@mozaik-ai/core 3.10.3 → 3.11.0

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;
@@ -385,12 +386,47 @@ 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>;
+/**
+ * Application-layer inference runner. Provider-agnostic: it depends on
+ * the {@link ModelRuntime} domain port and receives a concrete
+ * infrastructure adapter (OpenAIResponses, AnthropicMessages, …)
+ * through its constructor. The runner never instantiates a provider
+ * client itself — wiring happens at the composition root.
+ *
+ * Name retained for backward compatibility; the runner works with any
+ * `ModelRuntime`, not just OpenAI.
+ */
 declare class OpenAIInferenceRunner 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 {
@@ -441,4 +477,103 @@ 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[];
+}
+
+declare class AnthropicMessages implements ModelRuntime {
+    private readonly client;
+    constructor();
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    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 ClaudeOpus47 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 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;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    private readonly effort;
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+export { Agent, AgenticEnvironment, AnthropicMessages, BaseAgent, BaseHuman, BaseObserver, ClaudeHaiku45, ClaudeOpus47, ClaudeSonnet46, ContextItem, DefaultFunctionCallRunner, DeveloperMessageItem, FunctionCallItem, FunctionCallOutputItem, type FunctionCallRunner, type GenerativeModel, Gpt54, Gpt54Mini, Gpt54Nano, Gpt55, Human, InMemoryModelContextRepository, InferenceRequest, InferenceResponse, type InferenceRunner, InputTokenDetails, ModelContext, type ModelContextRepository, ModelMessageItem, type ModelRuntime, Observer, OpenAIInferenceRunner, OpenAIResponses, OutputTokenDetails, Participant, ReasoningItem, SemanticEvent, type StreamingRuntime, SystemMessageItem, TokenUsage, type Tool, UserMessageItem };

package/dist/index.d.ts

@@ -1,4 +1,5 @@
 import OpenAI from 'openai';
+import Anthropic from '@anthropic-ai/sdk';
 
 declare abstract class ContextItem {
     abstract readonly type: string;
@@ -385,12 +386,47 @@ 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>;
+/**
+ * Application-layer inference runner. Provider-agnostic: it depends on
+ * the {@link ModelRuntime} domain port and receives a concrete
+ * infrastructure adapter (OpenAIResponses, AnthropicMessages, …)
+ * through its constructor. The runner never instantiates a provider
+ * client itself — wiring happens at the composition root.
+ *
+ * Name retained for backward compatibility; the runner works with any
+ * `ModelRuntime`, not just OpenAI.
+ */
 declare class OpenAIInferenceRunner 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 {
@@ -441,4 +477,103 @@ 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[];
+}
+
+declare class AnthropicMessages implements ModelRuntime {
+    private readonly client;
+    constructor();
+    infer(inferenceRequest: InferenceRequest): Promise<InferenceResponse>;
+    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 ClaudeOpus47 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 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;
+    setStreaming(streaming: boolean): void;
+    getStreaming(): boolean;
+    setTools(tools: Tool[]): void;
+    getTools(): Tool[];
+    private readonly effort;
+    setReasoningEffort(effort: AnthropicReasoningEffortType): void;
+    getReasoningEffort(): AnthropicReasoningEffortType;
+}
+
+export { Agent, AgenticEnvironment, AnthropicMessages, BaseAgent, BaseHuman, BaseObserver, ClaudeHaiku45, ClaudeOpus47, ClaudeSonnet46, ContextItem, DefaultFunctionCallRunner, DeveloperMessageItem, FunctionCallItem, FunctionCallOutputItem, type FunctionCallRunner, type GenerativeModel, Gpt54, Gpt54Mini, Gpt54Nano, Gpt55, Human, InMemoryModelContextRepository, InferenceRequest, InferenceResponse, type InferenceRunner, InputTokenDetails, ModelContext, type ModelContextRepository, ModelMessageItem, type ModelRuntime, Observer, OpenAIInferenceRunner, OpenAIResponses, OutputTokenDetails, Participant, ReasoningItem, SemanticEvent, type StreamingRuntime, SystemMessageItem, TokenUsage, type Tool, UserMessageItem };