@tanstack/ai 0.59.0 → 0.61.0

package/dist/esm/activities/chat/adapter.d.ts

@@ -69,6 +69,14 @@ export interface TextAdapter<TModel extends string, TProviderOptions extends Rec
      * this is the declaration/validation surface only.
      */
     readonly requires?: ReadonlyArray<CapabilityHandle>;
+    /**
+     * Declares that this adapter can consume `{ type: 'file' }` content sources
+     * (provider Files API references). `chat()` rejects file sources in preflight
+     * for adapters that don't declare this, so an adapter written before the
+     * file arm existed fails closed instead of silently mis-mapping a reference
+     * onto its URL/data branch.
+     */
+    readonly supportsFileSources?: boolean;
     /**
      * @internal Type-only properties for inference. Not assigned at runtime.
      */
@@ -156,6 +164,7 @@ export declare abstract class BaseTextAdapter<TModel extends string, TProviderOp
     abstract readonly name: string;
     readonly model: TModel;
     readonly requires?: ReadonlyArray<CapabilityHandle>;
+    readonly supportsFileSources: boolean;
     '~types': {
         providerOptions: TProviderOptions;
         inputModalities: TInputModalities;

package/dist/esm/activities/chat/adapter.js

@@ -9,6 +9,7 @@ var BaseTextAdapter = class {
 	kind = "text";
 	model;
 	requires = void 0;
+	supportsFileSources = false;
 	config;
 	constructor(config = {}, model) {
 		this.config = config;

package/dist/esm/activities/chat/adapter.js.map

@@ -1 +1 @@
-{"version":3,"file":"adapter.js","names":[],"sources":["../../../../src/activities/chat/adapter.ts"],"sourcesContent":["import type {\n  DefaultMessageMetadataByModality,\n  JSONSchema,\n  Modality,\n  TextOptions,\n  TokenUsage,\n} from '../../types'\nimport type { AdapterYieldChunk } from '../../utilities/adapter-yield-chunk'\nimport type { CapabilityHandle } from './middleware/capabilities'\n\n/**\n * Configuration for adapter instances\n */\nexport interface TextAdapterConfig {\n  apiKey?: string\n  baseUrl?: string\n  timeout?: number\n  maxRetries?: number\n  headers?: Record<string, string>\n}\n\n/**\n * Options for structured output generation.\n *\n * The internal logger is threaded through `chatOptions.logger` (inherited from\n * `TextOptions`). Adapter implementations must call `logger.request()` before\n * SDK calls, `logger.provider()` for each chunk received, and `logger.errors()`\n * in catch blocks.\n */\nexport interface StructuredOutputOptions<TProviderOptions extends object> {\n  /** Text options for the request */\n  chatOptions: TextOptions<TProviderOptions>\n  /** JSON Schema for structured output - already converted from Zod in the ai layer */\n  outputSchema: JSONSchema\n}\n\n/**\n * Result from structured output generation\n */\nexport interface StructuredOutputResult<T = unknown> {\n  /** The parsed data conforming to the schema */\n  data: T\n  /** The raw text response from the model before parsing */\n  rawText: string\n  /** Token usage information (if provided by the adapter) */\n  usage?: TokenUsage\n}\n\n/**\n * Text adapter interface with pre-resolved generics.\n *\n * An adapter is created by a provider function: `provider('model')` → `adapter`\n * All type resolution happens at the provider call site, not in this interface.\n *\n * Generic parameters:\n * - TModel: The specific model name (e.g., 'gpt-4o')\n * - TProviderOptions: Provider-specific options for this model (already resolved)\n * - TInputModalities: Supported input modalities for this model (already resolved)\n * - TMessageMetadata: Metadata types for content parts (already resolved)\n * - TToolCapabilities: Tuple of tool-kind strings supported by this model, resolved from `supports.tools`\n * - TToolCallMetadata: Metadata type that round-trips with tool calls (e.g. Gemini's `thoughtSignature`)\n * - TSystemPromptMetadata: Provider-typed metadata accepted on each\n *   `systemPrompts[i]` entry (e.g. Anthropic `cache_control`). Defaults to\n *   `never` — adapters without per-prompt metadata reject the `metadata`\n *   field at the call site.\n */\nexport interface TextAdapter<\n  TModel extends string,\n  TProviderOptions extends Record<string, any>,\n  TInputModalities extends ReadonlyArray<Modality>,\n  TMessageMetadataByModality extends DefaultMessageMetadataByModality,\n  TToolCapabilities extends ReadonlyArray<string> = ReadonlyArray<string>,\n  TToolCallMetadata = unknown,\n  TSystemPromptMetadata = never,\n> {\n  /** Discriminator for adapter kind */\n  readonly kind: 'text'\n  /** Provider name identifier (e.g., 'openai', 'anthropic') */\n  readonly name: string\n  /** The model this adapter is configured for */\n  readonly model: TModel\n\n  /**\n   * Capabilities this adapter requires at runtime. `chat()` validates that the\n   * configured middleware provides each one. Model adapters omit this; harness\n   * adapters (e.g. a future `claudeCode()`) declare e.g. `[sandboxCapability]`.\n   * Runtime access to capabilities from inside the adapter is not yet wired —\n   * this is the declaration/validation surface only.\n   */\n  readonly requires?: ReadonlyArray<CapabilityHandle>\n\n  /**\n   * @internal Type-only properties for inference. Not assigned at runtime.\n   */\n  '~types': {\n    providerOptions: TProviderOptions\n    inputModalities: TInputModalities\n    messageMetadataByModality: TMessageMetadataByModality\n    toolCapabilities: TToolCapabilities\n    toolCallMetadata: TToolCallMetadata\n    systemPromptMetadata: TSystemPromptMetadata\n  }\n\n  /**\n   * Stream text completions from the model\n   */\n  chatStream: (\n    options: TextOptions<TProviderOptions>,\n  ) => AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Generate structured output using the provider's native structured output API.\n   * This method uses stream: false and sends the JSON schema to the provider\n   * to ensure the response conforms to the expected structure.\n   *\n   * @param options - Structured output options containing chat options and JSON schema\n   * @returns Promise with the raw data (validation is done in the chat function)\n   */\n  structuredOutput: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => Promise<StructuredOutputResult<unknown>>\n\n  /**\n   * Stream structured output using the provider's native streaming structured\n   * output API (stream + response_format json_schema in a single request).\n   *\n   * Optional — adapters without native streaming JSON omit this method and the\n   * activity layer synthesizes a stream around the non-streaming\n   * `structuredOutput` call.\n   *\n   * Implementations must emit standard AG-UI lifecycle events (RUN_STARTED,\n   * TEXT_MESSAGE_*, RUN_FINISHED) carrying raw JSON text deltas, plus a final\n   * `CUSTOM` event named `structured-output.complete` whose `value` is\n   * `{ object, raw, reasoning? }`. Events must be timestamped when emitted so\n   * their timestamps follow stream order.\n   */\n  structuredOutputStream?: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Declares whether the adapter supports combining `tools` and a\n   * schema-constrained final answer in a single streaming request.\n   *\n   * When `true`, the engine wires `outputSchema` into the regular\n   * `chatStream()` call and skips the separate `runStructuredFinalization`\n   * round-trip. The model's natural final turn carries the\n   * schema-constrained JSON text and the engine harvests it from the agent\n   * loop's accumulated content.\n   *\n   * When `false`, `undefined`, or the method is omitted, the engine runs\n   * the agent loop without `outputSchema` and then issues a separate\n   * `structuredOutput` / `structuredOutputStream` call against the JSON\n   * schema for finalization (the legacy path).\n   *\n   * The method receives the per-call `modelOptions` so providers whose\n   * support depends on the resolved upstream model (e.g. OpenRouter) can\n   * answer per-request. Most adapters can return a constant.\n   */\n  supportsCombinedToolsAndSchema?: (\n    modelOptions?: TProviderOptions | undefined,\n  ) => boolean\n\n  /**\n   * Where native-combined structured output is taken from.\n   *\n   * - `'text'` (default when omitted): the agent loop's accumulated\n   *   assistant text is schema JSON. The engine parses it after the loop.\n   *   HTTP adapters use this.\n   * - `'event'`: the adapter emits `structured-output.complete` during\n   *   `chatStream`. The engine must not parse accumulated prose. Harness\n   *   adapters use this.\n   */\n  combinedStructuredOutputSource?: (\n    modelOptions?: TProviderOptions | undefined,\n  ) => 'text' | 'event'\n}\n\n/**\n * A TextAdapter with any/unknown type parameters.\n * Useful as a constraint in generic functions and interfaces.\n */\nexport type AnyTextAdapter = TextAdapter<any, any, any, any, any, any, any>\n\n/**\n * Abstract base class for text adapters.\n * Extend this class to implement a text adapter for a specific provider.\n *\n * Generic parameters match TextAdapter - all pre-resolved by the provider function.\n */\nexport abstract class BaseTextAdapter<\n  TModel extends string,\n  TProviderOptions extends Record<string, any>,\n  TInputModalities extends ReadonlyArray<Modality>,\n  TMessageMetadataByModality extends DefaultMessageMetadataByModality,\n  TToolCapabilities extends ReadonlyArray<string> = ReadonlyArray<string>,\n  TToolCallMetadata = unknown,\n  TSystemPromptMetadata = never,\n> implements TextAdapter<\n  TModel,\n  TProviderOptions,\n  TInputModalities,\n  TMessageMetadataByModality,\n  TToolCapabilities,\n  TToolCallMetadata,\n  TSystemPromptMetadata\n> {\n  readonly kind = 'text' as const\n  abstract readonly name: string\n  readonly model: TModel\n  readonly requires?: ReadonlyArray<CapabilityHandle> = undefined\n\n  // Type-only property - never assigned at runtime\n  declare '~types': {\n    providerOptions: TProviderOptions\n    inputModalities: TInputModalities\n    messageMetadataByModality: TMessageMetadataByModality\n    toolCapabilities: TToolCapabilities\n    toolCallMetadata: TToolCallMetadata\n    systemPromptMetadata: TSystemPromptMetadata\n  }\n\n  protected config: TextAdapterConfig\n\n  constructor(config: TextAdapterConfig = {}, model: TModel) {\n    this.config = config\n    this.model = model\n  }\n\n  abstract chatStream(\n    options: TextOptions<TProviderOptions>,\n  ): AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Generate structured output using the provider's native structured output API.\n   * Concrete implementations should override this to use provider-specific structured output.\n   */\n  abstract structuredOutput(\n    options: StructuredOutputOptions<TProviderOptions>,\n  ): Promise<StructuredOutputResult<unknown>>\n\n  protected generateId(): string {\n    return `${this.name}-${Date.now()}-${Math.random().toString(36).substring(7)}`\n  }\n}\n"],"mappings":";;;;;;;AA8LA,IAAsB,kBAAtB,MAgBE;CACA,OAAgB;CAEhB;CACA,WAAsD,KAAA;CAYtD;CAEA,YAAY,SAA4B,CAAC,GAAG,OAAe;EACzD,KAAK,SAAS;EACd,KAAK,QAAQ;CACf;CAcA,aAA+B;EAC7B,OAAO,GAAG,KAAK,KAAK,GAAG,KAAK,IAAI,EAAE,GAAG,KAAK,OAAO,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,UAAU,CAAC;CAC7E;AACF"}
+{"version":3,"file":"adapter.js","names":[],"sources":["../../../../src/activities/chat/adapter.ts"],"sourcesContent":["import type {\n  DefaultMessageMetadataByModality,\n  JSONSchema,\n  Modality,\n  TextOptions,\n  TokenUsage,\n} from '../../types'\nimport type { AdapterYieldChunk } from '../../utilities/adapter-yield-chunk'\nimport type { CapabilityHandle } from './middleware/capabilities'\n\n/**\n * Configuration for adapter instances\n */\nexport interface TextAdapterConfig {\n  apiKey?: string\n  baseUrl?: string\n  timeout?: number\n  maxRetries?: number\n  headers?: Record<string, string>\n}\n\n/**\n * Options for structured output generation.\n *\n * The internal logger is threaded through `chatOptions.logger` (inherited from\n * `TextOptions`). Adapter implementations must call `logger.request()` before\n * SDK calls, `logger.provider()` for each chunk received, and `logger.errors()`\n * in catch blocks.\n */\nexport interface StructuredOutputOptions<TProviderOptions extends object> {\n  /** Text options for the request */\n  chatOptions: TextOptions<TProviderOptions>\n  /** JSON Schema for structured output - already converted from Zod in the ai layer */\n  outputSchema: JSONSchema\n}\n\n/**\n * Result from structured output generation\n */\nexport interface StructuredOutputResult<T = unknown> {\n  /** The parsed data conforming to the schema */\n  data: T\n  /** The raw text response from the model before parsing */\n  rawText: string\n  /** Token usage information (if provided by the adapter) */\n  usage?: TokenUsage\n}\n\n/**\n * Text adapter interface with pre-resolved generics.\n *\n * An adapter is created by a provider function: `provider('model')` → `adapter`\n * All type resolution happens at the provider call site, not in this interface.\n *\n * Generic parameters:\n * - TModel: The specific model name (e.g., 'gpt-4o')\n * - TProviderOptions: Provider-specific options for this model (already resolved)\n * - TInputModalities: Supported input modalities for this model (already resolved)\n * - TMessageMetadata: Metadata types for content parts (already resolved)\n * - TToolCapabilities: Tuple of tool-kind strings supported by this model, resolved from `supports.tools`\n * - TToolCallMetadata: Metadata type that round-trips with tool calls (e.g. Gemini's `thoughtSignature`)\n * - TSystemPromptMetadata: Provider-typed metadata accepted on each\n *   `systemPrompts[i]` entry (e.g. Anthropic `cache_control`). Defaults to\n *   `never` — adapters without per-prompt metadata reject the `metadata`\n *   field at the call site.\n */\nexport interface TextAdapter<\n  TModel extends string,\n  TProviderOptions extends Record<string, any>,\n  TInputModalities extends ReadonlyArray<Modality>,\n  TMessageMetadataByModality extends DefaultMessageMetadataByModality,\n  TToolCapabilities extends ReadonlyArray<string> = ReadonlyArray<string>,\n  TToolCallMetadata = unknown,\n  TSystemPromptMetadata = never,\n> {\n  /** Discriminator for adapter kind */\n  readonly kind: 'text'\n  /** Provider name identifier (e.g., 'openai', 'anthropic') */\n  readonly name: string\n  /** The model this adapter is configured for */\n  readonly model: TModel\n\n  /**\n   * Capabilities this adapter requires at runtime. `chat()` validates that the\n   * configured middleware provides each one. Model adapters omit this; harness\n   * adapters (e.g. a future `claudeCode()`) declare e.g. `[sandboxCapability]`.\n   * Runtime access to capabilities from inside the adapter is not yet wired —\n   * this is the declaration/validation surface only.\n   */\n  readonly requires?: ReadonlyArray<CapabilityHandle>\n\n  /**\n   * Declares that this adapter can consume `{ type: 'file' }` content sources\n   * (provider Files API references). `chat()` rejects file sources in preflight\n   * for adapters that don't declare this, so an adapter written before the\n   * file arm existed fails closed instead of silently mis-mapping a reference\n   * onto its URL/data branch.\n   */\n  readonly supportsFileSources?: boolean\n\n  /**\n   * @internal Type-only properties for inference. Not assigned at runtime.\n   */\n  '~types': {\n    providerOptions: TProviderOptions\n    inputModalities: TInputModalities\n    messageMetadataByModality: TMessageMetadataByModality\n    toolCapabilities: TToolCapabilities\n    toolCallMetadata: TToolCallMetadata\n    systemPromptMetadata: TSystemPromptMetadata\n  }\n\n  /**\n   * Stream text completions from the model\n   */\n  chatStream: (\n    options: TextOptions<TProviderOptions>,\n  ) => AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Generate structured output using the provider's native structured output API.\n   * This method uses stream: false and sends the JSON schema to the provider\n   * to ensure the response conforms to the expected structure.\n   *\n   * @param options - Structured output options containing chat options and JSON schema\n   * @returns Promise with the raw data (validation is done in the chat function)\n   */\n  structuredOutput: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => Promise<StructuredOutputResult<unknown>>\n\n  /**\n   * Stream structured output using the provider's native streaming structured\n   * output API (stream + response_format json_schema in a single request).\n   *\n   * Optional — adapters without native streaming JSON omit this method and the\n   * activity layer synthesizes a stream around the non-streaming\n   * `structuredOutput` call.\n   *\n   * Implementations must emit standard AG-UI lifecycle events (RUN_STARTED,\n   * TEXT_MESSAGE_*, RUN_FINISHED) carrying raw JSON text deltas, plus a final\n   * `CUSTOM` event named `structured-output.complete` whose `value` is\n   * `{ object, raw, reasoning? }`. Events must be timestamped when emitted so\n   * their timestamps follow stream order.\n   */\n  structuredOutputStream?: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Declares whether the adapter supports combining `tools` and a\n   * schema-constrained final answer in a single streaming request.\n   *\n   * When `true`, the engine wires `outputSchema` into the regular\n   * `chatStream()` call and skips the separate `runStructuredFinalization`\n   * round-trip. The model's natural final turn carries the\n   * schema-constrained JSON text and the engine harvests it from the agent\n   * loop's accumulated content.\n   *\n   * When `false`, `undefined`, or the method is omitted, the engine runs\n   * the agent loop without `outputSchema` and then issues a separate\n   * `structuredOutput` / `structuredOutputStream` call against the JSON\n   * schema for finalization (the legacy path).\n   *\n   * The method receives the per-call `modelOptions` so providers whose\n   * support depends on the resolved upstream model (e.g. OpenRouter) can\n   * answer per-request. Most adapters can return a constant.\n   */\n  supportsCombinedToolsAndSchema?: (\n    modelOptions?: TProviderOptions | undefined,\n  ) => boolean\n\n  /**\n   * Where native-combined structured output is taken from.\n   *\n   * - `'text'` (default when omitted): the agent loop's accumulated\n   *   assistant text is schema JSON. The engine parses it after the loop.\n   *   HTTP adapters use this.\n   * - `'event'`: the adapter emits `structured-output.complete` during\n   *   `chatStream`. The engine must not parse accumulated prose. Harness\n   *   adapters use this.\n   */\n  combinedStructuredOutputSource?: (\n    modelOptions?: TProviderOptions | undefined,\n  ) => 'text' | 'event'\n}\n\n/**\n * A TextAdapter with any/unknown type parameters.\n * Useful as a constraint in generic functions and interfaces.\n */\nexport type AnyTextAdapter = TextAdapter<any, any, any, any, any, any, any>\n\n/**\n * Abstract base class for text adapters.\n * Extend this class to implement a text adapter for a specific provider.\n *\n * Generic parameters match TextAdapter - all pre-resolved by the provider function.\n */\nexport abstract class BaseTextAdapter<\n  TModel extends string,\n  TProviderOptions extends Record<string, any>,\n  TInputModalities extends ReadonlyArray<Modality>,\n  TMessageMetadataByModality extends DefaultMessageMetadataByModality,\n  TToolCapabilities extends ReadonlyArray<string> = ReadonlyArray<string>,\n  TToolCallMetadata = unknown,\n  TSystemPromptMetadata = never,\n> implements TextAdapter<\n  TModel,\n  TProviderOptions,\n  TInputModalities,\n  TMessageMetadataByModality,\n  TToolCapabilities,\n  TToolCallMetadata,\n  TSystemPromptMetadata\n> {\n  readonly kind = 'text' as const\n  abstract readonly name: string\n  readonly model: TModel\n  readonly requires?: ReadonlyArray<CapabilityHandle> = undefined\n  readonly supportsFileSources: boolean = false\n\n  // Type-only property - never assigned at runtime\n  declare '~types': {\n    providerOptions: TProviderOptions\n    inputModalities: TInputModalities\n    messageMetadataByModality: TMessageMetadataByModality\n    toolCapabilities: TToolCapabilities\n    toolCallMetadata: TToolCallMetadata\n    systemPromptMetadata: TSystemPromptMetadata\n  }\n\n  protected config: TextAdapterConfig\n\n  constructor(config: TextAdapterConfig = {}, model: TModel) {\n    this.config = config\n    this.model = model\n  }\n\n  abstract chatStream(\n    options: TextOptions<TProviderOptions>,\n  ): AsyncIterable<AdapterYieldChunk>\n\n  /**\n   * Generate structured output using the provider's native structured output API.\n   * Concrete implementations should override this to use provider-specific structured output.\n   */\n  abstract structuredOutput(\n    options: StructuredOutputOptions<TProviderOptions>,\n  ): Promise<StructuredOutputResult<unknown>>\n\n  protected generateId(): string {\n    return `${this.name}-${Date.now()}-${Math.random().toString(36).substring(7)}`\n  }\n}\n"],"mappings":";;;;;;;AAuMA,IAAsB,kBAAtB,MAgBE;CACA,OAAgB;CAEhB;CACA,WAAsD,KAAA;CACtD,sBAAwC;CAYxC;CAEA,YAAY,SAA4B,CAAC,GAAG,OAAe;EACzD,KAAK,SAAS;EACd,KAAK,QAAQ;CACf;CAcA,aAA+B;EAC7B,OAAO,GAAG,KAAK,KAAK,GAAG,KAAK,IAAI,EAAE,GAAG,KAAK,OAAO,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,UAAU,CAAC;CAC7E;AACF"}

package/dist/esm/activities/chat/index.js

@@ -17,12 +17,13 @@ import { hashSchemaInput, normalizeApprovalSchema } from "./tools/approval-schem
 import { INTERRUPT_BINDING_METADATA_KEY, InterruptResumeValidationError, readInterruptBinding, readUnopenedInterruptBinding, validateInterruptResumeBatch } from "../../interrupt-resume.js";
 import { INTERRUPT_PAYLOAD_METADATA_KEY, createInterruptBinding, rehydrateInterruptRequest } from "../../interrupt-definition.js";
 import { readGenericInterruptContinuation } from "../../generic-interrupt-continuation.js";
-import { normalizeToolResult } from "../../utilities/tool-result.js";
-import { subagentHostMessageId } from "../../utilities/subagent-wire.js";
 import { isProviderExecutedToolCall } from "../../utilities/provider-executed.js";
+import { normalizeToolResult, parseToolOutput, toolResultErrorText } from "../../utilities/tool-result.js";
+import { subagentHostMessageId } from "../../utilities/subagent-wire.js";
 import { appendUiResourceToModelMessages, convertMessagesToModelMessages, generateMessageId, modelMessagesToUIMessages, safeJsonStringify, uiResourcePartFromCustomValue } from "./messages.js";
 import { uiMessagesToWire } from "../../utilities/ag-ui-wire.js";
 import { restorePublicUsage } from "../../utilities/restore-inbound-chunk.js";
+import { assertMessagesFileSourceSupport } from "../../utilities/content-source.js";
 import { LazyToolManager } from "./tools/lazy-tool-manager.js";
 import { assertUniqueToolNames } from "./tools/unique-tool-names.js";
 import { MiddlewareAbortError, ToolCallManager, executeToolCalls } from "./tools/tool-calls.js";
@@ -50,12 +51,6 @@ var kind = "text";
 var interruptBindingMetadataKey = INTERRUPT_BINDING_METADATA_KEY;
 /** Resume entries a subagent tool call owns. The parent run skips them. */
 var CHILD_RESUME_IDS = Symbol("tanstack.ai.childResumeIds");
-function assertNoFileSources(adapterName, messages) {
-	for (const message of messages) {
-		if (!Array.isArray(message.content)) continue;
-		for (const part of message.content) if ("source" in part && part.source.type === "file") throw new Error(`${adapterName} does not support provider file-handle sources ({ type: 'file' }). Pass a data or url source.`);
-	}
-}
 function isInterruptSubmissionError(value) {
 	if (value === null || typeof value !== "object" || Array.isArray(value)) return false;
 	if (!("scope" in value) || !("code" in value) || !("message" in value) || !("source" in value) || !("retryable" in value) || !("threadId" in value) || !("interruptedRunId" in value) || !("generation" in value) || typeof value.code !== "string" || typeof value.message !== "string" || typeof value.retryable !== "boolean" || typeof value.threadId !== "string" || typeof value.interruptedRunId !== "string" || typeof value.generation !== "number") return false;
@@ -143,6 +138,15 @@ var TextEngine = class {
 	streamIdentityCaptured = false;
 	accumulatedContent = "";
 	accumulatedThinking = [];
+	/**
+	* Arrival order of this iteration's thinking steps, text and tool calls.
+	* A ModelMessage keeps `thinking` apart from `content`/`toolCalls`, so a
+	* provider turn that thinks between provider-executed tools would otherwise
+	* be recorded as "all thinking, then text, then tools" and the provider
+	* rejects the replay (signed thinking must keep its position). `null` once
+	* the order could no longer be tracked (callers fall back to one message).
+	*/
+	turnParts = [];
 	currentThinkingContent = "";
 	currentThinkingSignature = "";
 	eventOptions;
@@ -501,6 +505,7 @@ var TextEngine = class {
 		this.streamIdentityCaptured = false;
 		this.accumulatedContent = "";
 		this.accumulatedThinking = [];
+		this.turnParts = [];
 		this.currentThinkingContent = "";
 		this.currentThinkingSignature = "";
 		this.finishedEvent = null;
@@ -531,7 +536,7 @@ var TextEngine = class {
 		const { approvals } = this.collectClientState();
 		const adapterApprovals = /* @__PURE__ */ new Map();
 		for (const [approvalId, resolution] of approvals) adapterApprovals.set(approvalId, typeof resolution === "boolean" ? resolution : resolution.approved);
-		assertNoFileSources(this.adapter.name, this.messages);
+		assertMessagesFileSourceSupport(this.adapter, this.messages);
 		for await (const raw of this.adapter.chatStream({
 			model: this.params.model,
 			messages: this.providerMessages,
@@ -666,9 +671,23 @@ var TextEngine = class {
 	}
 	handleTextMessageContentEvent(chunk) {
 		const extra = chunk;
+		const before = this.accumulatedContent;
 		if (typeof extra.content === "string" && extra.content !== "") this.accumulatedContent = extra.content;
 		else this.accumulatedContent += chunk.delta;
 		this.middlewareCtx.accumulatedContent = this.accumulatedContent;
+		if (!this.turnParts) return;
+		if (!this.accumulatedContent.startsWith(before)) {
+			this.turnParts = null;
+			return;
+		}
+		const delta = this.accumulatedContent.slice(before.length);
+		if (delta === "") return;
+		const last = this.turnParts[this.turnParts.length - 1];
+		if (last && last.type === "text") last.content += delta;
+		else this.turnParts.push({
+			type: "text",
+			content: delta
+		});
 	}
 	captureStreamMessageIdentity(messageId) {
 		this.currentMessageId = messageId;
@@ -685,6 +704,11 @@ var TextEngine = class {
 	handleToolCallStartEvent(chunk) {
 		if (typeof chunk.parentMessageId === "string" && chunk.parentMessageId !== "") this.captureStreamMessageIdentity(chunk.parentMessageId);
 		this.toolCallManager.addToolCallStartEvent(chunk);
+		if (this.turnParts && !this.turnParts.some((part) => part.type === "call" && part.id === chunk.toolCallId)) this.turnParts.push({
+			type: "call",
+			id: chunk.toolCallId,
+			providerExecuted: isProviderExecutedToolCall({ metadata: chunk.metadata })
+		});
 		const metadata = chunk.metadata;
 		const thoughtSignature = metadata != null && typeof metadata === "object" && "thoughtSignature" in metadata && typeof metadata.thoughtSignature === "string" && metadata.thoughtSignature !== "" ? metadata.thoughtSignature : void 0;
 		if (thoughtSignature === void 0) return;
@@ -746,17 +770,42 @@ var TextEngine = class {
 				content: this.currentThinkingContent,
 				...this.currentThinkingSignature && { signature: this.currentThinkingSignature }
 			});
+			if (this.turnParts) {
+				const placeholder = [...this.turnParts].reverse().find((part) => part.type === "thinking" && part.index === -1);
+				const index = this.accumulatedThinking.length - 1;
+				if (placeholder) placeholder.index = index;
+				else this.turnParts.push({
+					type: "thinking",
+					index
+				});
+			}
 			this.currentThinkingContent = "";
 			this.currentThinkingSignature = "";
 		}
 	}
+	/**
+	* Record where the current thinking step sits among this turn's parts. A
+	* step is finalized only when the next step starts (or the turn ends), by
+	* which time later tool calls have already arrived, so the position has to
+	* be noted when the step's first content or signature shows up.
+	*/
+	noteThinkingStepPosition() {
+		if (this.turnParts && this.currentThinkingContent === "" && this.currentThinkingSignature === "") this.turnParts.push({
+			type: "thinking",
+			index: -1
+		});
+	}
 	handleStepStartedEvent() {
 		this.finalizeCurrentThinkingStep();
 	}
 	handleStepFinishedEvent(chunk) {
-		if (typeof chunk.signature === "string" && chunk.signature !== "") this.currentThinkingSignature = chunk.signature;
+		if (typeof chunk.signature === "string" && chunk.signature !== "") {
+			this.noteThinkingStepPosition();
+			this.currentThinkingSignature = chunk.signature;
+		}
 	}
 	handleReasoningMessageContentEvent(chunk) {
+		this.noteThinkingStepPosition();
 		this.currentThinkingContent += chunk.delta;
 	}
 	handleReasoningEncryptedValueEvent(chunk) {
@@ -768,6 +817,7 @@ var TextEngine = class {
 			};
 			return;
 		}
+		this.noteThinkingStepPosition();
 		this.currentThinkingSignature = chunk.encryptedValue;
 	}
 	/**
@@ -1013,16 +1063,67 @@ var TextEngine = class {
 	shouldExecuteToolPhase() {
 		return this.lastFinishReason === "tool_calls" && this.tools.length > 0 && this.toolCallManager.hasToolCalls();
 	}
+	/**
+	* Split this iteration into assistant ModelMessages that keep the provider's
+	* block order: a new segment starts at every thinking step that follows a
+	* provider-executed tool call (the rule buildAssistantMessages applies to
+	* UIMessages). Segments after the first get `${id}-segment-${n}` ids.
+	* Returns null when no split is needed or the order could not be tracked,
+	* so callers fall back to the single-message shape.
+	*/
+	buildOrderedAssistantSegments(toolCalls, id, createdAt) {
+		const parts = this.turnParts;
+		if (!parts) return null;
+		const providerCallIds = new Set(parts.flatMap((part) => part.type === "call" && part.providerExecuted ? [part.id] : []));
+		let current = {
+			thinking: [],
+			text: "",
+			callIds: []
+		};
+		const segments = [current];
+		let split = false;
+		for (const part of parts) if (part.type === "thinking") {
+			const thinking = this.accumulatedThinking[part.index];
+			if (!thinking) return null;
+			if (current.callIds.some((callId) => providerCallIds.has(callId))) {
+				current = {
+					thinking: [thinking],
+					text: "",
+					callIds: []
+				};
+				segments.push(current);
+				split = true;
+			} else current.thinking.push(thinking);
+		} else if (part.type === "text") current.text += part.content;
+		else current.callIds.push(part.id);
+		if (!split) return null;
+		if (segments.map((segment) => segment.text).join("") !== this.accumulatedContent) return null;
+		if (segments.reduce((n, segment) => n + segment.thinking.length, 0) !== this.accumulatedThinking.length) return null;
+		const placed = new Set(segments.flatMap((segment) => segment.callIds));
+		for (const toolCall of toolCalls) if (!placed.has(toolCall.id)) current.callIds.push(toolCall.id);
+		return segments.map((segment, index) => {
+			const segmentCalls = toolCalls.filter((toolCall) => segment.callIds.includes(toolCall.id));
+			return {
+				role: "assistant",
+				content: segment.text || null,
+				...segmentCalls.length > 0 && { toolCalls: segmentCalls },
+				id: id === void 0 ? void 0 : index === 0 ? id : `${id}-segment-${index}`,
+				createdAt,
+				...segment.thinking.length > 0 && { thinking: segment.thinking }
+			};
+		});
+	}
 	addAssistantToolCallMessage(toolCalls) {
 		this.finalizeCurrentThinkingStep();
-		this.messages = [...this.messages, {
+		const segments = this.buildOrderedAssistantSegments(toolCalls, this.currentMessageId ?? void 0, this.currentMessageCreatedAt ?? void 0);
+		this.messages = [...this.messages, ...segments ?? [{
 			role: "assistant",
 			content: this.accumulatedContent || null,
 			toolCalls,
 			id: this.currentMessageId ?? void 0,
 			createdAt: this.currentMessageCreatedAt ?? void 0,
 			...this.accumulatedThinking.length > 0 && { thinking: this.accumulatedThinking }
-		}];
+		}]];
 		this.middlewareCtx.messages = this.messages;
 	}
 	addTerminalAssistantMessages() {
@@ -1063,13 +1164,17 @@ var TextEngine = class {
 				...thinking ? { thinking } : {}
 			});
 		} else {
-			if (!currentTurnAlreadyRecorded && (this.accumulatedContent !== "" || thinking)) messages.push({
-				role: "assistant",
-				content: this.accumulatedContent || null,
-				id: this.currentMessageId ?? this.createId("msg"),
-				createdAt: this.currentMessageCreatedAt ?? /* @__PURE__ */ new Date(),
-				...thinking ? { thinking } : {}
-			});
+			if (!currentTurnAlreadyRecorded && (this.accumulatedContent !== "" || thinking)) {
+				const id = this.currentMessageId ?? this.createId("msg");
+				const createdAt = this.currentMessageCreatedAt ?? /* @__PURE__ */ new Date();
+				messages.push(...this.buildOrderedAssistantSegments(this.toolCallManager.getToolCalls(), id, createdAt) ?? [{
+					role: "assistant",
+					content: this.accumulatedContent || null,
+					id,
+					createdAt,
+					...thinking ? { thinking } : {}
+				}]);
+			}
 			if (structuredOutput) messages.push({
 				role: "assistant",
 				content: raw || null,
@@ -1360,6 +1465,7 @@ var TextEngine = class {
 		const approvalRequests = [];
 		const clientRequests = [];
 		for (const toolCall of toolCalls) {
+			if (isProviderExecutedToolCall(toolCall)) continue;
 			const tool = this.resolveExecutableTools([toolCall]).find((candidate) => candidate.name === toolCall.function.name);
 			if (!tool) continue;
 			let input = {};
@@ -1460,7 +1566,8 @@ var TextEngine = class {
 			const newToolMessage = {
 				role: "tool",
 				content,
-				toolCallId: result.toolCallId
+				toolCallId: result.toolCallId,
+				...result.state === "output-error" && { error: toolResultErrorText(parseToolOutput(wireContent)) }
 			};
 			if (placeholderIdx >= 0) this.messages = [
 				...this.messages.slice(0, placeholderIdx),
@@ -1682,7 +1789,7 @@ var TextEngine = class {
 			tools: baseConfig.tools
 		}));
 		this.applyMiddlewareConfig(postOnConfig);
-		assertNoFileSources(this.adapter.name, this.messages);
+		assertMessagesFileSourceSupport(this.adapter, this.messages);
 		const structuredCallOptions = {
 			chatOptions: {
 				model: this.params.model,
@@ -1832,7 +1939,12 @@ var TextEngine = class {
 	async *harvestCombinedStructuredOutput() {
 		if (!this.finalStructuredOutput) throw new Error("harvestCombinedStructuredOutput called without finalStructuredOutput config");
 		const yieldChunks = this.finalStructuredOutput.yieldChunks;
-		if ((this.finalStructuredOutput.source ?? "text") === "event") {
+		const source = this.finalStructuredOutput.source ?? "text";
+		if (this.lastFinishReason === "length") this.finalizationError = {
+			message: "The response was cut off because the maximum token limit was reached (finish_reason=length); raise the output token limit.",
+			code: "max_tokens"
+		};
+		else if (source === "event") {
 			if (!this.structuredOutputResult) this.finalizationError = {
 				message: "missing structured result",
 				code: "structured-output-missing-result"