@ai-sdk/workflow 1.0.0-beta.3 → 1.0.0-beta.30

package/src/workflow-agent.ts

@@ -7,21 +7,24 @@ import type {
   SharedV4ProviderOptions,
 } from '@ai-sdk/provider';
 import {
-  type Experimental_LanguageModelStreamPart as ModelCallStreamPart,
+  Output,
+  experimental_filterActiveTools as filterActiveTools,
   type FinishReason,
   type LanguageModelResponseMetadata,
   type LanguageModelUsage,
+  type Experimental_LanguageModelStreamPart as ModelCallStreamPart,
   type ModelMessage,
-  Output,
   type StepResult,
   type StopCondition,
-  type StreamTextOnStepFinishCallback,
+  type GenerateTextOnStepFinishCallback,
   type SystemModelMessage,
+  type ActiveTools,
   type ToolCallRepairFunction,
   type ToolChoice,
   type ToolSet,
   type UIMessage,
-  LanguageModel,
+  type LanguageModel,
+  type Prompt,
 } from 'ai';
 import {
   convertToLanguageModelPrompt,
@@ -30,19 +33,18 @@ import {
   standardizePrompt,
 } from 'ai/internal';
 import { streamTextIterator } from './stream-text-iterator.js';
-import type { CompatibleLanguageModel } from './types.js';
 
 // Re-export for consumers
 export type { CompatibleLanguageModel } from './types.js';
 
 /**
  * Callback function to be called after each step completes.
- * Alias for the AI SDK's StreamTextOnStepFinishCallback, using
+ * Alias for the AI SDK's GenerateTextOnStepFinishCallback, using
  * WorkflowAgent-consistent naming.
  */
 export type WorkflowAgentOnStepFinishCallback<
   TTools extends ToolSet = ToolSet,
-> = StreamTextOnStepFinishCallback<TTools, any>;
+> = GenerateTextOnStepFinishCallback<TTools, any>;
 
 /**
  * Infer the type of the tools of a workflow agent.
@@ -92,7 +94,7 @@ export type ProviderOptions = SharedV4ProviderOptions;
 /**
  * Telemetry settings for observability.
  */
-export interface TelemetrySettings {
+export interface TelemetryOptions {
   /**
    * Enable or disable telemetry. Defaults to true.
    */
@@ -339,7 +341,11 @@ export interface PrepareCallOptions<
   tools: TTools;
   instructions?: string | SystemModelMessage | Array<SystemModelMessage>;
   toolChoice?: ToolChoice<TTools>;
-  experimental_telemetry?: TelemetrySettings;
+  telemetry?: TelemetryOptions;
+  /**
+   * @deprecated Use `telemetry` instead. This alias will be removed in a future major release.
+   */
+  experimental_telemetry?: TelemetryOptions;
   experimental_context?: unknown;
   messages: ModelMessage[];
 }
@@ -367,6 +373,11 @@ export type PrepareCallCallback<TTools extends ToolSet = ToolSet> = (
 export interface WorkflowAgentOptions<
   TTools extends ToolSet = ToolSet,
 > extends GenerationSettings {
+  /**
+   * The id of the agent.
+   */
+  id?: string;
+
   /**
    * The model provider to use for the agent.
    *
@@ -400,9 +411,16 @@ export interface WorkflowAgentOptions<
   toolChoice?: ToolChoice<TTools>;
 
   /**
-   * Optional telemetry configuration (experimental).
+   * Optional telemetry configuration.
    */
-  experimental_telemetry?: TelemetrySettings;
+  telemetry?: TelemetryOptions;
+
+  /**
+   * Optional telemetry configuration.
+   *
+   * @deprecated Use `telemetry` instead. This alias will be removed in a future major release.
+   */
+  experimental_telemetry?: TelemetryOptions;
 
   /**
    * Default context that is passed into tool execution for every stream call on this agent.
@@ -413,6 +431,46 @@ export interface WorkflowAgentOptions<
    */
   experimental_context?: unknown;
 
+  /**
+   * Default stop condition for the agent loop. When the condition is an array,
+   * any of the conditions can be met to stop the generation.
+   *
+   * Per-stream `stopWhen` values passed to `stream()` override this default.
+   */
+  stopWhen?:
+    | StopCondition<NoInfer<ToolSet>, any>
+    | Array<StopCondition<NoInfer<ToolSet>, any>>;
+
+  /**
+   * Default set of active tools that limits which tools the model can call,
+   * without changing the tool call and result types in the result.
+   *
+   * Per-stream `activeTools` values passed to `stream()` override this default.
+   */
+  activeTools?: ActiveTools<NoInfer<TTools>>;
+
+  /**
+   * Default output specification for structured outputs.
+   * Use `Output.object({ schema })` for structured output or `Output.text()` for text output.
+   *
+   * Per-stream `output` values passed to `stream()` override this default.
+   */
+  output?: OutputSpecification<any, any>;
+
+  /**
+   * Default function that attempts to repair a tool call that failed to parse.
+   *
+   * Per-stream `experimental_repairToolCall` values passed to `stream()` override this default.
+   */
+  experimental_repairToolCall?: ToolCallRepairFunction<TTools>;
+
+  /**
+   * Default custom download function to use for URLs.
+   *
+   * Per-stream `experimental_download` values passed to `stream()` override this default.
+   */
+  experimental_download?: DownloadFunction;
+
   /**
    * Default callback function called before each step in the agent loop.
    * Use this to modify settings, manage context, or inject messages dynamically
@@ -445,12 +503,12 @@ export interface WorkflowAgentOptions<
   /**
    * Callback called before a tool's execute function runs.
    */
-  experimental_onToolCallStart?: WorkflowAgentOnToolCallStartCallback;
+  experimental_onToolExecutionStart?: WorkflowAgentOnToolExecutionStartCallback;
 
   /**
    * Callback called after a tool execution completes.
    */
-  experimental_onToolCallFinish?: WorkflowAgentOnToolCallFinishCallback;
+  experimental_onToolExecutionEnd?: WorkflowAgentOnToolExecutionEndCallback;
 
   /**
    * Prepare the parameters for the stream call.
@@ -535,250 +593,305 @@ export type WorkflowAgentOnStartCallback = (event: {
 /**
  * Callback that is called before each step (LLM call) begins.
  */
-export type WorkflowAgentOnStepStartCallback = (event: {
-  /** The current step number (0-based) */
-  readonly stepNumber: number;
-  /** The model being used for this step */
-  readonly model: LanguageModel;
-  /** The messages being sent for this step */
-  readonly messages: ModelMessage[];
-}) => PromiseLike<void> | void;
+export type WorkflowAgentOnStepStartCallback<TTools extends ToolSet = ToolSet> =
+  (event: {
+    /** The current step number (0-based) */
+    readonly stepNumber: number;
+    /** The model being used for this step */
+    readonly model: LanguageModel;
+    /** The messages being sent for this step */
+    readonly messages: ModelMessage[];
+    /** Results from all previously finished steps */
+    readonly steps: ReadonlyArray<StepResult<TTools, any>>;
+  }) => PromiseLike<void> | void;
 
 /**
  * Callback that is called before a tool's execute function runs.
  */
-export type WorkflowAgentOnToolCallStartCallback = (event: {
+export type WorkflowAgentOnToolExecutionStartCallback = (event: {
   /** The tool call being executed */
   readonly toolCall: ToolCall;
+  /** The current step number (0-based) */
+  readonly stepNumber: number;
 }) => PromiseLike<void> | void;
 
 /**
  * Callback that is called after a tool execution completes.
+ * Uses a discriminated union pattern: check `success` to determine
+ * whether `output` or `error` is available.
  */
-export type WorkflowAgentOnToolCallFinishCallback = (event: {
-  /** The tool call that was executed */
-  readonly toolCall: ToolCall;
-  /** The tool result (undefined if execution failed) */
-  readonly result?: unknown;
-  /** The error if execution failed */
-  readonly error?: unknown;
-}) => PromiseLike<void> | void;
+export type WorkflowAgentOnToolExecutionEndCallback = (
+  event:
+    | {
+        /** The tool call that was executed */
+        readonly toolCall: ToolCall;
+        /** The current step number (0-based) */
+        readonly stepNumber: number;
+        /** Execution time in milliseconds */
+        readonly durationMs: number;
+        /** Whether the tool call succeeded */
+        readonly success: true;
+        /** The tool result */
+        readonly output: unknown;
+        readonly error?: never;
+      }
+    | {
+        /** The tool call that was executed */
+        readonly toolCall: ToolCall;
+        /** The current step number (0-based) */
+        readonly stepNumber: number;
+        /** Execution time in milliseconds */
+        readonly durationMs: number;
+        /** Whether the tool call succeeded */
+        readonly success: false;
+        /** The error that occurred */
+        readonly error: unknown;
+        readonly output?: never;
+      },
+) => PromiseLike<void> | void;
 
 /**
  * Options for the {@link WorkflowAgent.stream} method.
  */
-export interface WorkflowAgentStreamOptions<
+export type WorkflowAgentStreamOptions<
   TTools extends ToolSet = ToolSet,
   OUTPUT = never,
   PARTIAL_OUTPUT = never,
-> extends Partial<GenerationSettings> {
-  /**
-   * The conversation messages to process. Should follow the AI SDK's ModelMessage format.
-   */
-  messages: ModelMessage[];
-
-  /**
-   * Optional system prompt override. If provided, overrides the system prompt from the constructor.
-   */
-  system?: string;
+> = Partial<GenerationSettings> &
+  (
+    | {
+        /**
+         * A prompt. It can be either a text prompt or a list of messages.
+         *
+         * You can either use `prompt` or `messages` but not both.
+         */
+        prompt: string | Array<ModelMessage>;
+
+        /**
+         * A list of messages.
+         *
+         * You can either use `prompt` or `messages` but not both.
+         */
+        messages?: never;
+      }
+    | {
+        /**
+         * The conversation messages to process. Should follow the AI SDK's ModelMessage format.
+         *
+         * You can either use `prompt` or `messages` but not both.
+         */
+        messages: Array<ModelMessage>;
+
+        /**
+         * A prompt. It can be either a text prompt or a list of messages.
+         *
+         * You can either use `prompt` or `messages` but not both.
+         */
+        prompt?: never;
+      }
+  ) & {
+    /**
+     * Optional system prompt override. If provided, overrides the system prompt from the constructor.
+     */
+    system?: string;
 
-  /**
-   * A WritableStream that receives raw LanguageModelV4StreamPart chunks in real-time
-   * as the model generates them. This enables streaming to the client without
-   * coupling WorkflowAgent to UIMessageChunk format.
-   *
-   * Convert to UIMessageChunks at the response boundary using
-   * `createUIMessageChunkTransform()` from `@ai-sdk/workflow`.
-   *
-   * @example
-   * ```typescript
-   * // In the workflow:
-   * await agent.stream({
-   *   messages,
-   *   writable: getWritable<ModelCallStreamPart>(),
-   * });
-   *
-   * // In the route handler:
-   * return createUIMessageStreamResponse({
-   *   stream: run.readable.pipeThrough(createModelCallToUIChunkTransform()),
-   * });
-   * ```
-   */
-  writable?: WritableStream<ModelCallStreamPart<ToolSet>>;
+    /**
+     * A WritableStream that receives raw LanguageModelV4StreamPart chunks in real-time
+     * as the model generates them. This enables streaming to the client without
+     * coupling WorkflowAgent to UIMessageChunk format.
+     *
+     * Convert to UIMessageChunks at the response boundary using
+     * `createUIMessageChunkTransform()` from `@ai-sdk/workflow`.
+     *
+     * @example
+     * ```typescript
+     * // In the workflow:
+     * await agent.stream({
+     *   messages,
+     *   writable: getWritable<ModelCallStreamPart>(),
+     * });
+     *
+     * // In the route handler:
+     * return createUIMessageStreamResponse({
+     *   stream: run.readable.pipeThrough(createModelCallToUIChunkTransform()),
+     * });
+     * ```
+     */
+    writable?: WritableStream<ModelCallStreamPart<ToolSet>>;
 
-  /**
-   * Condition for stopping the generation when there are tool results in the last step.
-   * When the condition is an array, any of the conditions can be met to stop the generation.
-   */
-  stopWhen?:
-    | StopCondition<NoInfer<ToolSet>, any>
-    | Array<StopCondition<NoInfer<ToolSet>, any>>;
+    /**
+     * Condition for stopping the generation when there are tool results in the last step.
+     * When the condition is an array, any of the conditions can be met to stop the generation.
+     */
+    stopWhen?:
+      | StopCondition<NoInfer<ToolSet>, any>
+      | Array<StopCondition<NoInfer<ToolSet>, any>>;
 
-  /**
-   * Maximum number of sequential LLM calls (steps), e.g. when you use tool calls.
-   * A maximum number can be set to prevent infinite loops in the case of misconfigured tools.
-   * By default, it's unlimited (the agent loops until completion).
-   */
-  maxSteps?: number;
+    /**
+     * The tool choice strategy. Default: 'auto'.
+     * Overrides the toolChoice from the constructor if provided.
+     */
+    toolChoice?: ToolChoice<TTools>;
 
-  /**
-   * The tool choice strategy. Default: 'auto'.
-   * Overrides the toolChoice from the constructor if provided.
-   */
-  toolChoice?: ToolChoice<TTools>;
+    /**
+     * Limits the tools that are available for the model to call without
+     * changing the tool call and result types in the result.
+     */
+    activeTools?: ActiveTools<NoInfer<TTools>>;
 
-  /**
-   * Limits the tools that are available for the model to call without
-   * changing the tool call and result types in the result.
-   */
-  activeTools?: Array<keyof NoInfer<TTools>>;
+    /**
+     * Optional telemetry configuration.
+     */
+    telemetry?: TelemetryOptions;
 
-  /**
-   * Optional telemetry configuration (experimental).
-   */
-  experimental_telemetry?: TelemetrySettings;
+    /**
+     * Optional telemetry configuration.
+     *
+     * @deprecated Use `telemetry` instead. This alias will be removed in a future major release.
+     */
+    experimental_telemetry?: TelemetryOptions;
 
-  /**
-   * Context that is passed into tool execution.
-   * Experimental (can break in patch releases).
-   * @default undefined
-   */
-  experimental_context?: unknown;
+    /**
+     * Context that is passed into tool execution.
+     * Experimental (can break in patch releases).
+     * @default undefined
+     */
+    experimental_context?: unknown;
 
-  /**
-   * Optional specification for parsing structured outputs from the LLM response.
-   * Use `Output.object({ schema })` for structured output or `Output.text()` for text output.
-   *
-   * @example
-   * ```typescript
-   * import { Output } from '@workflow/ai';
-   * import { z } from 'zod';
-   *
-   * const result = await agent.stream({
-   *   messages: [...],
-   *   writable: getWritable(),
-   *   output: Output.object({
-   *     schema: z.object({
-   *       sentiment: z.enum(['positive', 'negative', 'neutral']),
-   *       confidence: z.number(),
-   *     }),
-   *   }),
-   * });
-   *
-   * console.log(result.output); // { sentiment: 'positive', confidence: 0.95 }
-   * ```
-   */
-  output?: OutputSpecification<OUTPUT, PARTIAL_OUTPUT>;
+    /**
+     * Optional specification for parsing structured outputs from the LLM response.
+     * Use `Output.object({ schema })` for structured output or `Output.text()` for text output.
+     *
+     * @example
+     * ```typescript
+     * import { Output } from '@workflow/ai';
+     * import { z } from 'zod';
+     *
+     * const result = await agent.stream({
+     *   messages: [...],
+     *   writable: getWritable(),
+     *   output: Output.object({
+     *     schema: z.object({
+     *       sentiment: z.enum(['positive', 'negative', 'neutral']),
+     *       confidence: z.number(),
+     *     }),
+     *   }),
+     * });
+     *
+     * console.log(result.output); // { sentiment: 'positive', confidence: 0.95 }
+     * ```
+     */
+    output?: OutputSpecification<OUTPUT, PARTIAL_OUTPUT>;
 
-  /**
-   * Whether to include raw chunks from the provider in the stream.
-   * When enabled, you will receive raw chunks with type 'raw' that contain the unprocessed data from the provider.
-   * This allows access to cutting-edge provider features not yet wrapped by the AI SDK.
-   * Defaults to false.
-   */
-  includeRawChunks?: boolean;
+    /**
+     * Whether to include raw chunks from the provider in the stream.
+     * When enabled, you will receive raw chunks with type 'raw' that contain the unprocessed data from the provider.
+     * This allows access to cutting-edge provider features not yet wrapped by the AI SDK.
+     * Defaults to false.
+     */
+    includeRawChunks?: boolean;
 
-  /**
-   * A function that attempts to repair a tool call that failed to parse.
-   */
-  experimental_repairToolCall?: ToolCallRepairFunction<TTools>;
+    /**
+     * A function that attempts to repair a tool call that failed to parse.
+     */
+    experimental_repairToolCall?: ToolCallRepairFunction<TTools>;
 
-  /**
-   * Optional stream transformations.
-   * They are applied in the order they are provided.
-   * The stream transformations must maintain the stream structure for streamText to work correctly.
-   */
-  experimental_transform?:
-    | StreamTextTransform<TTools>
-    | Array<StreamTextTransform<TTools>>;
+    /**
+     * Optional stream transformations.
+     * They are applied in the order they are provided.
+     * The stream transformations must maintain the stream structure for streamText to work correctly.
+     */
+    experimental_transform?:
+      | StreamTextTransform<TTools>
+      | Array<StreamTextTransform<TTools>>;
 
-  /**
-   * Custom download function to use for URLs.
-   * By default, files are downloaded if the model does not support the URL for the given media type.
-   */
-  experimental_download?: DownloadFunction;
+    /**
+     * Custom download function to use for URLs.
+     * By default, files are downloaded if the model does not support the URL for the given media type.
+     */
+    experimental_download?: DownloadFunction;
 
-  /**
-   * Callback function to be called after each step completes.
-   */
-  onStepFinish?: WorkflowAgentOnStepFinishCallback<TTools>;
+    /**
+     * Callback function to be called after each step completes.
+     */
+    onStepFinish?: WorkflowAgentOnStepFinishCallback<TTools>;
 
-  /**
-   * Callback that is invoked when an error occurs during streaming.
-   * You can use it to log errors.
-   */
-  onError?: WorkflowAgentOnErrorCallback;
+    /**
+     * Callback that is invoked when an error occurs during streaming.
+     * You can use it to log errors.
+     */
+    onError?: WorkflowAgentOnErrorCallback;
 
-  /**
-   * Callback that is called when the LLM response and all request tool executions
-   * (for tools that have an `execute` function) are finished.
-   */
-  onFinish?: WorkflowAgentOnFinishCallback<TTools, OUTPUT>;
+    /**
+     * Callback that is called when the LLM response and all request tool executions
+     * (for tools that have an `execute` function) are finished.
+     */
+    onFinish?: WorkflowAgentOnFinishCallback<TTools, OUTPUT>;
 
-  /**
-   * Callback that is called when the operation is aborted.
-   */
-  onAbort?: WorkflowAgentOnAbortCallback<TTools>;
+    /**
+     * Callback that is called when the operation is aborted.
+     */
+    onAbort?: WorkflowAgentOnAbortCallback<TTools>;
 
-  /**
-   * Callback called when the agent starts streaming, before any LLM calls.
-   */
-  experimental_onStart?: WorkflowAgentOnStartCallback;
+    /**
+     * Callback called when the agent starts streaming, before any LLM calls.
+     */
+    experimental_onStart?: WorkflowAgentOnStartCallback;
 
-  /**
-   * Callback called before each step (LLM call) begins.
-   */
-  experimental_onStepStart?: WorkflowAgentOnStepStartCallback;
+    /**
+     * Callback called before each step (LLM call) begins.
+     */
+    experimental_onStepStart?: WorkflowAgentOnStepStartCallback;
 
-  /**
-   * Callback called before a tool's execute function runs.
-   */
-  experimental_onToolCallStart?: WorkflowAgentOnToolCallStartCallback;
+    /**
+     * Callback called before a tool's execute function runs.
+     */
+    experimental_onToolExecutionStart?: WorkflowAgentOnToolExecutionStartCallback;
 
-  /**
-   * Callback called after a tool execution completes.
-   */
-  experimental_onToolCallFinish?: WorkflowAgentOnToolCallFinishCallback;
+    /**
+     * Callback called after a tool execution completes.
+     */
+    experimental_onToolExecutionEnd?: WorkflowAgentOnToolExecutionEndCallback;
 
-  /**
-   * Callback function called before each step in the agent loop.
-   * Use this to modify settings, manage context, or inject messages dynamically.
-   *
-   * @example
-   * ```typescript
-   * prepareStep: async ({ messages, stepNumber }) => {
-   *   // Inject messages from a queue
-   *   const queuedMessages = await getQueuedMessages();
-   *   if (queuedMessages.length > 0) {
-   *     return {
-   *       messages: [...messages, ...queuedMessages],
-   *     };
-   *   }
-   *   return {};
-   * }
-   * ```
-   */
-  prepareStep?: PrepareStepCallback<TTools>;
+    /**
+     * Callback function called before each step in the agent loop.
+     * Use this to modify settings, manage context, or inject messages dynamically.
+     *
+     * @example
+     * ```typescript
+     * prepareStep: async ({ messages, stepNumber }) => {
+     *   // Inject messages from a queue
+     *   const queuedMessages = await getQueuedMessages();
+     *   if (queuedMessages.length > 0) {
+     *     return {
+     *       messages: [...messages, ...queuedMessages],
+     *     };
+     *   }
+     *   return {};
+     * }
+     * ```
+     */
+    prepareStep?: PrepareStepCallback<TTools>;
 
-  /**
-   * Timeout in milliseconds for the stream operation.
-   * When specified, creates an AbortSignal that will abort the operation after the given time.
-   * If both `timeout` and `abortSignal` are provided, whichever triggers first will abort.
-   */
-  timeout?: number;
+    /**
+     * Timeout in milliseconds for the stream operation.
+     * When specified, creates an AbortSignal that will abort the operation after the given time.
+     * If both `timeout` and `abortSignal` are provided, whichever triggers first will abort.
+     */
+    timeout?: number;
 
-  /**
-   * Whether to send a 'finish' chunk to the writable stream when streaming completes.
-   * @default true
-   */
-  sendFinish?: boolean;
+    /**
+     * Whether to send a 'finish' chunk to the writable stream when streaming completes.
+     * @default true
+     */
+    sendFinish?: boolean;
 
-  /**
-   * Whether to prevent the writable stream from being closed after streaming completes.
-   * @default false
-   */
-  preventClose?: boolean;
-}
+    /**
+     * Whether to prevent the writable stream from being closed after streaming completes.
+     * @default false
+     */
+    preventClose?: boolean;
+  };
 
 /**
  * A tool call made by the model. Matches the AI SDK's tool call shape.
@@ -889,6 +1002,11 @@ export interface WorkflowAgentStreamResult<
  * ```
  */
 export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
+  /**
+   * The id of the agent.
+   */
+  public readonly id: string | undefined;
+
   private model: LanguageModel;
   /**
    * The tool set configured for this agent.
@@ -900,32 +1018,47 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
     | Array<SystemModelMessage>;
   private generationSettings: GenerationSettings;
   private toolChoice?: ToolChoice<TBaseTools>;
-  private telemetry?: TelemetrySettings;
+  private telemetry?: TelemetryOptions;
   private experimentalContext: unknown;
+  private stopWhen?:
+    | StopCondition<ToolSet, any>
+    | Array<StopCondition<ToolSet, any>>;
+  private activeTools?: ActiveTools<TBaseTools>;
+  private output?: OutputSpecification<any, any>;
+  private experimentalRepairToolCall?: ToolCallRepairFunction<TBaseTools>;
+  private experimentalDownload?: DownloadFunction;
   private prepareStep?: PrepareStepCallback<TBaseTools>;
   private constructorOnStepFinish?: WorkflowAgentOnStepFinishCallback<ToolSet>;
   private constructorOnFinish?: WorkflowAgentOnFinishCallback<ToolSet>;
   private constructorOnStart?: WorkflowAgentOnStartCallback;
   private constructorOnStepStart?: WorkflowAgentOnStepStartCallback;
-  private constructorOnToolCallStart?: WorkflowAgentOnToolCallStartCallback;
-  private constructorOnToolCallFinish?: WorkflowAgentOnToolCallFinishCallback;
+  private constructorOnToolExecutionStart?: WorkflowAgentOnToolExecutionStartCallback;
+  private constructorOnToolExecutionEnd?: WorkflowAgentOnToolExecutionEndCallback;
   private prepareCall?: PrepareCallCallback<TBaseTools>;
 
   constructor(options: WorkflowAgentOptions<TBaseTools>) {
+    this.id = options.id;
     this.model = options.model;
     this.tools = (options.tools ?? {}) as TBaseTools;
     // `instructions` takes precedence over deprecated `system`
     this.instructions = options.instructions ?? options.system;
     this.toolChoice = options.toolChoice;
-    this.telemetry = options.experimental_telemetry;
+    this.telemetry = options.telemetry ?? options.experimental_telemetry;
     this.experimentalContext = options.experimental_context;
+    this.stopWhen = options.stopWhen;
+    this.activeTools = options.activeTools;
+    this.output = options.output;
+    this.experimentalRepairToolCall = options.experimental_repairToolCall;
+    this.experimentalDownload = options.experimental_download;
     this.prepareStep = options.prepareStep;
     this.constructorOnStepFinish = options.onStepFinish;
     this.constructorOnFinish = options.onFinish;
     this.constructorOnStart = options.experimental_onStart;
     this.constructorOnStepStart = options.experimental_onStepStart;
-    this.constructorOnToolCallStart = options.experimental_onToolCallStart;
-    this.constructorOnToolCallFinish = options.experimental_onToolCallFinish;
+    this.constructorOnToolExecutionStart =
+      options.experimental_onToolExecutionStart;
+    this.constructorOnToolExecutionEnd =
+      options.experimental_onToolExecutionEnd;
     this.prepareCall = options.prepareCall;
 
     // Extract generation settings
@@ -959,13 +1092,23 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
     // Call prepareCall to transform parameters before the agent loop
     let effectiveModel: LanguageModel = this.model;
     let effectiveInstructions = options.system ?? this.instructions;
-    let effectiveMessages = options.messages;
+    let effectivePrompt: string | Array<ModelMessage> | undefined =
+      options.prompt;
+    let effectiveMessages: Array<ModelMessage> | undefined = options.messages;
     let effectiveGenerationSettings = { ...this.generationSettings };
     let effectiveExperimentalContext =
       options.experimental_context ?? this.experimentalContext;
     let effectiveToolChoiceFromPrepare = options.toolChoice ?? this.toolChoice;
     let effectiveTelemetryFromPrepare =
-      options.experimental_telemetry ?? this.telemetry;
+      options.telemetry ?? options.experimental_telemetry ?? this.telemetry;
+
+    // Resolve messages for prepareCall: use messages directly, or convert prompt
+    const resolvedMessagesForPrepareCall: ModelMessage[] =
+      effectiveMessages ??
+      (typeof effectivePrompt === 'string'
+        ? [{ role: 'user' as const, content: effectivePrompt }]
+        : (effectivePrompt as ModelMessage[])) ??
+      [];
 
     if (this.prepareCall) {
       const prepared = await this.prepareCall({
@@ -973,24 +1116,28 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
         tools: this.tools,
         instructions: effectiveInstructions,
         toolChoice: effectiveToolChoiceFromPrepare as ToolChoice<TBaseTools>,
+        telemetry: effectiveTelemetryFromPrepare,
         experimental_telemetry: effectiveTelemetryFromPrepare,
         experimental_context: effectiveExperimentalContext,
-        messages: effectiveMessages as ModelMessage[],
+        messages: resolvedMessagesForPrepareCall,
         ...effectiveGenerationSettings,
       } as PrepareCallOptions<TBaseTools>);
 
       if (prepared.model !== undefined) effectiveModel = prepared.model;
       if (prepared.instructions !== undefined)
         effectiveInstructions = prepared.instructions;
-      if (prepared.messages !== undefined)
-        effectiveMessages =
-          prepared.messages as WorkflowAgentStreamOptions<TTools>['messages'];
+      if (prepared.messages !== undefined) {
+        effectiveMessages = prepared.messages as Array<ModelMessage>;
+        effectivePrompt = undefined; // messages from prepareCall take precedence
+      }
       if (prepared.experimental_context !== undefined)
         effectiveExperimentalContext = prepared.experimental_context;
       if (prepared.toolChoice !== undefined)
         effectiveToolChoiceFromPrepare =
           prepared.toolChoice as ToolChoice<TBaseTools>;
-      if (prepared.experimental_telemetry !== undefined)
+      if (prepared.telemetry !== undefined)
+        effectiveTelemetryFromPrepare = prepared.telemetry;
+      else if (prepared.experimental_telemetry !== undefined)
         effectiveTelemetryFromPrepare = prepared.experimental_telemetry;
       if (prepared.maxOutputTokens !== undefined)
         effectiveGenerationSettings.maxOutputTokens = prepared.maxOutputTokens;
@@ -1017,8 +1164,11 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
 
     const prompt = await standardizePrompt({
       system: effectiveInstructions,
-      messages: effectiveMessages,
-    });
+      allowSystemInMessages: true, // TODO: consider exposing this as a parameter
+      ...(effectivePrompt != null
+        ? { prompt: effectivePrompt }
+        : { messages: effectiveMessages! }),
+    } as Prompt);
 
     // Process tool approval responses before starting the agent loop.
     // This mirrors how stream-text.ts handles tool-approval-response parts:
@@ -1146,14 +1296,12 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
     const modelPrompt = await convertToLanguageModelPrompt({
       prompt,
       supportedUrls: {},
-      download: options.experimental_download,
+      download: options.experimental_download ?? this.experimentalDownload,
     });
 
     const effectiveAbortSignal = mergeAbortSignals(
       options.abortSignal ?? effectiveGenerationSettings.abortSignal,
-      options.timeout != null
-        ? AbortSignal.timeout(options.timeout)
-        : undefined,
+      options.timeout,
     );
 
     // Merge generation settings: constructor defaults < prepareCall < stream options
@@ -1210,13 +1358,13 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
       this.constructorOnStepStart,
       options.experimental_onStepStart,
     );
-    const mergedOnToolCallStart = mergeCallbacks(
-      this.constructorOnToolCallStart,
-      options.experimental_onToolCallStart,
+    const mergedOnToolExecutionStart = mergeCallbacks(
+      this.constructorOnToolExecutionStart,
+      options.experimental_onToolExecutionStart,
     );
-    const mergedOnToolCallFinish = mergeCallbacks(
-      this.constructorOnToolCallFinish,
-      options.experimental_onToolCallFinish,
+    const mergedOnToolExecutionEnd = mergeCallbacks(
+      this.constructorOnToolExecutionEnd,
+      options.experimental_onToolExecutionEnd,
     );
 
     // Determine effective tool choice
@@ -1225,10 +1373,14 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
     // Merge telemetry settings
     const effectiveTelemetry = effectiveTelemetryFromPrepare;
 
-    // Filter tools if activeTools is specified
+    // Filter tools if activeTools is specified (stream-level overrides constructor default)
+    const effectiveActiveTools = options.activeTools ?? this.activeTools;
     const effectiveTools =
-      options.activeTools && options.activeTools.length > 0
-        ? filterTools(this.tools, options.activeTools as string[])
+      effectiveActiveTools && effectiveActiveTools.length > 0
+        ? (filterActiveTools({
+            tools: this.tools,
+            activeTools: effectiveActiveTools,
+          }) ?? this.tools)
         : this.tools;
 
     // Initialize context
@@ -1244,69 +1396,77 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
     if (mergedOnStart) {
       await mergedOnStart({
         model: effectiveModel,
-        messages: effectiveMessages as ModelMessage[],
+        messages: prompt.messages,
       });
     }
 
-    // Helper to wrap executeTool with onToolCallStart/onToolCallFinish callbacks
+    // Helper to wrap executeTool with onToolExecutionStart/onToolExecutionEnd callbacks
     const executeToolWithCallbacks = async (
       toolCall: { toolCallId: string; toolName: string; input: unknown },
       tools: ToolSet,
       messages: LanguageModelV4Prompt,
       context?: unknown,
+      currentStepNumber: number = 0,
     ): Promise<LanguageModelV4ToolResultPart> => {
-      if (mergedOnToolCallStart) {
-        await mergedOnToolCallStart({
-          toolCall: {
-            type: 'tool-call',
-            toolCallId: toolCall.toolCallId,
-            toolName: toolCall.toolName,
-            input: toolCall.input,
-          },
+      const toolCallEvent: ToolCall = {
+        type: 'tool-call',
+        toolCallId: toolCall.toolCallId,
+        toolName: toolCall.toolName,
+        input: toolCall.input,
+      };
+
+      if (mergedOnToolExecutionStart) {
+        await mergedOnToolExecutionStart({
+          toolCall: toolCallEvent,
+          stepNumber: currentStepNumber,
         });
       }
+
+      const startTime = Date.now();
       let result: LanguageModelV4ToolResultPart;
       try {
         result = await executeTool(toolCall, tools, messages, context);
       } catch (err) {
-        if (mergedOnToolCallFinish) {
-          await mergedOnToolCallFinish({
-            toolCall: {
-              type: 'tool-call',
-              toolCallId: toolCall.toolCallId,
-              toolName: toolCall.toolName,
-              input: toolCall.input,
-            },
+        const durationMs = Date.now() - startTime;
+        if (mergedOnToolExecutionEnd) {
+          await mergedOnToolExecutionEnd({
+            toolCall: toolCallEvent,
+            stepNumber: currentStepNumber,
+            durationMs,
+            success: false,
             error: err,
           });
         }
         throw err;
       }
-      if (mergedOnToolCallFinish) {
+
+      const durationMs = Date.now() - startTime;
+      if (mergedOnToolExecutionEnd) {
         const isError =
           result.output &&
           'type' in result.output &&
           (result.output.type === 'error-text' ||
             result.output.type === 'error-json');
-        await mergedOnToolCallFinish({
-          toolCall: {
-            type: 'tool-call',
-            toolCallId: toolCall.toolCallId,
-            toolName: toolCall.toolName,
-            input: toolCall.input,
-          },
-          ...(isError
-            ? {
-                error:
-                  'value' in result.output ? result.output.value : undefined,
-              }
-            : {
-                result:
-                  result.output && 'value' in result.output
-                    ? result.output.value
-                    : undefined,
-              }),
-        });
+        if (isError) {
+          await mergedOnToolExecutionEnd({
+            toolCall: toolCallEvent,
+            stepNumber: currentStepNumber,
+            durationMs,
+            success: false,
+            error: 'value' in result.output ? result.output.value : undefined,
+          });
+        } else {
+          await mergedOnToolExecutionEnd({
+            toolCall: toolCallEvent,
+            stepNumber: currentStepNumber,
+            durationMs,
+            success: true,
+            output:
+              result.output && 'value' in result.output
+                ? result.output.value
+                : undefined,
+          });
+        }
       }
       return result;
     };
@@ -1317,7 +1477,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
         await options.onAbort({ steps });
       }
       return {
-        messages: options.messages as unknown as ModelMessage[],
+        messages: prompt.messages,
         steps,
         toolCalls: [],
         toolResults: [],
@@ -1330,8 +1490,8 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
       tools: effectiveTools as ToolSet,
       writable: options.writable,
       prompt: modelPrompt,
-      stopConditions: options.stopWhen,
-      maxSteps: options.maxSteps,
+      stopConditions: options.stopWhen ?? this.stopWhen,
+
       onStepFinish: mergedOnStepFinish,
       onStepStart: mergedOnStepStart,
       onError: options.onError,
@@ -1341,11 +1501,13 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
       generationSettings: mergedGenerationSettings,
       toolChoice: effectiveToolChoice as ToolChoice<ToolSet>,
       experimental_context: experimentalContext,
-      experimental_telemetry: effectiveTelemetry,
+      telemetry: effectiveTelemetry,
       includeRawChunks: options.includeRawChunks ?? false,
-      repairToolCall:
-        options.experimental_repairToolCall as ToolCallRepairFunction<ToolSet>,
-      responseFormat: await options.output?.responseFormat,
+      repairToolCall: (options.experimental_repairToolCall ??
+        this.experimentalRepairToolCall) as
+        | ToolCallRepairFunction<ToolSet>
+        | undefined,
+      responseFormat: await (options.output ?? this.output)?.responseFormat,
     });
 
     // Track the final conversation messages from the iterator
@@ -1372,6 +1534,8 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
           context,
           providerExecutedToolResults,
         } = result.value;
+        // Capture current step number before pushing (0-based)
+        const currentStepNumber = steps.length;
         if (step) {
           steps.push(step as unknown as StepResult<TTools, any>);
         }
@@ -1381,11 +1545,16 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
 
         // Only execute tools if there are tool calls
         if (toolCalls.length > 0) {
+          const invalidToolCalls = toolCalls.filter(tc => tc.invalid === true);
+          const validToolCalls = toolCalls.filter(tc => tc.invalid !== true);
+
           // Separate provider-executed tool calls from client-executed ones
-          const nonProviderToolCalls = toolCalls.filter(
+          const nonProviderToolCalls = validToolCalls.filter(
             tc => !tc.providerExecuted,
           );
-          const providerToolCalls = toolCalls.filter(tc => tc.providerExecuted);
+          const providerToolCalls = validToolCalls.filter(
+            tc => tc.providerExecuted,
+          );
 
           // Check which tools need approval (can be async)
           const approvalNeeded = await Promise.all(
@@ -1435,6 +1604,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
                     effectiveTools as ToolSet,
                     iterMessages,
                     experimentalContext,
+                    currentStepNumber,
                   ),
               ),
             );
@@ -1448,7 +1618,15 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
                 ),
               );
 
-            const resolvedResults = [...executableResults, ...providerResults];
+            const continuationInvalidResults = invalidToolCalls.map(
+              createInvalidToolResult,
+            );
+            const resolvedResults = [
+              ...executableResults,
+              ...providerResults,
+              ...continuationInvalidResults,
+            ];
+            const executedResults = [...executableResults, ...providerResults];
 
             const allToolCalls: ToolCall[] = toolCalls.map(tc => ({
               type: 'tool-call' as const,
@@ -1457,7 +1635,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
               input: tc.input,
             }));
 
-            const allToolResults: ToolResult[] = resolvedResults.map(r => ({
+            const allToolResults: ToolResult[] = executedResults.map(r => ({
               type: 'tool-result' as const,
               toolCallId: r.toolCallId,
               toolName: r.toolName,
@@ -1535,6 +1713,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
                   effectiveTools as ToolSet,
                   iterMessages,
                   experimentalContext,
+                  currentStepNumber,
                 ),
             ),
           );
@@ -1544,25 +1723,32 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
             providerToolCalls.map(toolCall =>
               resolveProviderToolResult(toolCall, providerExecutedToolResults),
             );
+          const continuationInvalidToolResults = invalidToolCalls.map(
+            createInvalidToolResult,
+          );
 
-          // Combine results in the original order
-          const toolResults = toolCalls.map(tc => {
+          // Combine executable/provider results in the original order,
+          // while preserving invalid tool calls as error results for the
+          // next model step without emitting them as synthetic UI success.
+          const continuationToolResults = toolCalls.flatMap(tc => {
+            const invalidResult = continuationInvalidToolResults.find(
+              r => r.toolCallId === tc.toolCallId,
+            );
+            if (invalidResult) return [invalidResult];
             const clientResult = clientToolResults.find(
               r => r.toolCallId === tc.toolCallId,
             );
-            if (clientResult) return clientResult;
+            if (clientResult) return [clientResult];
             const providerResult = providerToolResults.find(
               r => r.toolCallId === tc.toolCallId,
             );
-            if (providerResult) return providerResult;
-            // This should never happen, but return empty result as fallback
-            return {
-              type: 'tool-result' as const,
-              toolCallId: tc.toolCallId,
-              toolName: tc.toolName,
-              output: { type: 'text' as const, value: '' },
-            };
+            if (providerResult) return [providerResult];
+            return [];
           });
+          const executedToolResults = continuationToolResults.filter(
+            result =>
+              !invalidToolCalls.some(tc => tc.toolCallId === result.toolCallId),
+          );
 
           // Write tool results and step boundaries to the stream so the
           // UI can transition tool parts to output-available state and
@@ -1570,7 +1756,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
           if (options.writable) {
             await writeToolResultsWithStepBoundary(
               options.writable,
-              toolResults.map(r => ({
+              executedToolResults.map(r => ({
                 toolCallId: r.toolCallId,
                 toolName: r.toolName,
                 input: toolCalls.find(tc => tc.toolCallId === r.toolCallId)
@@ -1587,7 +1773,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
             toolName: tc.toolName,
             input: tc.input,
           }));
-          lastStepToolResults = toolResults.map(r => ({
+          lastStepToolResults = executedToolResults.map(r => ({
             type: 'tool-result' as const,
             toolCallId: r.toolCallId,
             toolName: r.toolName,
@@ -1595,7 +1781,7 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
             output: 'value' in r.output ? r.output.value : undefined,
           }));
 
-          result = await iterator.next(toolResults);
+          result = await iterator.next(continuationToolResults);
         } else {
           // Final step with no tool calls - reset tracking
           lastStepToolCalls = [];
@@ -1623,18 +1809,19 @@ export class WorkflowAgent<TBaseTools extends ToolSet = ToolSet> {
       // Don't throw yet - we want to call onFinish first
     }
 
-    // Use the final messages from the iterator, or fall back to original messages
+    // Use the final messages from the iterator, or fall back to standardized messages
     const messages = (finalMessages ??
-      options.messages) as unknown as ModelMessage[];
+      prompt.messages) as unknown as ModelMessage[];
 
-    // Parse structured output if output is specified
+    // Parse structured output if output is specified (stream-level overrides constructor default)
+    const effectiveOutput = options.output ?? this.output;
     let experimentalOutput: OUTPUT = undefined as OUTPUT;
-    if (options.output && steps.length > 0) {
+    if (effectiveOutput && steps.length > 0) {
       const lastStep = steps[steps.length - 1];
       const text = lastStep.text;
       if (text) {
         try {
-          experimentalOutput = await options.output.parseCompleteOutput(
+          experimentalOutput = await effectiveOutput.parseCompleteOutput(
             { text },
             {
               response: lastStep.response,
@@ -1832,19 +2019,6 @@ function aggregateUsage(steps: StepResult<any, any>[]): LanguageModelUsage {
   } as LanguageModelUsage;
 }
 
-function filterTools<TTools extends ToolSet>(
-  tools: TTools,
-  activeTools: string[],
-): ToolSet {
-  const filtered: ToolSet = {};
-  for (const toolName of activeTools) {
-    if (toolName in tools) {
-      filtered[toolName] = tools[toolName];
-    }
-  }
-  return filtered;
-}
-
 // Matches AI SDK's getErrorMessage from @ai-sdk/provider-utils
 function getErrorMessage(error: unknown): string {
   if (error == null) {
@@ -1909,6 +2083,22 @@ function resolveProviderToolResult(
   };
 }
 
+function createInvalidToolResult(toolCall: {
+  toolCallId: string;
+  toolName: string;
+  error?: unknown;
+}): LanguageModelV4ToolResultPart {
+  return {
+    type: 'tool-result' as const,
+    toolCallId: toolCall.toolCallId,
+    toolName: toolCall.toolName,
+    output: {
+      type: 'error-text' as const,
+      value: getErrorMessage(toolCall.error),
+    },
+  };
+}
+
 async function executeTool(
   toolCall: { toolCallId: string; toolName: string; input: unknown },
   tools: ToolSet,