llmist 2.3.0 → 2.5.0

package/dist/index.d.cts

@@ -1,7 +1,7 @@
 import { ZodType, ZodTypeAny } from 'zod';
 export { z } from 'zod';
-import { A as AgentHooks, q as ModelRegistry, I as IConversationManager, b as LLMMessage, S as StreamEvent, T as TokenUsage, G as GadgetRegistry, r as LLMist, a as LLMStreamChunk, C as CompactionStrategy, R as ResolvedCompactionConfig, s as CompactionContext, t as CompactionResult, u as CompactionConfig, v as CompactionEvent, w as CompactionStats, H as HintTemplate, E as ExecutionContext, x as GadgetExecuteReturn, y as GadgetExample, B as BaseGadget, P as ParsedGadgetCall, z as GadgetExecutionResult, D as ProviderAdapter, F as ModelDescriptor, J as ModelSpec, K as LLMGenerationOptions, L as LLMStream } from './mock-stream-wRfUqXx4.cjs';
-export { Z as AfterGadgetExecutionAction, _ as AfterGadgetExecutionControllerContext, $ as AfterLLMCallAction, a0 as AfterLLMCallControllerContext, a1 as AfterLLMErrorAction, U as AgentBuilder, a2 as AgentOptions, a3 as BeforeGadgetExecutionAction, a4 as BeforeLLMCallAction, a5 as ChunkInterceptorContext, a6 as Controllers, as as CostEstimate, aM as CostReportingLLMist, an as DEFAULT_COMPACTION_CONFIG, aC as DEFAULT_HINTS, aD as DEFAULT_PROMPTS, ao as DEFAULT_SUMMARIZATION_PROMPT, V as EventHandlers, aK as GadgetClass, aN as GadgetExecuteResult, a7 as GadgetExecutionControllerContext, aL as GadgetOrClass, a8 as GadgetParameterInterceptorContext, a9 as GadgetResultInterceptorContext, ay as HintContext, N as HistoryMessage, aa as Interceptors, ab as LLMCallControllerContext, ac as LLMErrorControllerContext, ar as LLMMessageBuilder, aq as LLMRole, ap as LLMistOptions, ad as MessageInterceptorContext, al as MessageTurn, d as MockBuilder, f as MockManager, j as MockMatcher, k as MockMatcherContext, l as MockOptions, M as MockProviderAdapter, n as MockRegistration, o as MockResponse, p as MockStats, at as ModelFeatures, ax as ModelIdentifierParser, au as ModelLimits, av as ModelPricing, ae as ObserveChunkContext, am as ObserveCompactionContext, af as ObserveGadgetCompleteContext, ag as ObserveGadgetStartContext, ah as ObserveLLMCallContext, ai as ObserveLLMCompleteContext, aj as ObserveLLMErrorContext, ak as Observers, az as PromptConfig, aA as PromptContext, aB as PromptTemplate, aw as ProviderIdentifier, aH as QuickOptions, aO as TextOnlyAction, aP as TextOnlyContext, aQ as TextOnlyCustomHandler, aR as TextOnlyGadgetConfig, aS as TextOnlyHandler, aT as TextOnlyStrategy, O as TrailingMessage, Q as TrailingMessageContext, W as collectEvents, X as collectText, aI as complete, c as createMockAdapter, e as createMockClient, h as createMockStream, i as createTextMockStream, g as getMockManager, m as mockLLM, aE as resolveHintTemplate, aF as resolvePromptTemplate, aG as resolveRulesTemplate, Y as runWithHandlers, aJ as stream } from './mock-stream-wRfUqXx4.cjs';
+import { A as AgentHooks, s as ModelRegistry, I as IConversationManager, b as LLMMessage, t as MessageContent, S as StreamEvent, T as TokenUsage, G as GadgetRegistry, u as LLMist, a as LLMStreamChunk, C as CompactionStrategy, R as ResolvedCompactionConfig, v as CompactionContext, w as CompactionResult, x as CompactionConfig, y as CompactionEvent, z as CompactionStats, H as HintTemplate, E as ExecutionContext, D as GadgetExecuteReturn, F as GadgetExample, B as BaseGadget, P as ParsedGadgetCall, J as GadgetExecutionResult, K as ProviderAdapter, N as ModelDescriptor, O as ModelSpec, Q as LLMGenerationOptions, L as LLMStream, U as ImageModelSpec, V as ImageGenerationOptions, W as ImageGenerationResult, X as SpeechModelSpec, Y as SpeechGenerationOptions, Z as SpeechGenerationResult } from './mock-stream-ga4KIiwX.cjs';
+export { a6 as AfterGadgetExecutionAction, a7 as AfterGadgetExecutionControllerContext, a8 as AfterLLMCallAction, a9 as AfterLLMCallControllerContext, aa as AfterLLMErrorAction, a1 as AgentBuilder, ab as AgentOptions, az as AudioContentPart, aA as AudioMimeType, aB as AudioSource, ac as BeforeGadgetExecutionAction, ad as BeforeLLMCallAction, ae as ChunkInterceptorContext, aC as ContentPart, af as Controllers, b1 as CostEstimate, bl as CostReportingLLMist, aw as DEFAULT_COMPACTION_CONFIG, bb as DEFAULT_HINTS, bc as DEFAULT_PROMPTS, ax as DEFAULT_SUMMARIZATION_PROMPT, a2 as EventHandlers, bj as GadgetClass, bm as GadgetExecuteResult, ag as GadgetExecutionControllerContext, bk as GadgetOrClass, ah as GadgetParameterInterceptorContext, ai as GadgetResultInterceptorContext, bn as GadgetSkippedEvent, b7 as HintContext, _ as HistoryMessage, aD as ImageBase64Source, aE as ImageContentPart, aF as ImageMimeType, aG as ImageSource, aH as ImageUrlSource, aj as Interceptors, ak as LLMCallControllerContext, al as LLMErrorControllerContext, aZ as LLMMessageBuilder, aX as LLMRole, ay as LLMistOptions, am as MessageInterceptorContext, au as MessageTurn, d as MockBuilder, f as MockManager, l as MockMatcher, n as MockMatcherContext, o as MockOptions, M as MockProviderAdapter, p as MockRegistration, q as MockResponse, r as MockStats, b2 as ModelFeatures, b6 as ModelIdentifierParser, b3 as ModelLimits, b4 as ModelPricing, an as ObserveChunkContext, av as ObserveCompactionContext, ao as ObserveGadgetCompleteContext, ap as ObserveGadgetStartContext, aq as ObserveLLMCallContext, ar as ObserveLLMCompleteContext, as as ObserveLLMErrorContext, at as Observers, b8 as PromptConfig, b9 as PromptContext, ba as PromptTemplate, b5 as ProviderIdentifier, bg as QuickOptions, aI as TextContentPart, bo as TextOnlyAction, bp as TextOnlyContext, bq as TextOnlyCustomHandler, br as TextOnlyGadgetConfig, bs as TextOnlyHandler, bt as TextOnlyStrategy, $ as TrailingMessage, a0 as TrailingMessageContext, a$ as VisionAnalyzeOptions, b0 as VisionAnalyzeResult, aJ as audioFromBase64, aK as audioFromBuffer, a3 as collectEvents, a4 as collectText, bh as complete, c as createMockAdapter, e as createMockClient, h as createMockStream, i as createTextMockStream, aL as detectAudioMimeType, aM as detectImageMimeType, aY as extractText, g as getMockManager, aN as imageFromBase64, aO as imageFromBuffer, aP as imageFromUrl, aQ as isAudioPart, aR as isDataUrl, aS as isImagePart, aT as isTextPart, m as mockLLM, a_ as normalizeContent, aU as parseDataUrl, bd as resolveHintTemplate, be as resolvePromptTemplate, bf as resolveRulesTemplate, a5 as runWithHandlers, bi as stream, aV as text, aW as toBase64 } from './mock-stream-ga4KIiwX.cjs';
 import { Logger, ILogObj } from 'tslog';
 import { MessageCreateParamsStreaming, MessageStreamEvent } from '@anthropic-ai/sdk/resources/messages';
 import OpenAI from 'openai';
@@ -740,7 +740,7 @@ declare class ConversationManager implements IConversationManager {
     private readonly endPrefix?;
     private readonly argPrefix?;
     constructor(baseMessages: LLMMessage[], initialMessages: LLMMessage[], options?: ConversationManagerOptions);
-    addUserMessage(content: string): void;
+    addUserMessage(content: MessageContent): void;
     addAssistantMessage(content: string): void;
     addGadgetCall(gadgetName: string, parameters: Record<string, unknown>, result: string): void;
     getMessages(): LLMMessage[];
@@ -843,6 +843,12 @@ declare class StreamProcessor {
     private accumulatedText;
     private shouldStopExecution;
     private observerFailureCount;
+    /** Gadgets waiting for their dependencies to complete */
+    private pendingGadgets;
+    /** Completed gadget results, keyed by invocation ID */
+    private completedResults;
+    /** Invocation IDs of gadgets that have failed (error or skipped due to dependency) */
+    private failedInvocations;
     constructor(options: StreamProcessorOptions);
     /**
      * Process an LLM stream and return structured results.
@@ -857,9 +863,28 @@ declare class StreamProcessor {
      */
     private processTextEvent;
     /**
-     * Process a gadget call through the full lifecycle.
+     * Process a gadget call through the full lifecycle, handling dependencies.
+     *
+     * Gadgets without dependencies (or with all dependencies satisfied) execute immediately.
+     * Gadgets with unsatisfied dependencies are queued for later execution.
+     * After each execution, pending gadgets are checked to see if they can now run.
      */
     private processGadgetCall;
+    /**
+     * Execute a gadget through the full hook lifecycle.
+     * This is the core execution logic, extracted from processGadgetCall.
+     */
+    private executeGadgetWithHooks;
+    /**
+     * Handle a gadget that cannot execute because a dependency failed.
+     * Calls the onDependencySkipped controller to allow customization.
+     */
+    private handleFailedDependency;
+    /**
+     * Process pending gadgets whose dependencies are now satisfied.
+     * Executes ready gadgets in parallel and continues until no more can be triggered.
+     */
+    private processPendingGadgets;
     /**
      * Safely execute an observer, catching and logging any errors.
      * Observers are non-critical, so errors are logged but don't crash the system.
@@ -1682,8 +1707,14 @@ declare class StreamParser {
     constructor(options?: StreamParserOptions);
     private takeTextUntil;
     /**
-     * Parse gadget name, handling both old format (name:invocationId) and new format (just name).
-     * For new format, generates a unique invocation ID.
+     * Parse gadget name with optional invocation ID and dependencies.
+     *
+     * Supported formats:
+     * - `GadgetName` - Auto-generate ID, no dependencies
+     * - `GadgetName:my_id` - Explicit ID, no dependencies
+     * - `GadgetName:my_id:dep1,dep2` - Explicit ID with dependencies
+     *
+     * Dependencies must be comma-separated invocation IDs.
      */
     private parseGadgetName;
     /**
@@ -2072,7 +2103,20 @@ declare class AnthropicMessagesProvider extends BaseProviderAdapter {
     readonly providerId: "anthropic";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    supportsImageGeneration(_modelId: string): boolean;
+    generateImage(): Promise<never>;
+    supportsSpeechGeneration(_modelId: string): boolean;
+    generateSpeech(): Promise<never>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): MessageCreateParamsStreaming;
+    /**
+     * Convert llmist content to Anthropic's content block format.
+     * Handles text, images (base64 only), and applies cache_control.
+     */
+    private convertToAnthropicContent;
+    /**
+     * Convert an image content part to Anthropic's image block format.
+     */
+    private convertImagePart;
     protected executeStreamRequest(payload: MessageCreateParamsStreaming, signal?: AbortSignal): Promise<AsyncIterable<MessageStreamEvent>>;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     /**
@@ -2104,6 +2148,24 @@ declare function createAnthropicProviderFromEnv(): AnthropicMessagesProvider | n
 
 declare function discoverProviderAdapters(): ProviderAdapter[];
 
+/**
+ * Gemini content part - can be text or inline data (images/audio).
+ */
+type GeminiPart = {
+    text: string;
+} | {
+    inlineData: {
+        mimeType: string;
+        data: string;
+    };
+};
+/**
+ * Gemini content with role and multimodal parts.
+ */
+type GeminiContent = {
+    role: string;
+    parts: GeminiPart[];
+};
 type GeminiChunk = {
     text?: () => string;
     candidates?: Array<{
@@ -2125,24 +2187,20 @@ declare class GeminiGenerativeProvider extends BaseProviderAdapter {
     readonly providerId: "gemini";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    getImageModelSpecs(): ImageModelSpec[];
+    supportsImageGeneration(modelId: string): boolean;
+    generateImage(options: ImageGenerationOptions): Promise<ImageGenerationResult>;
+    getSpeechModelSpecs(): SpeechModelSpec[];
+    supportsSpeechGeneration(modelId: string): boolean;
+    generateSpeech(options: SpeechGenerationOptions): Promise<SpeechGenerationResult>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, _spec: ModelSpec | undefined, messages: LLMMessage[]): {
         model: string;
-        contents: Array<{
-            role: string;
-            parts: Array<{
-                text: string;
-            }>;
-        }>;
+        contents: GeminiContent[];
         config: Record<string, unknown>;
     };
     protected executeStreamRequest(payload: {
         model: string;
-        contents: Array<{
-            role: string;
-            parts: Array<{
-                text: string;
-            }>;
-        }>;
+        contents: GeminiContent[];
         config: Record<string, unknown>;
     }, signal?: AbortSignal): Promise<AsyncIterable<GeminiChunk>>;
     /**
@@ -2160,8 +2218,16 @@ declare class GeminiGenerativeProvider extends BaseProviderAdapter {
      * - Model: "Understood."
      */
     private convertMessagesToContents;
+    /**
+     * Merge consecutive messages with the same role (required by Gemini).
+     * Handles multimodal content by converting to Gemini's part format.
+     */
     private mergeConsecutiveMessages;
-    private convertContentsForNewSDK;
+    /**
+     * Convert llmist content to Gemini's part format.
+     * Handles text, images, and audio (Gemini supports all three).
+     */
+    private convertToGeminiParts;
     private buildGenerationConfig;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     private extractText;
@@ -2198,7 +2264,30 @@ declare class OpenAIChatProvider extends BaseProviderAdapter {
     readonly providerId: "openai";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    getImageModelSpecs(): ImageModelSpec[];
+    supportsImageGeneration(modelId: string): boolean;
+    generateImage(options: ImageGenerationOptions): Promise<ImageGenerationResult>;
+    getSpeechModelSpecs(): SpeechModelSpec[];
+    supportsSpeechGeneration(modelId: string): boolean;
+    generateSpeech(options: SpeechGenerationOptions): Promise<SpeechGenerationResult>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): Parameters<OpenAI["chat"]["completions"]["create"]>[0];
+    /**
+     * Convert an LLMMessage to OpenAI's ChatCompletionMessageParam.
+     * Handles role-specific content type requirements:
+     * - system/assistant: string content only
+     * - user: string or multimodal array content
+     */
+    private convertToOpenAIMessage;
+    /**
+     * Convert llmist content to OpenAI's content format.
+     * Optimizes by returning string for text-only content, array for multimodal.
+     */
+    private convertToOpenAIContent;
+    /**
+     * Convert an image content part to OpenAI's image_url format.
+     * Supports both URLs and base64 data URLs.
+     */
+    private convertImagePart;
     protected executeStreamRequest(payload: Parameters<OpenAI["chat"]["completions"]["create"]>[0], signal?: AbortSignal): Promise<AsyncIterable<ChatCompletionChunk>>;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     /**
@@ -2228,4 +2317,4 @@ declare class OpenAIChatProvider extends BaseProviderAdapter {
 }
 declare function createOpenAIProviderFromEnv(): OpenAIChatProvider | null;
 
-export { AbortError, AgentHooks, AnthropicMessagesProvider, BaseGadget, BreakLoopException, CompactionConfig, CompactionContext, CompactionEvent, CompactionManager, CompactionResult, CompactionStats, CompactionStrategy, ConversationManager, type CreateGadgetConfig, ExecutionContext, Gadget, type GadgetConfig, GadgetExample, GadgetExecuteReturn, GadgetExecutionResult, GadgetExecutor, GadgetOutputStore, GadgetRegistry, GeminiGenerativeProvider, HintTemplate, type HintsConfig, HookPresets, HumanInputException, HybridStrategy, IConversationManager, type IterationHintOptions, LLMGenerationOptions, LLMMessage, LLMStream, LLMStreamChunk, LLMist, type LoggerOptions, type LoggingOptions, MODEL_ALIASES, ModelDescriptor, ModelRegistry, ModelSpec, OpenAIChatProvider, type ParallelGadgetHintOptions, ParsedGadgetCall, ProviderAdapter, ResolvedCompactionConfig, SlidingWindowStrategy, type StoredOutput, StreamEvent, StreamParser, type StreamProcessingResult, StreamProcessor, type StreamProcessorOptions, SummarizationStrategy, TokenUsage, type ValidationIssue, type ValidationResult, createAnthropicProviderFromEnv, createGadget, createGadgetOutputViewer, createGeminiProviderFromEnv, createHints, createLogger, createOpenAIProviderFromEnv, defaultLogger, discoverProviderAdapters, getModelId, getProvider, hasProviderPrefix, iterationProgressHint, parallelGadgetHint, resolveModel, validateAndApplyDefaults, validateGadgetParams };
+export { AbortError, AgentHooks, AnthropicMessagesProvider, BaseGadget, BreakLoopException, CompactionConfig, CompactionContext, CompactionEvent, CompactionManager, CompactionResult, CompactionStats, CompactionStrategy, ConversationManager, type CreateGadgetConfig, ExecutionContext, Gadget, type GadgetConfig, GadgetExample, GadgetExecuteReturn, GadgetExecutionResult, GadgetExecutor, GadgetOutputStore, GadgetRegistry, GeminiGenerativeProvider, HintTemplate, type HintsConfig, HookPresets, HumanInputException, HybridStrategy, IConversationManager, type IterationHintOptions, LLMGenerationOptions, LLMMessage, LLMStream, LLMStreamChunk, LLMist, type LoggerOptions, type LoggingOptions, MODEL_ALIASES, MessageContent, ModelDescriptor, ModelRegistry, ModelSpec, OpenAIChatProvider, type ParallelGadgetHintOptions, ParsedGadgetCall, ProviderAdapter, ResolvedCompactionConfig, SlidingWindowStrategy, type StoredOutput, StreamEvent, StreamParser, type StreamProcessingResult, StreamProcessor, type StreamProcessorOptions, SummarizationStrategy, TokenUsage, type ValidationIssue, type ValidationResult, createAnthropicProviderFromEnv, createGadget, createGadgetOutputViewer, createGeminiProviderFromEnv, createHints, createLogger, createOpenAIProviderFromEnv, defaultLogger, discoverProviderAdapters, getModelId, getProvider, hasProviderPrefix, iterationProgressHint, parallelGadgetHint, resolveModel, validateAndApplyDefaults, validateGadgetParams };

package/dist/index.d.ts

@@ -1,7 +1,7 @@
 import { ZodType, ZodTypeAny } from 'zod';
 export { z } from 'zod';
-import { A as AgentHooks, q as ModelRegistry, I as IConversationManager, b as LLMMessage, S as StreamEvent, T as TokenUsage, G as GadgetRegistry, r as LLMist, a as LLMStreamChunk, C as CompactionStrategy, R as ResolvedCompactionConfig, s as CompactionContext, t as CompactionResult, u as CompactionConfig, v as CompactionEvent, w as CompactionStats, H as HintTemplate, E as ExecutionContext, x as GadgetExecuteReturn, y as GadgetExample, B as BaseGadget, P as ParsedGadgetCall, z as GadgetExecutionResult, D as ProviderAdapter, F as ModelDescriptor, J as ModelSpec, K as LLMGenerationOptions, L as LLMStream } from './mock-stream-wRfUqXx4.js';
-export { Z as AfterGadgetExecutionAction, _ as AfterGadgetExecutionControllerContext, $ as AfterLLMCallAction, a0 as AfterLLMCallControllerContext, a1 as AfterLLMErrorAction, U as AgentBuilder, a2 as AgentOptions, a3 as BeforeGadgetExecutionAction, a4 as BeforeLLMCallAction, a5 as ChunkInterceptorContext, a6 as Controllers, as as CostEstimate, aM as CostReportingLLMist, an as DEFAULT_COMPACTION_CONFIG, aC as DEFAULT_HINTS, aD as DEFAULT_PROMPTS, ao as DEFAULT_SUMMARIZATION_PROMPT, V as EventHandlers, aK as GadgetClass, aN as GadgetExecuteResult, a7 as GadgetExecutionControllerContext, aL as GadgetOrClass, a8 as GadgetParameterInterceptorContext, a9 as GadgetResultInterceptorContext, ay as HintContext, N as HistoryMessage, aa as Interceptors, ab as LLMCallControllerContext, ac as LLMErrorControllerContext, ar as LLMMessageBuilder, aq as LLMRole, ap as LLMistOptions, ad as MessageInterceptorContext, al as MessageTurn, d as MockBuilder, f as MockManager, j as MockMatcher, k as MockMatcherContext, l as MockOptions, M as MockProviderAdapter, n as MockRegistration, o as MockResponse, p as MockStats, at as ModelFeatures, ax as ModelIdentifierParser, au as ModelLimits, av as ModelPricing, ae as ObserveChunkContext, am as ObserveCompactionContext, af as ObserveGadgetCompleteContext, ag as ObserveGadgetStartContext, ah as ObserveLLMCallContext, ai as ObserveLLMCompleteContext, aj as ObserveLLMErrorContext, ak as Observers, az as PromptConfig, aA as PromptContext, aB as PromptTemplate, aw as ProviderIdentifier, aH as QuickOptions, aO as TextOnlyAction, aP as TextOnlyContext, aQ as TextOnlyCustomHandler, aR as TextOnlyGadgetConfig, aS as TextOnlyHandler, aT as TextOnlyStrategy, O as TrailingMessage, Q as TrailingMessageContext, W as collectEvents, X as collectText, aI as complete, c as createMockAdapter, e as createMockClient, h as createMockStream, i as createTextMockStream, g as getMockManager, m as mockLLM, aE as resolveHintTemplate, aF as resolvePromptTemplate, aG as resolveRulesTemplate, Y as runWithHandlers, aJ as stream } from './mock-stream-wRfUqXx4.js';
+import { A as AgentHooks, s as ModelRegistry, I as IConversationManager, b as LLMMessage, t as MessageContent, S as StreamEvent, T as TokenUsage, G as GadgetRegistry, u as LLMist, a as LLMStreamChunk, C as CompactionStrategy, R as ResolvedCompactionConfig, v as CompactionContext, w as CompactionResult, x as CompactionConfig, y as CompactionEvent, z as CompactionStats, H as HintTemplate, E as ExecutionContext, D as GadgetExecuteReturn, F as GadgetExample, B as BaseGadget, P as ParsedGadgetCall, J as GadgetExecutionResult, K as ProviderAdapter, N as ModelDescriptor, O as ModelSpec, Q as LLMGenerationOptions, L as LLMStream, U as ImageModelSpec, V as ImageGenerationOptions, W as ImageGenerationResult, X as SpeechModelSpec, Y as SpeechGenerationOptions, Z as SpeechGenerationResult } from './mock-stream-ga4KIiwX.js';
+export { a6 as AfterGadgetExecutionAction, a7 as AfterGadgetExecutionControllerContext, a8 as AfterLLMCallAction, a9 as AfterLLMCallControllerContext, aa as AfterLLMErrorAction, a1 as AgentBuilder, ab as AgentOptions, az as AudioContentPart, aA as AudioMimeType, aB as AudioSource, ac as BeforeGadgetExecutionAction, ad as BeforeLLMCallAction, ae as ChunkInterceptorContext, aC as ContentPart, af as Controllers, b1 as CostEstimate, bl as CostReportingLLMist, aw as DEFAULT_COMPACTION_CONFIG, bb as DEFAULT_HINTS, bc as DEFAULT_PROMPTS, ax as DEFAULT_SUMMARIZATION_PROMPT, a2 as EventHandlers, bj as GadgetClass, bm as GadgetExecuteResult, ag as GadgetExecutionControllerContext, bk as GadgetOrClass, ah as GadgetParameterInterceptorContext, ai as GadgetResultInterceptorContext, bn as GadgetSkippedEvent, b7 as HintContext, _ as HistoryMessage, aD as ImageBase64Source, aE as ImageContentPart, aF as ImageMimeType, aG as ImageSource, aH as ImageUrlSource, aj as Interceptors, ak as LLMCallControllerContext, al as LLMErrorControllerContext, aZ as LLMMessageBuilder, aX as LLMRole, ay as LLMistOptions, am as MessageInterceptorContext, au as MessageTurn, d as MockBuilder, f as MockManager, l as MockMatcher, n as MockMatcherContext, o as MockOptions, M as MockProviderAdapter, p as MockRegistration, q as MockResponse, r as MockStats, b2 as ModelFeatures, b6 as ModelIdentifierParser, b3 as ModelLimits, b4 as ModelPricing, an as ObserveChunkContext, av as ObserveCompactionContext, ao as ObserveGadgetCompleteContext, ap as ObserveGadgetStartContext, aq as ObserveLLMCallContext, ar as ObserveLLMCompleteContext, as as ObserveLLMErrorContext, at as Observers, b8 as PromptConfig, b9 as PromptContext, ba as PromptTemplate, b5 as ProviderIdentifier, bg as QuickOptions, aI as TextContentPart, bo as TextOnlyAction, bp as TextOnlyContext, bq as TextOnlyCustomHandler, br as TextOnlyGadgetConfig, bs as TextOnlyHandler, bt as TextOnlyStrategy, $ as TrailingMessage, a0 as TrailingMessageContext, a$ as VisionAnalyzeOptions, b0 as VisionAnalyzeResult, aJ as audioFromBase64, aK as audioFromBuffer, a3 as collectEvents, a4 as collectText, bh as complete, c as createMockAdapter, e as createMockClient, h as createMockStream, i as createTextMockStream, aL as detectAudioMimeType, aM as detectImageMimeType, aY as extractText, g as getMockManager, aN as imageFromBase64, aO as imageFromBuffer, aP as imageFromUrl, aQ as isAudioPart, aR as isDataUrl, aS as isImagePart, aT as isTextPart, m as mockLLM, a_ as normalizeContent, aU as parseDataUrl, bd as resolveHintTemplate, be as resolvePromptTemplate, bf as resolveRulesTemplate, a5 as runWithHandlers, bi as stream, aV as text, aW as toBase64 } from './mock-stream-ga4KIiwX.js';
 import { Logger, ILogObj } from 'tslog';
 import { MessageCreateParamsStreaming, MessageStreamEvent } from '@anthropic-ai/sdk/resources/messages';
 import OpenAI from 'openai';
@@ -740,7 +740,7 @@ declare class ConversationManager implements IConversationManager {
     private readonly endPrefix?;
     private readonly argPrefix?;
     constructor(baseMessages: LLMMessage[], initialMessages: LLMMessage[], options?: ConversationManagerOptions);
-    addUserMessage(content: string): void;
+    addUserMessage(content: MessageContent): void;
     addAssistantMessage(content: string): void;
     addGadgetCall(gadgetName: string, parameters: Record<string, unknown>, result: string): void;
     getMessages(): LLMMessage[];
@@ -843,6 +843,12 @@ declare class StreamProcessor {
     private accumulatedText;
     private shouldStopExecution;
     private observerFailureCount;
+    /** Gadgets waiting for their dependencies to complete */
+    private pendingGadgets;
+    /** Completed gadget results, keyed by invocation ID */
+    private completedResults;
+    /** Invocation IDs of gadgets that have failed (error or skipped due to dependency) */
+    private failedInvocations;
     constructor(options: StreamProcessorOptions);
     /**
      * Process an LLM stream and return structured results.
@@ -857,9 +863,28 @@ declare class StreamProcessor {
      */
     private processTextEvent;
     /**
-     * Process a gadget call through the full lifecycle.
+     * Process a gadget call through the full lifecycle, handling dependencies.
+     *
+     * Gadgets without dependencies (or with all dependencies satisfied) execute immediately.
+     * Gadgets with unsatisfied dependencies are queued for later execution.
+     * After each execution, pending gadgets are checked to see if they can now run.
      */
     private processGadgetCall;
+    /**
+     * Execute a gadget through the full hook lifecycle.
+     * This is the core execution logic, extracted from processGadgetCall.
+     */
+    private executeGadgetWithHooks;
+    /**
+     * Handle a gadget that cannot execute because a dependency failed.
+     * Calls the onDependencySkipped controller to allow customization.
+     */
+    private handleFailedDependency;
+    /**
+     * Process pending gadgets whose dependencies are now satisfied.
+     * Executes ready gadgets in parallel and continues until no more can be triggered.
+     */
+    private processPendingGadgets;
     /**
      * Safely execute an observer, catching and logging any errors.
      * Observers are non-critical, so errors are logged but don't crash the system.
@@ -1682,8 +1707,14 @@ declare class StreamParser {
     constructor(options?: StreamParserOptions);
     private takeTextUntil;
     /**
-     * Parse gadget name, handling both old format (name:invocationId) and new format (just name).
-     * For new format, generates a unique invocation ID.
+     * Parse gadget name with optional invocation ID and dependencies.
+     *
+     * Supported formats:
+     * - `GadgetName` - Auto-generate ID, no dependencies
+     * - `GadgetName:my_id` - Explicit ID, no dependencies
+     * - `GadgetName:my_id:dep1,dep2` - Explicit ID with dependencies
+     *
+     * Dependencies must be comma-separated invocation IDs.
      */
     private parseGadgetName;
     /**
@@ -2072,7 +2103,20 @@ declare class AnthropicMessagesProvider extends BaseProviderAdapter {
     readonly providerId: "anthropic";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    supportsImageGeneration(_modelId: string): boolean;
+    generateImage(): Promise<never>;
+    supportsSpeechGeneration(_modelId: string): boolean;
+    generateSpeech(): Promise<never>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): MessageCreateParamsStreaming;
+    /**
+     * Convert llmist content to Anthropic's content block format.
+     * Handles text, images (base64 only), and applies cache_control.
+     */
+    private convertToAnthropicContent;
+    /**
+     * Convert an image content part to Anthropic's image block format.
+     */
+    private convertImagePart;
     protected executeStreamRequest(payload: MessageCreateParamsStreaming, signal?: AbortSignal): Promise<AsyncIterable<MessageStreamEvent>>;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     /**
@@ -2104,6 +2148,24 @@ declare function createAnthropicProviderFromEnv(): AnthropicMessagesProvider | n
 
 declare function discoverProviderAdapters(): ProviderAdapter[];
 
+/**
+ * Gemini content part - can be text or inline data (images/audio).
+ */
+type GeminiPart = {
+    text: string;
+} | {
+    inlineData: {
+        mimeType: string;
+        data: string;
+    };
+};
+/**
+ * Gemini content with role and multimodal parts.
+ */
+type GeminiContent = {
+    role: string;
+    parts: GeminiPart[];
+};
 type GeminiChunk = {
     text?: () => string;
     candidates?: Array<{
@@ -2125,24 +2187,20 @@ declare class GeminiGenerativeProvider extends BaseProviderAdapter {
     readonly providerId: "gemini";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    getImageModelSpecs(): ImageModelSpec[];
+    supportsImageGeneration(modelId: string): boolean;
+    generateImage(options: ImageGenerationOptions): Promise<ImageGenerationResult>;
+    getSpeechModelSpecs(): SpeechModelSpec[];
+    supportsSpeechGeneration(modelId: string): boolean;
+    generateSpeech(options: SpeechGenerationOptions): Promise<SpeechGenerationResult>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, _spec: ModelSpec | undefined, messages: LLMMessage[]): {
         model: string;
-        contents: Array<{
-            role: string;
-            parts: Array<{
-                text: string;
-            }>;
-        }>;
+        contents: GeminiContent[];
         config: Record<string, unknown>;
     };
     protected executeStreamRequest(payload: {
         model: string;
-        contents: Array<{
-            role: string;
-            parts: Array<{
-                text: string;
-            }>;
-        }>;
+        contents: GeminiContent[];
         config: Record<string, unknown>;
     }, signal?: AbortSignal): Promise<AsyncIterable<GeminiChunk>>;
     /**
@@ -2160,8 +2218,16 @@ declare class GeminiGenerativeProvider extends BaseProviderAdapter {
      * - Model: "Understood."
      */
     private convertMessagesToContents;
+    /**
+     * Merge consecutive messages with the same role (required by Gemini).
+     * Handles multimodal content by converting to Gemini's part format.
+     */
     private mergeConsecutiveMessages;
-    private convertContentsForNewSDK;
+    /**
+     * Convert llmist content to Gemini's part format.
+     * Handles text, images, and audio (Gemini supports all three).
+     */
+    private convertToGeminiParts;
     private buildGenerationConfig;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     private extractText;
@@ -2198,7 +2264,30 @@ declare class OpenAIChatProvider extends BaseProviderAdapter {
     readonly providerId: "openai";
     supports(descriptor: ModelDescriptor): boolean;
     getModelSpecs(): ModelSpec[];
+    getImageModelSpecs(): ImageModelSpec[];
+    supportsImageGeneration(modelId: string): boolean;
+    generateImage(options: ImageGenerationOptions): Promise<ImageGenerationResult>;
+    getSpeechModelSpecs(): SpeechModelSpec[];
+    supportsSpeechGeneration(modelId: string): boolean;
+    generateSpeech(options: SpeechGenerationOptions): Promise<SpeechGenerationResult>;
     protected buildRequestPayload(options: LLMGenerationOptions, descriptor: ModelDescriptor, spec: ModelSpec | undefined, messages: LLMMessage[]): Parameters<OpenAI["chat"]["completions"]["create"]>[0];
+    /**
+     * Convert an LLMMessage to OpenAI's ChatCompletionMessageParam.
+     * Handles role-specific content type requirements:
+     * - system/assistant: string content only
+     * - user: string or multimodal array content
+     */
+    private convertToOpenAIMessage;
+    /**
+     * Convert llmist content to OpenAI's content format.
+     * Optimizes by returning string for text-only content, array for multimodal.
+     */
+    private convertToOpenAIContent;
+    /**
+     * Convert an image content part to OpenAI's image_url format.
+     * Supports both URLs and base64 data URLs.
+     */
+    private convertImagePart;
     protected executeStreamRequest(payload: Parameters<OpenAI["chat"]["completions"]["create"]>[0], signal?: AbortSignal): Promise<AsyncIterable<ChatCompletionChunk>>;
     protected wrapStream(iterable: AsyncIterable<unknown>): LLMStream;
     /**
@@ -2228,4 +2317,4 @@ declare class OpenAIChatProvider extends BaseProviderAdapter {
 }
 declare function createOpenAIProviderFromEnv(): OpenAIChatProvider | null;
 
-export { AbortError, AgentHooks, AnthropicMessagesProvider, BaseGadget, BreakLoopException, CompactionConfig, CompactionContext, CompactionEvent, CompactionManager, CompactionResult, CompactionStats, CompactionStrategy, ConversationManager, type CreateGadgetConfig, ExecutionContext, Gadget, type GadgetConfig, GadgetExample, GadgetExecuteReturn, GadgetExecutionResult, GadgetExecutor, GadgetOutputStore, GadgetRegistry, GeminiGenerativeProvider, HintTemplate, type HintsConfig, HookPresets, HumanInputException, HybridStrategy, IConversationManager, type IterationHintOptions, LLMGenerationOptions, LLMMessage, LLMStream, LLMStreamChunk, LLMist, type LoggerOptions, type LoggingOptions, MODEL_ALIASES, ModelDescriptor, ModelRegistry, ModelSpec, OpenAIChatProvider, type ParallelGadgetHintOptions, ParsedGadgetCall, ProviderAdapter, ResolvedCompactionConfig, SlidingWindowStrategy, type StoredOutput, StreamEvent, StreamParser, type StreamProcessingResult, StreamProcessor, type StreamProcessorOptions, SummarizationStrategy, TokenUsage, type ValidationIssue, type ValidationResult, createAnthropicProviderFromEnv, createGadget, createGadgetOutputViewer, createGeminiProviderFromEnv, createHints, createLogger, createOpenAIProviderFromEnv, defaultLogger, discoverProviderAdapters, getModelId, getProvider, hasProviderPrefix, iterationProgressHint, parallelGadgetHint, resolveModel, validateAndApplyDefaults, validateGadgetParams };
+export { AbortError, AgentHooks, AnthropicMessagesProvider, BaseGadget, BreakLoopException, CompactionConfig, CompactionContext, CompactionEvent, CompactionManager, CompactionResult, CompactionStats, CompactionStrategy, ConversationManager, type CreateGadgetConfig, ExecutionContext, Gadget, type GadgetConfig, GadgetExample, GadgetExecuteReturn, GadgetExecutionResult, GadgetExecutor, GadgetOutputStore, GadgetRegistry, GeminiGenerativeProvider, HintTemplate, type HintsConfig, HookPresets, HumanInputException, HybridStrategy, IConversationManager, type IterationHintOptions, LLMGenerationOptions, LLMMessage, LLMStream, LLMStreamChunk, LLMist, type LoggerOptions, type LoggingOptions, MODEL_ALIASES, MessageContent, ModelDescriptor, ModelRegistry, ModelSpec, OpenAIChatProvider, type ParallelGadgetHintOptions, ParsedGadgetCall, ProviderAdapter, ResolvedCompactionConfig, SlidingWindowStrategy, type StoredOutput, StreamEvent, StreamParser, type StreamProcessingResult, StreamProcessor, type StreamProcessorOptions, SummarizationStrategy, TokenUsage, type ValidationIssue, type ValidationResult, createAnthropicProviderFromEnv, createGadget, createGadgetOutputViewer, createGeminiProviderFromEnv, createHints, createLogger, createOpenAIProviderFromEnv, defaultLogger, discoverProviderAdapters, getModelId, getProvider, hasProviderPrefix, iterationProgressHint, parallelGadgetHint, resolveModel, validateAndApplyDefaults, validateGadgetParams };

package/dist/index.js

@@ -5,7 +5,7 @@ import {
   iterationProgressHint,
   parallelGadgetHint,
   z
-} from "./chunk-ZDNV7DDO.js";
+} from "./chunk-IHSZUAYN.js";
 import {
   AbortError,
   AgentBuilder,
@@ -37,6 +37,8 @@ import {
   StreamParser,
   StreamProcessor,
   SummarizationStrategy,
+  audioFromBase64,
+  audioFromBuffer,
   collectEvents,
   collectText,
   complete,
@@ -51,21 +53,35 @@ import {
   createOpenAIProviderFromEnv,
   createTextMockStream,
   defaultLogger,
+  detectAudioMimeType,
+  detectImageMimeType,
   discoverProviderAdapters,
+  extractText,
   getMockManager,
   getModelId,
   getProvider,
   hasProviderPrefix,
+  imageFromBase64,
+  imageFromBuffer,
+  imageFromUrl,
+  isAudioPart,
+  isDataUrl,
+  isImagePart,
+  isTextPart,
   mockLLM,
+  normalizeContent,
+  parseDataUrl,
   resolveHintTemplate,
   resolveModel,
   resolvePromptTemplate,
   resolveRulesTemplate,
   runWithHandlers,
   stream,
+  text,
+  toBase64,
   validateAndApplyDefaults,
   validateGadgetParams
-} from "./chunk-GANXNBIZ.js";
+} from "./chunk-YHS2DYXP.js";
 export {
   AbortError,
   AgentBuilder,
@@ -99,6 +115,8 @@ export {
   StreamParser,
   StreamProcessor,
   SummarizationStrategy,
+  audioFromBase64,
+  audioFromBuffer,
   collectEvents,
   collectText,
   complete,
@@ -114,20 +132,34 @@ export {
   createOpenAIProviderFromEnv,
   createTextMockStream,
   defaultLogger,
+  detectAudioMimeType,
+  detectImageMimeType,
   discoverProviderAdapters,
+  extractText,
   getMockManager,
   getModelId,
   getProvider,
   hasProviderPrefix,
+  imageFromBase64,
+  imageFromBuffer,
+  imageFromUrl,
+  isAudioPart,
+  isDataUrl,
+  isImagePart,
+  isTextPart,
   iterationProgressHint,
   mockLLM,
+  normalizeContent,
   parallelGadgetHint,
+  parseDataUrl,
   resolveHintTemplate,
   resolveModel,
   resolvePromptTemplate,
   resolveRulesTemplate,
   runWithHandlers,
   stream,
+  text,
+  toBase64,
   validateAndApplyDefaults,
   validateGadgetParams,
   z