & ToolCallingCapability;
+    readonly context: ModelContext;
+    constructor(model: GenerativeModel & ReasoningEffort<string> & ToolCallingCapability, context: ModelContext);
+}
+
+declare class InferenceResponse {
+    readonly contextItems: ContextItem[];
+    readonly tokenUsage?: TokenUsage;
+    constructor(contextItems: ContextItem[], tokenUsage: TokenUsage | undefined);
+}
+
+/**
+ * Provider-agnostic port for a single round of inference.
+ *
+ * Lives in the domain layer; concrete adapters (OpenAI Responses,
+ * Anthropic Messages, DeepSeek Chat Completions, …) live in the
+ * infrastructure layer and are injected into application-layer runners
+ * through their constructor. The application layer must depend on this
+ * interface, never on a concrete provider client.
+ */
+interface ModelRuntime {
+    infer(request: InferenceRequest): Promise<InferenceResponse>;
+}
+
 type InferenceItem = ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>;
-declare class OpenAIInferenceRunner implements InferenceRunner {
+/**
+ * Provider-agnostic inference runner: it depends on
+ * the {@link ModelRuntime} domain port and receives a concrete
+ * infrastructure adapter (OpenAIResponses, AnthropicMessages, …)
+ * through its constructor.
+ */
+declare class DefaultInferenceRunner implements InferenceRunner {
+    private readonly runtime;
+    constructor(runtime: ModelRuntime);
     run(context: ModelContext, model: GenerativeModel, signal?: AbortSignal): AsyncIterable<InferenceItem>;
-    extractTokenUsage(response: OpenAI.Responses.Response): TokenUsage | undefined;
-    mapContextToRequest(context: ModelContext): any[];
-    extractContextItems(response: any): ContextItem[];
+    private isStreamingRuntime;
 }
 
 declare class DefaultFunctionCallRunner implements FunctionCallRunner {
@@ -411,11 +450,11 @@ declare class Gpt55 implements GenerativeModel {
     };
     private tools;
     private streaming;
+    private readonly effort;
     setStreaming(streaming: boolean): void;
     getStreaming(): boolean;
     setTools(tools: Tool[]): void;
     getTools(): Tool[];
-    private readonly effort;
     setReasoningEffort(effort: OpenAIReasoningEffortType): void;
     getReasoningEffort(): OpenAIReasoningEffortType;
 }
@@ -441,4 +480,299 @@ declare class BaseObserver extends Observer {
     onExternalEvent(source: Participant, item: SemanticEvent<unknown>): void;
 }
 
-export { Agent, AgenticEnvironment, BaseAgent, BaseHuman, BaseObserver, ContextItem, DefaultFunctionCallRunner, DeveloperMessageItem, FunctionCallItem, FunctionCallOutputItem, type FunctionCallRunner, type GenerativeModel, Gpt54, Gpt54Mini, Gpt54Nano, Gpt55, Human, InMemoryModelContextRepository, type InferenceRunner, InputTokenDetails, ModelContext, type ModelContextRepository, ModelMessageItem, Observer, OpenAIInferenceRunner, OutputTokenDetails, Participant, ReasoningItem, SemanticEvent, SystemMessageItem, TokenUsage, type Tool, UserMessageItem };
+/**
+ * Optional streaming companion to {@link ModelRuntime}. Adapters whose
+ * provider can stream a round implement this in addition to
+ * `ModelRuntime`; runners delegate to `stream` only when the model has
+ * streaming enabled and the injected runtime actually implements it.
+ */
+interface StreamingRuntime {
+    stream(request: InferenceRequest, signal?: AbortSignal): AsyncIterable<ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>>;
+}
+
+declare class OpenAIResponses implements ModelRuntime, StreamingRuntime {
+    private readonly client;
+    constructor();
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    stream(inferenceRequest: InferenceRequest, signal?: AbortSignal): AsyncIterable<ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>>;
+    private buildRequest;
+    extractTokenUsage(response: OpenAI.Responses.Response): TokenUsage | undefined;
+    mapContextToRequest(context: ModelContext): any[];
+    extractContextItems(response: any): ContextItem[];
+}
+
+/**
+ * Optional connection config. When omitted, the `openai` SDK reads
+ * `OPENAI_API_KEY` and `OPENAI_BASE_URL` from the environment, so the
+ * default-constructed runtime targets whatever `OPENAI_BASE_URL` points
+ * at (real OpenAI when unset). Provider presets (e.g. DeepSeek) pass an
+ * explicit base URL + credential.
+ */
+interface OpenAICompatibleConfig {
+    baseURL?: string;
+    apiKey?: string;
+    /**
+     * Extra request-body fields merged into every `chat.completions`
+     * call. This is how provider-specific quirks are handled **without
+     * a per-provider subclass in mozaik** — the consumer supplies the
+     * vendor-only fields (e.g. DeepSeek's `{ thinking: { type } }`,
+     * safety flags, routing hints). Standard fields the runtime already
+     * sets (`model`, `messages`, `tools`, `reasoning_effort`) take
+     * precedence and are not overwritten.
+     */
+    extraBody?: Record<string, unknown>;
+}
+/**
+ * Generic adapter for any **OpenAI Chat Completions**-compatible
+ * endpoint — real OpenAI, DeepSeek, Xiaomi MiMo, OpenRouter, vLLM,
+ * Ollama, etc. It speaks `/chat/completions` (not the Responses API),
+ * which is the dialect third-party OpenAI-compatible providers expose.
+ *
+ * The base URL and credential are configurable; everything else (the
+ * `ModelContext` ⇄ chat-message conversion, tool-call round-trip, token
+ * usage extraction) is provider-agnostic. Provider-specific request
+ * shaping (e.g. DeepSeek's `thinking` field) is supplied by the
+ * consumer via {@link OpenAICompatibleConfig.extraBody} — mozaik stays
+ * generic and gains no per-provider subclasses.
+ *
+ * Was `DeepSeekChatCompletions`; generalized so consumers can point it
+ * at any OpenAI-compatible endpoint.
+ */
+declare class OpenAICompatibleChatCompletions implements ModelRuntime, StreamingRuntime {
+    private readonly client;
+    private readonly extraBody;
+    constructor(config?: OpenAICompatibleConfig);
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    stream(inferenceRequest: InferenceRequest, signal?: AbortSignal): AsyncIterable<ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>>;
+    buildRequest(inferenceRequest: InferenceRequest): any;
+    extractTokenUsage(response: any): TokenUsage | undefined;
+    mapContextToRequest(context: ModelContext): any[];
+    extractContextItems(response: any): ContextItem[];
+}
+
+declare class AnthropicMessages implements ModelRuntime, StreamingRuntime {
+    private readonly client;
+    constructor();
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    stream(inferenceRequest: InferenceRequest, signal?: AbortSignal): AsyncIterable<ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>>;
+    private buildRequest;
+    extractTokenUsage(response: Anthropic.Messages.Message): TokenUsage | undefined;
+    mapContextToRequest(context: ModelContext): {
+        messages: any[];
+        system?: string;
+    };
+    extractContextItems(response: Anthropic.Messages.Message): ContextItem[];
+    private addContentBlock;
+}
+
+type AnthropicReasoningEffortType = "high" | "medium" | "low" | "none";
+
+declare class ClaudeOpus48 implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: AnthropicReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+declare class ClaudeOpus47 implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: AnthropicReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+declare class ClaudeSonnet46 implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: AnthropicReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    private readonly effort;
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+declare class ClaudeHaiku45 implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: AnthropicReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+/**
+ * DeepSeek V4 reasoning levels. `"none"` selects non-thinking mode
+ * (`thinking: { type: "disabled" }`); any other value enables thinking
+ * mode and is forwarded as the `reasoning_effort` request parameter.
+ */
+type DeepSeekReasoningEffortType = "max" | "high" | "medium" | "low" | "none";
+
+declare class DeepSeekV4Flash implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: DeepSeekReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: DeepSeekReasoningEffortType): void;
+    getReasoningEffort(): DeepSeekReasoningEffortType;
+}
+
+declare class DeepSeekV4Pro implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: DeepSeekReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: DeepSeekReasoningEffortType): void;
+    getReasoningEffort(): DeepSeekReasoningEffortType;
+}
+
+/**
+ * Native Gemini adapter on the `@google/genai` SDK (`generateContent` /
+ * `generateContentStream`). Unlike an OpenAI-compat shim this maps our
+ * domain context to Gemini's native `contents`/`parts` shape, system
+ * instruction, `functionDeclarations`, and `thinkingConfig`, and reads
+ * thought parts + native usage metadata back out. Another `ModelRuntime`
+ * — no runner or port changes.
+ */
+declare class GeminiGenerateContent implements ModelRuntime, StreamingRuntime {
+    private readonly client;
+    constructor();
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    stream(inferenceRequest: InferenceRequest, signal?: AbortSignal): AsyncIterable<ReasoningItem | FunctionCallItem | ModelMessageItem | SemanticEvent<unknown>>;
+    private buildRequest;
+    extractTokenUsage(response: any): TokenUsage | undefined;
+    mapContextToRequest(context: ModelContext): {
+        contents: any[];
+        systemInstruction?: string;
+    };
+    extractContextItems(response: any): ContextItem[];
+    private parseResponse;
+    private addPart;
+}
+
+/**
+ * Gemini reasoning levels, forwarded as the `reasoning_effort` request
+ * parameter on the OpenAI-compatible endpoint. `"none"` only disables
+ * thinking on Gemini 2.5 models; Gemini 3.x always thinks, so for those
+ * the adapter simply omits `reasoning_effort` when set to `"none"` and
+ * lets the model use its default thinking level.
+ */
+type GeminiReasoningEffortType = "high" | "medium" | "low" | "minimal" | "none";
+
+declare class Gemini35Flash implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: GeminiReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: GeminiReasoningEffortType): void;
+    getReasoningEffort(): GeminiReasoningEffortType;
+}
+
+declare class Gemini31Pro implements GenerativeModel {
+    readonly specification: {
+        name: string;
+        supportReasoningEffort: boolean;
+        defaultReasoningEffort: GeminiReasoningEffortType;
+        supportStreaming: boolean;
+        contextWindowSize: number;
+        maxOutputTokens: number;
+        supportFunctionCalling: boolean;
+    };
+    private tools;
+    private streaming;
+    private readonly effort;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    setReasoningEffort(effort: GeminiReasoningEffortType): void;
+    getReasoningEffort(): GeminiReasoningEffortType;
+}
+
+export { Agent, AgenticEnvironment, AnthropicMessages, BaseAgent, BaseHuman, BaseObserver, ClaudeHaiku45, ClaudeOpus47, ClaudeOpus48, ClaudeSonnet46, ContextItem, DeepSeekV4Flash, DeepSeekV4Pro, DefaultFunctionCallRunner, DefaultInferenceRunner, DeveloperMessageItem, FunctionCallItem, FunctionCallOutputItem, type FunctionCallRunner, Gemini31Pro, Gemini35Flash, GeminiGenerateContent, type GenerativeModel, Gpt54, Gpt54Mini, Gpt54Nano, Gpt55, Human, InMemoryModelContextRepository, InferenceRequest, InferenceResponse, type InferenceRunner, InputTokenDetails, ModelContext, type ModelContextRepository, ModelMessageItem, type ModelRuntime, Observer, OpenAICompatibleChatCompletions, OpenAIResponses, OutputTokenDetails, Participant, ReasoningItem, SemanticEvent, type StreamingRuntime, SystemMessageItem, TokenUsage, type Tool, UserMessageItem };