qlogicagent 2.6.1 → 2.9.0

package/dist/types/llm/transport.d.ts

@@ -1,281 +0,0 @@
-/**
- * LLMTransport 鈥?abstract interface for LLM inference calls.
- *
- * Aligned with Hermes `ProviderTransport` ABC:
- *   stream(request, apiKey, signal) 鈫?AsyncGenerator<LLMChunk>
- *
- * Two concrete implementations:
- *   - OpenAI Chat Completions (covers 95% of providers)
- *   - Anthropic Messages API
- */
-import type { ChatMessage, ToolDefinition } from "../protocol/wire/index.js";
-export type StructuredOutputConfig = {
-    mode: "json_object";
-} | {
-    mode: "json_schema";
-    name: string;
-    schema: Record<string, unknown>;
-    strict?: boolean;
-};
-export interface CachingConfig {
-    type: "enabled" | "disabled";
-    /** Enable prefix caching mode (搂20.3). Requires store=true and stream=false. */
-    prefix?: boolean;
-}
-export type ContextEdit = {
-    type: "clear_thinking";
-    keep?: "all" | {
-        type: "thinking_turns";
-        value: number;
-    };
-} | {
-    type: "clear_tool_uses";
-    trigger?: {
-        type: "tool_uses";
-        value: number;
-    };
-    keep?: {
-        type: "tool_uses";
-        value: number;
-    };
-    excludeTools?: string[];
-    clearToolInput?: boolean;
-};
-export interface ContextManagementConfig {
-    edits: ContextEdit[];
-}
-export interface LLMRequest {
-    model: string;
-    messages: ChatMessage[];
-    tools?: ToolDefinition[];
-    toolChoice?: "auto" | "none" | "required" | {
-        type: "function";
-        name: string;
-    };
-    temperature?: number;
-    /** Nucleus sampling: controls diversity via cumulative probability cutoff. */
-    topP?: number;
-    maxTokens?: number;
-    reasoning?: {
-        effort: "minimal" | "low" | "medium" | "high" | "xhigh";
-        /** Request encrypted original reasoning content (Volcengine 搂17.7). */
-        includeEncryptedReasoning?: boolean;
-    };
-    /** Volcengine: max builtin tool calls per turn (搂19.15). */
-    maxToolCalls?: number;
-    /**
-     * DeepSeek prefix completion: force model to continue from this prefix.
-     * Requires `/beta` endpoint; adds a trailing assistant message with `prefix: true`.
-     */
-    prefixMessage?: string;
-    /**
-     * Model requires streaming 鈥?disable non-streaming fallback in transports.
-     * When true, transports must NOT fall back to non-streaming requests on failure.
-     * Set for models like QwQ/Omni where the provider rejects non-streaming calls.
-     */
-    streamRequired?: boolean;
-    /**
-     * Disable injection of provider-native builtin tools (web_search, code_interpreter)
-     * for this specific request. Allows session-level control over GLM/Kimi builtin tools.
-     */
-    disableBuiltinTools?: boolean;
-    /**
-     * Volcengine builtin tools to inject (web_search, image_process, knowledge_search).
-     * Each entry specifies a tool type and optional config.
-     * These are platform-executed tools requiring beta headers.
-     */
-    builtinTools?: Array<{
-        type: "builtin_web_search" | "builtin_image_process" | "builtin_knowledge_search" | "builtin_doubao_app";
-        config?: Record<string, unknown>;
-    }>;
-    /**
-     * Server-side context continuation via response chain (搂5).
-     * When set, the server automatically includes previous context,
-     * so messages[] only needs to contain the NEW user message.
-     */
-    previousResponseId?: string;
-    /**
-     * Control server-side storage of this request's input/output (搂5.1).
-     * Default: true (server stores for 3 days).
-     */
-    store?: boolean;
-    /** Expiration time for stored response (Unix seconds, max 7 days from now) */
-    storeExpireAt?: number;
-    /**
-     * Per-turn system instruction augmentation (搂8).
-     * Temporarily overlays persona or adds constraints for this turn only.
-     * NOTE: Incompatible with caching 鈥?do not use both together.
-     */
-    instructions?: string;
-    /**
-     * Structured output format (搂16).
-     * Forces model to produce JSON conforming to the specified schema.
-     */
-    structuredOutput?: StructuredOutputConfig;
-    /**
-     * Caching configuration (搂20).
-     * Controls prefix/session caching behavior.
-     * NOTE: Incompatible with instructions, json_schema, and builtin tools.
-     */
-    caching?: CachingConfig;
-    /**
-     * Context management edits (搂21, beta).
-     * Server-side trimming of historical thinking chains and tool call traces.
-     */
-    contextManagement?: ContextManagementConfig;
-    /**
-     * Gemini explicit cache reference (gemini-ProviderMax 搂8).
-     * Passes a pre-created cache name (e.g. "cachedContents/abc123") to
-     * generateContent so the server uses cached tokens instead of re-processing.
-     * Create caches via GeminiCacheAPI.createCache() first.
-     */
-    cachedContent?: string;
-    /**
-     * Predicted output for speculative decoding (openai-ProviderMax 搂11).
-     * When editing code, pass the existing content so the model can diff efficiently.
-     * Reduces latency by 3-5x when prediction matches. Falls back when it doesn't.
-     * Works with OpenAI GPT-5.x models via Responses API and Chat Completions.
-     */
-    prediction?: {
-        type: "content";
-        content: string;
-    };
-    /**
-     * Prompt cache bucketing key (openai-ProviderMax 搂11).
-     * Replaces the deprecated `user` field. Helps OpenAI group similar requests
-     * for higher cache hit rates.
-     */
-    promptCacheKey?: string;
-    /**
-     * Prompt cache retention policy (openai-ProviderMax 搂11).
-     * "in_memory" = default 5-10 min, "24h" = extended up to 24 hours.
-     */
-    promptCacheRetention?: "in_memory" | "24h";
-    /**
-     * Service tier for request scheduling (openai-ProviderMax 搂14).
-     * "auto" = project default, "flex" = 50% cheaper / higher latency,
-     * "priority" = guaranteed low latency.
-     */
-    serviceTier?: "auto" | "default" | "flex" | "priority";
-    /**
-     * OpenAI Responses API built-in tools (openai-ProviderMax 搂7).
-     * Platform-executed tools like web_search, file_search, code_interpreter, etc.
-     */
-    openaiBuiltinTools?: Array<{
-        type: "web_search_preview" | "file_search" | "code_interpreter" | "computer_use_preview";
-        [key: string]: unknown;
-    }>;
-    /**
-     * OpenAI Responses API conversation ID (openai-ProviderMax 搂2.1).
-     * Alternative to previous_response_id 鈥?persistent server-side conversation.
-     * Cannot be used together with previousResponseId.
-     */
-    conversationId?: string;
-    /**
-     * Disable parallel tool calling for this request.
-     * When false, the model must call tools sequentially.
-     */
-    parallelToolCalls?: boolean;
-    /**
-     * Text output verbosity hint (openai-ProviderMax 搂5).
-     * Controls how detailed the model's textual output should be.
-     */
-    textVerbosity?: "low" | "medium" | "high";
-}
-/**
- * FIM completion request 鈥?DeepSeek Beta Completions API.
- * POST /beta/v1/completions with prompt + suffix.
- * Only works with non-thinking mode.
- */
-export interface FIMRequest {
-    model: string;
-    /** Text before the cursor (prefix context) */
-    prompt: string;
-    /** Text after the cursor (suffix context) */
-    suffix?: string;
-    /** Max tokens to generate for the infill */
-    maxTokens?: number;
-    /** Sampling temperature */
-    temperature?: number;
-    /** Stop sequences */
-    stop?: string[];
-}
-export type FIMChunk = {
-    type: "delta";
-    text: string;
-} | {
-    type: "done";
-    finishReason: string;
-};
-export type LLMChunk = {
-    type: "delta";
-    text: string;
-} | {
-    type: "tool_call_delta";
-    index: number;
-    id?: string;
-    name?: string;
-    arguments: string;
-} | {
-    type: "reasoning_delta";
-    text: string;
-} | {
-    type: "reasoning_block_complete";
-    thinking: string;
-    signature: string;
-} | {
-    type: "usage";
-    promptTokens: number;
-    completionTokens: number;
-    reasoningTokens?: number;
-    cacheReadTokens?: number;
-    cacheCreationTokens?: number;
-} | {
-    type: "response_id";
-    id: string;
-} | {
-    /** Informational status from platform-executed builtin tools (web_search, image_process). */
-    type: "builtin_tool_status";
-    toolType: string;
-    event: string;
-    data?: Record<string, unknown>;
-} | {
-    /** Web search citation annotations from provider builtin tools (Volcengine web_search, Gemini grounding). */
-    type: "annotations";
-    annotations: Array<{
-        type: string;
-        url?: string;
-        title?: string;
-        [key: string]: unknown;
-    }>;
-} | {
-    type: "error";
-    message: string;
-} | {
-    type: "done";
-    finishReason: string;
-};
-export interface AccumulatedToolCall {
-    id: string;
-    name: string;
-    arguments: string;
-}
-export interface LLMTransport {
-    /**
-     * Stream an LLM inference request.
-     * apiKey is passed explicitly (from agent.turn.config, not env).
-     */
-    stream(request: LLMRequest, apiKey: string, signal?: AbortSignal): AsyncGenerator<LLMChunk>;
-    /**
-     * FIM (Fill-In-Middle) completion 鈥?optional capability.
-     * Only implemented by providers that support it (DeepSeek /beta endpoint).
-     */
-    complete?(request: FIMRequest, apiKey: string, signal?: AbortSignal): AsyncGenerator<FIMChunk>;
-}
-/**
- * Accumulate tool_call_delta chunks into complete ToolCall objects.
- * Modeled after admin-infer-proxy-client's Map<index, toolCall> accumulator.
- */
-export declare function accumulateToolCalls(accumulator: Map<number, AccumulatedToolCall>, chunk: LLMChunk & {
-    type: "tool_call_delta";
-}): void;

package/dist/types/llm/transports/anthropic-messages.d.ts

@@ -1,65 +0,0 @@
-/**
- * Anthropic Messages Transport 鈥?SSE streaming for Claude API.
- *
- * Aligned with CC (claude-code-haha) src/services/api/claude.ts:
- *   - cache_control ephemeral injection on system prompt blocks
- *   - ensureToolResultPairing() conversation repair before every request
- *   - Retry with exponential backoff on transient errors (429/529/overloaded)
- *   - Non-streaming fallback when stream errors out
- *   - 90s idle watchdog timeout for silently dropped connections
- *   - Adaptive/budget thinking with temperature omit
- *   - Cache token extraction with >0 guard (CC updateUsage parity)
- *   - signature_delta handling for thinking blocks
- */
-import type { LLMChunk, LLMRequest, LLMTransport } from "../transport.js";
-import type { ProviderQuirks } from "../provider-def.js";
-import type { FileUploadAdapter } from "../file-upload-service.js";
-export interface AnthropicTransportConfig {
-    baseUrl: string;
-    /** anthropic-version header (default "2023-06-01") */
-    apiVersion?: string;
-    /** Per-request timeout in ms (default 180_000) */
-    timeoutMs?: number;
-    /** Stream idle watchdog timeout in ms (default 90_000, CC parity) */
-    streamIdleTimeoutMs?: number;
-    /** Enable prompt caching via cache_control ephemeral (default true) */
-    enablePromptCaching?: boolean;
-    /** Max retry attempts on transient errors (default 3) */
-    maxRetries?: number;
-    /** Omit temperature when it equals 0 鈥?MiniMax rejects temperature=0 */
-    omitZeroTemperature?: boolean;
-    /** Provider-specific quirks for conditional logic (CC/altcode parity) */
-    quirks?: ProviderQuirks;
-    /** File upload adapter (Hub OSS relay for Anthropic). */
-    fileUploadAdapter?: FileUploadAdapter;
-}
-export declare class AnthropicMessagesTransport implements LLMTransport {
-    private baseUrl;
-    private apiVersion;
-    private timeoutMs;
-    private streamIdleTimeoutMs;
-    private enablePromptCaching;
-    private maxRetries;
-    private omitZeroTemperature;
-    private quirks;
-    private fileUploadAdapter?;
-    constructor(config: AnthropicTransportConfig);
-    stream(request: LLMRequest, apiKey: string, signal?: AbortSignal): AsyncGenerator<LLMChunk>;
-    /**
-     * Stream with idle watchdog timer (CC parity: 90s default).
-     * Throws if no chunks received for streamIdleTimeoutMs.
-     */
-    private streamWithWatchdog;
-    /**
-     * Non-streaming fallback (CC executeNonStreamingRequest parity).
-     * Used when streaming fails after all retries.
-     * Caps max_tokens at 64K and adjusts thinking budget accordingly.
-     */
-    private nonStreamingFallback;
-    /**
-     * Convert a non-streaming API response to LLMChunk sequence.
-     */
-    private mapNonStreamingResponse;
-    private parseSSEStream;
-    private mapEvent;
-}

package/dist/types/llm/transports/gemini-cache-api.d.ts

@@ -1,86 +0,0 @@
-/**
- * GeminiCacheAPI — Explicit Context Caching for Gemini generateContent.
- *
- * Manages named cached content resources that can be referenced in
- * generateContent requests via the `cachedContent` field.
- *
- * REST endpoints:
- *   POST   /v1beta/cachedContents          — create cache
- *   GET    /v1beta/cachedContents           — list caches
- *   GET    /v1beta/cachedContents/{name}    — get cache metadata
- *   PATCH  /v1beta/cachedContents/{name}    — update TTL
- *   DELETE /v1beta/cachedContents/{name}    — delete cache
- *
- * Minimum cacheable content: 1024 tokens (Flash) / 4096 tokens (Pro).
- * TTL default: 1 hour. Storage: ~$1.00/hour/MTok (Flash series).
- *
- * Docs: https://ai.google.dev/gemini-api/docs/caching
- */
-export interface GeminiCacheCreateOptions {
-    /** Model to use, e.g. "models/gemini-3-flash-preview" */
-    model: string;
-    /** Contents to cache (same format as generateContent contents) */
-    contents: Array<Record<string, unknown>>;
-    /** Optional system instruction to include in cache */
-    systemInstruction?: Record<string, unknown>;
-    /** Time-to-live, e.g. "300s" for 5 minutes. Default: "3600s" (1 hour) */
-    ttl?: string;
-    /** Human-readable name for identifying the cache */
-    displayName?: string;
-}
-export interface GeminiCachedContent {
-    /** Resource name, e.g. "cachedContents/abc123" */
-    name: string;
-    /** Model this cache is bound to */
-    model: string;
-    /** Display name (if set) */
-    displayName?: string;
-    /** Token usage metadata */
-    usageMetadata?: {
-        totalTokenCount?: number;
-    };
-    /** Creation time (ISO 8601) */
-    createTime?: string;
-    /** Last update time (ISO 8601) */
-    updateTime?: string;
-    /** Expiration time (ISO 8601) */
-    expireTime?: string;
-}
-export declare class GeminiCacheAPI {
-    private baseUrl;
-    private timeoutMs;
-    constructor(config: {
-        baseUrl: string;
-        timeoutMs?: number;
-    });
-    /**
-     * Create a new cached content resource.
-     * The cache name returned can be passed as `cachedContent` in generateContent.
-     */
-    createCache(options: GeminiCacheCreateOptions, apiKey: string, signal?: AbortSignal): Promise<GeminiCachedContent>;
-    /**
-     * Get metadata for a cached content resource.
-     */
-    getCache(name: string, apiKey: string, signal?: AbortSignal): Promise<GeminiCachedContent>;
-    /**
-     * List all cached content resources.
-     */
-    listCaches(apiKey: string, options?: {
-        pageSize?: number;
-        pageToken?: string;
-    }, signal?: AbortSignal): Promise<{
-        cachedContents: GeminiCachedContent[];
-        nextPageToken?: string;
-    }>;
-    /**
-     * Update a cache's TTL or expiration time.
-     */
-    updateCache(name: string, update: {
-        ttl?: string;
-        expireTime?: string;
-    }, apiKey: string, signal?: AbortSignal): Promise<GeminiCachedContent>;
-    /**
-     * Delete a cached content resource.
-     */
-    deleteCache(name: string, apiKey: string, signal?: AbortSignal): Promise<void>;
-}

package/dist/types/llm/transports/gemini-file-api.d.ts

@@ -1,90 +0,0 @@
-/**
- * GeminiFileAPI — Gemini File API for uploading and managing files.
- *
- * Files uploaded via this API can be referenced in generateContent requests
- * using `file_data: { file_uri, mime_type }` parts.
- *
- * Upload uses the resumable upload protocol (2-step):
- *   1. POST /upload/v1beta/files → get upload URL (response header)
- *   2. PUT {upload_url} with file bytes → get file info
- *
- * Other operations:
- *   GET    /v1beta/files           — list files
- *   GET    /v1beta/files/{name}    — get file metadata
- *   DELETE /v1beta/files/{name}    — delete file
- *
- * Files expire after 48 hours. Max 2GB per file, 20GB per project.
- *
- * Docs: https://ai.google.dev/gemini-api/docs/files
- */
-export interface GeminiFileInfo {
-    /** Resource name, e.g. "files/abc123" */
-    name: string;
-    /** Display name (set during upload) */
-    displayName?: string;
-    /** MIME type */
-    mimeType: string;
-    /** File size in bytes */
-    sizeBytes?: string;
-    /** File URI for use in generateContent, e.g. "https://generativelanguage.googleapis.com/v1beta/files/abc123" */
-    uri: string;
-    /** Processing state: PROCESSING | ACTIVE | FAILED */
-    state: string;
-    /** Creation time (ISO 8601) */
-    createTime?: string;
-    /** Last update time (ISO 8601) */
-    updateTime?: string;
-    /** Expiration time (ISO 8601) */
-    expirationTime?: string;
-    /** Error info if state is FAILED */
-    error?: {
-        code: number;
-        message: string;
-    };
-}
-export declare class GeminiFileAPI {
-    private baseUrl;
-    private timeoutMs;
-    constructor(config: {
-        baseUrl: string;
-        timeoutMs?: number;
-    });
-    /**
-     * Upload a file using the resumable upload protocol.
-     *
-     * Step 1: Initiate upload → get upload URL from response header
-     * Step 2: Upload bytes to that URL → get file metadata
-     *
-     * @returns GeminiFileInfo with .uri for use in generateContent
-     */
-    uploadFile(file: Blob | Buffer, apiKey: string, options?: {
-        mimeType?: string;
-        displayName?: string;
-    }, signal?: AbortSignal): Promise<GeminiFileInfo>;
-    /**
-     * Wait for a file to finish processing (state → ACTIVE).
-     * Some file types (video, audio) require server-side processing.
-     */
-    waitForProcessing(name: string, apiKey: string, options?: {
-        pollIntervalMs?: number;
-        maxWaitMs?: number;
-    }): Promise<GeminiFileInfo>;
-    /**
-     * Get metadata for an uploaded file.
-     */
-    getFile(name: string, apiKey: string, signal?: AbortSignal): Promise<GeminiFileInfo>;
-    /**
-     * List uploaded files.
-     */
-    listFiles(apiKey: string, options?: {
-        pageSize?: number;
-        pageToken?: string;
-    }, signal?: AbortSignal): Promise<{
-        files: GeminiFileInfo[];
-        nextPageToken?: string;
-    }>;
-    /**
-     * Delete an uploaded file.
-     */
-    deleteFile(name: string, apiKey: string, signal?: AbortSignal): Promise<void>;
-}

package/dist/types/llm/transports/gemini-generatecontent.d.ts

@@ -1,56 +0,0 @@
-/**
- * Gemini generateContent Transport 鈥?Native Gemini API streaming implementation.
- *
- * Targets Gemini 3 series exclusively (3.1 Pro, 3 Flash, 3.1 Flash-Lite).
- * Uses the native Gemini REST API instead of the OpenAI compatibility layer,
- * unlocking Gemini-exclusive features unavailable via the compat endpoint:
- *   - thinkingConfig (thinkingLevel 鈥?G3 native control)
- *   - Google Search / Maps Grounding
- *   - Code Execution
- *   - Safety Settings fine-grained control
- *   - Thought Signatures for multi-turn reasoning continuity
- *   - URL Context / File Search
- *   - systemInstruction top-level field
- *
- * Streaming endpoint: POST .../models/{model}:streamGenerateContent?alt=sse
- * Non-streaming:      POST .../models/{model}:generateContent
- * Auth: x-goog-api-key header
- *
- * Protocol reference: https://ai.google.dev/gemini-api/docs
- * Aligned with gemini-ProviderMax.md native protocol strategy.
- */
-import type { LLMChunk, LLMRequest, LLMTransport } from "../transport.js";
-import type { ProviderQuirks } from "../provider-def.js";
-import type { FileUploadAdapter } from "../file-upload-service.js";
-export interface GeminiGenerateContentTransportConfig {
-    /** Base URL, e.g. "https://generativelanguage.googleapis.com/v1beta" */
-    baseUrl: string;
-    /** Per-request timeout in ms (default 180_000) */
-    timeoutMs?: number;
-    /** Provider-specific quirks */
-    quirks?: ProviderQuirks;
-    /** File upload adapter for resolving local media URLs via upload instead of base64. */
-    fileUploadAdapter?: FileUploadAdapter;
-}
-export declare class GeminiGenerateContentTransport implements LLMTransport {
-    private baseUrl;
-    private timeoutMs;
-    private quirks;
-    private fileUploadAdapter?;
-    constructor(config: GeminiGenerateContentTransportConfig);
-    stream(request: LLMRequest, apiKey: string, signal?: AbortSignal): AsyncGenerator<LLMChunk>;
-    private buildRequestBody;
-    private buildTools;
-    private buildToolConfig;
-    private buildGenerationConfig;
-    private fetchAndStream;
-    /**
-     * Parse SSE stream with 90s idle watchdog (CC parity).
-     */
-    private parseSSEStreamWithWatchdog;
-    /**
-     * Process a single Gemini response chunk, yielding LLMChunk events.
-     */
-    private processResponse;
-    private nonStreamingFallback;
-}

package/dist/types/llm/transports/gemini-lyria-realtime.d.ts

@@ -1,117 +0,0 @@
-/**
- * GeminiLyriaRealtimeSession — WebSocket-based real-time streaming music generation.
- *
- * Uses Lyria RealTime (`lyria-realtime-exp`) via the Gemini Live API WebSocket.
- * Provides both a full interactive session API and a one-shot convenience method.
- *
- * Protocol:
- *   - WebSocket URL: wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent
- *   - Auth: API key as query parameter
- *   - Client → Server: setup, musicInput (weightedPrompts, musicGenerationConfig, playbackControl)
- *   - Server → Client: serverContent.audioChunks (base64 PCM s16le, 48kHz, stereo)
- *
- * Docs: https://ai.google.dev/gemini-api/docs/realtime-music-generation
- */
-export interface WeightedPrompt {
-    text: string;
-    weight: number;
-}
-export type MusicScale = "C_MAJOR_A_MINOR" | "D_FLAT_MAJOR_B_FLAT_MINOR" | "D_MAJOR_B_MINOR" | "E_FLAT_MAJOR_C_MINOR" | "E_MAJOR_D_FLAT_MINOR" | "F_MAJOR_D_MINOR" | "G_FLAT_MAJOR_E_FLAT_MINOR" | "G_MAJOR_E_MINOR" | "A_FLAT_MAJOR_F_MINOR" | "A_MAJOR_G_FLAT_MINOR" | "B_FLAT_MAJOR_G_MINOR" | "B_MAJOR_A_FLAT_MINOR" | "SCALE_UNSPECIFIED";
-export type MusicGenerationMode = "QUALITY" | "DIVERSITY" | "VOCALIZATION";
-export interface MusicGenerationConfig {
-    bpm?: number;
-    density?: number;
-    brightness?: number;
-    guidance?: number;
-    scale?: MusicScale;
-    temperature?: number;
-    topK?: number;
-    seed?: number;
-    muteBass?: boolean;
-    muteDrums?: boolean;
-    onlyBassAndDrums?: boolean;
-    musicGenerationMode?: MusicGenerationMode;
-    audioFormat?: string;
-    sampleRateHz?: number;
-}
-export interface LyriaRealtimeConfig {
-    /** Base URL (REST), e.g. "https://generativelanguage.googleapis.com/v1beta" */
-    baseUrl: string;
-}
-export interface LyriaRealtimeSessionOptions {
-    model?: string;
-    prompts: WeightedPrompt[];
-    config?: MusicGenerationConfig;
-}
-/** Audio chunk received from the server. */
-export interface AudioChunk {
-    /** Raw PCM s16le data (48kHz, stereo) */
-    data: Buffer;
-}
-/**
- * Interactive Lyria RealTime session over WebSocket.
- *
- * Usage:
- *   const session = new GeminiLyriaRealtimeSession({ baseUrl: "..." });
- *   await session.connect(apiKey, { prompts: [{ text: "jazz", weight: 1 }] });
- *   session.onAudioChunk = (chunk) => { ... };
- *   await session.play();
- *   // ... later
- *   await session.stop();
- *   session.close();
- */
-export declare class GeminiLyriaRealtimeSession {
-    private ws;
-    private wsUrl;
-    private model;
-    /** Called for each audio chunk received from the server. */
-    onAudioChunk: ((chunk: AudioChunk) => void) | null;
-    /** Called when the server reports a filtered prompt. */
-    onFilteredPrompt: ((reason: string) => void) | null;
-    /** Called on WebSocket error. */
-    onError: ((error: Error) => void) | null;
-    /** Called when the WebSocket connection closes. */
-    onClose: (() => void) | null;
-    constructor(config: LyriaRealtimeConfig);
-    /**
-     * Connect to the Lyria RealTime WebSocket and send setup + initial config.
-     */
-    connect(apiKey: string, options: LyriaRealtimeSessionOptions): Promise<void>;
-    /** Set or update weighted prompts (smooth transition). */
-    setWeightedPrompts(prompts: WeightedPrompt[]): Promise<void>;
-    /** Set or update music generation config. */
-    setMusicGenerationConfig(config: MusicGenerationConfig): Promise<void>;
-    /** Start streaming music. */
-    play(): Promise<void>;
-    /** Pause music streaming (can resume with play). */
-    pause(): Promise<void>;
-    /** Stop music streaming (terminates the current piece). */
-    stop(): Promise<void>;
-    /** Reset model context (for BPM/scale changes). */
-    resetContext(): Promise<void>;
-    /** Close the WebSocket connection. */
-    close(): void;
-    /** Whether the session is connected. */
-    get connected(): boolean;
-    private send;
-    private handleMessage;
-}
-/**
- * Generate a fixed-duration music clip using Lyria RealTime.
- *
- * Connects, plays for the specified duration (default 30s), collects all
- * audio chunks, assembles into a WAV file, and returns the file path.
- *
- * Output: 48kHz, stereo, 16-bit PCM wrapped in WAV.
- */
-export declare function generateRealtimeMusic(apiKey: string, config: LyriaRealtimeConfig, options: {
-    prompts: WeightedPrompt[];
-    durationSeconds?: number;
-    musicConfig?: MusicGenerationConfig;
-    model?: string;
-    signal?: AbortSignal;
-    onProgress?: (percent: number, status: string) => void;
-}): Promise<{
-    filePath: string;
-    durationMs: number;
-}>;