@tanstack/ai 0.28.0 → 0.31.0

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

@@ -1,4 +1,5 @@
 import { DefaultMessageMetadataByModality, JSONSchema, Modality, StreamChunk, TextOptions, TokenUsage } from '../../types.js';
+import { CapabilityHandle } from './middleware/capabilities.js';
 /**
  * Configuration for adapter instances
  */
@@ -59,6 +60,14 @@ export interface TextAdapter<TModel extends string, TProviderOptions extends Rec
     readonly name: string;
     /** The model this adapter is configured for */
     readonly model: TModel;
+    /**
+     * Capabilities this adapter requires at runtime. `chat()` validates that the
+     * configured middleware provides each one. Model adapters omit this; harness
+     * adapters (e.g. a future `claudeCode()`) declare e.g. `[sandboxCapability]`.
+     * Runtime access to capabilities from inside the adapter is not yet wired —
+     * this is the declaration/validation surface only.
+     */
+    readonly requires?: ReadonlyArray<CapabilityHandle>;
     /**
      * @internal Type-only properties for inference. Not assigned at runtime.
      */
@@ -133,6 +142,7 @@ export declare abstract class BaseTextAdapter<TModel extends string, TProviderOp
     readonly kind: "text";
     abstract readonly name: string;
     readonly model: TModel;
+    readonly requires?: ReadonlyArray<CapabilityHandle>;
     '~types': {
         providerOptions: TProviderOptions;
         inputModalities: TInputModalities;

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

@@ -1,6 +1,7 @@
 class BaseTextAdapter {
   kind = "text";
   model;
+  requires = void 0;
   config;
   constructor(config = {}, model) {
     this.config = config;

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

@@ -1 +1 @@
-{"version":3,"file":"adapter.js","sources":["../../../../src/activities/chat/adapter.ts"],"sourcesContent":["import type {\n  DefaultMessageMetadataByModality,\n  JSONSchema,\n  Modality,\n  StreamChunk,\n  TextOptions,\n  TokenUsage,\n} from '../../types'\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   * @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<StreamChunk>\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? }`.\n   */\n  structuredOutputStream?: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => AsyncIterable<StreamChunk>\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/**\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\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<StreamChunk>\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"],"names":[],"mappings":"AAqKO,MAAe,gBAgBpB;AAAA,EACS,OAAO;AAAA,EAEP;AAAA,EAYC;AAAA,EAEV,YAAY,SAA4B,CAAA,GAAI,OAAe;AACzD,SAAK,SAAS;AACd,SAAK,QAAQ;AAAA,EACf;AAAA,EAcU,aAAqB;AAC7B,WAAO,GAAG,KAAK,IAAI,IAAI,KAAK,KAAK,IAAI,KAAK,OAAA,EAAS,SAAS,EAAE,EAAE,UAAU,CAAC,CAAC;AAAA,EAC9E;AACF;"}
+{"version":3,"file":"adapter.js","sources":["../../../../src/activities/chat/adapter.ts"],"sourcesContent":["import type {\n  DefaultMessageMetadataByModality,\n  JSONSchema,\n  Modality,\n  StreamChunk,\n  TextOptions,\n  TokenUsage,\n} from '../../types'\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<StreamChunk>\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? }`.\n   */\n  structuredOutputStream?: (\n    options: StructuredOutputOptions<TProviderOptions>,\n  ) => AsyncIterable<StreamChunk>\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/**\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<StreamChunk>\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"],"names":[],"mappings":"AA+KO,MAAe,gBAgBpB;AAAA,EACS,OAAO;AAAA,EAEP;AAAA,EACA,WAA6C;AAAA,EAY5C;AAAA,EAEV,YAAY,SAA4B,CAAA,GAAI,OAAe;AACzD,SAAK,SAAS;AACd,SAAK,QAAQ;AAAA,EACf;AAAA,EAcU,aAAqB;AAC7B,WAAO,GAAG,KAAK,IAAI,IAAI,KAAK,KAAK,IAAI,KAAK,OAAA,EAAS,SAAS,EAAE,EAAE,UAAU,CAAC,CAAC;AAAA,EAC9E;AACF;"}

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

@@ -1,6 +1,7 @@
 import { AnyTextAdapter } from './adapter.js';
 import { AnyTool, ConstrainedModelMessage, InferSchemaType, ModelMessage, SchemaInput, StreamChunk, StructuredOutputStream, TextOptions, UIMessage } from '../../types.js';
-import { ChatMiddleware } from './middleware/types.js';
+import { AnyChatMiddleware, ChatMiddleware } from './middleware/types.js';
+import { CheckCoverage } from './middleware/builder.js';
 import { SystemPrompt } from '../../system-prompts.js';
 import { DebugOption } from '../../logger/types.js';
 import { ProviderTool } from '../../tools/provider-tool.js';
@@ -30,7 +31,7 @@ type RequiredContextFromInputs<TTools, TMiddleware> = [
 };
 type TextActivityOptionsWithContext<TAdapter extends AnyTextAdapter, TSchema extends SchemaInput | undefined, TStream extends boolean, TTools extends TextActivityOptions<TAdapter, TSchema, TStream, any>['tools'], TMiddleware extends TextActivityOptions<TAdapter, TSchema, TStream, any>['middleware']> = Omit<TextActivityOptions<TAdapter, TSchema, TStream, any>, 'tools' | 'middleware' | 'context'> & {
     tools?: TTools;
-    middleware?: TMiddleware;
+    middleware?: TMiddleware & CheckCoverage<Extract<TMiddleware, ReadonlyArray<AnyChatMiddleware>>>;
 } & RequiredContextFromInputs<TTools, TMiddleware>;
 /**
  * Options for the text activity.

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

@@ -9,6 +9,8 @@ import { convertSchemaToJsonSchema, isStandardSchema, parseWithStandardSchema }
 import { maxIterations } from "./agent-loop-strategies.js";
 import { convertMessagesToModelMessages, generateMessageId } from "./messages.js";
 import { MiddlewareRunner } from "./middleware/compose.js";
+import { CapabilityRegistry } from "./middleware/capabilities.js";
+import { validateCapabilities } from "./middleware/validate.js";
 import { MCPManager } from "./mcp/manager.js";
 import { EventType } from "@ag-ui/core";
 const kind = "text";
@@ -170,7 +172,15 @@ class TextEngine {
       accumulatedContent: "",
       // References
       messages: this.messages,
-      createId: (prefix) => this.createId(prefix)
+      createId: (prefix) => this.createId(prefix),
+      // Capability bookkeeping for this request (populated by middleware setup)
+      capabilities: new CapabilityRegistry(),
+      // Convenience accessors that delegate to a capability handle's own
+      // tuple getter/provider, keyed by this context. `getX(ctx)` and
+      // `ctx.get(X)` are interchangeable.
+      get: (capability) => capability[0](this.middlewareCtx),
+      getOptional: (capability) => capability[0](this.middlewareCtx, { optional: true }),
+      provide: (capability, value) => capability[1](this.middlewareCtx, value)
     };
   }
   /** Get the accumulated content after the chat loop completes */
@@ -205,6 +215,7 @@ class TextEngine {
       threadId: this.middlewareCtx.threadId
     });
     try {
+      await this.middlewareRunner.runSetup(this.middlewareCtx);
       this.middlewareCtx.phase = "init";
       const initialConfig = this.buildMiddlewareConfig();
       const transformedConfig = await this.middlewareRunner.runOnConfig(
@@ -896,17 +907,17 @@ class TextEngine {
           delta: args,
           args
         });
+        chunks.push({
+          type: "TOOL_CALL_END",
+          timestamp: Date.now(),
+          model: finishEvent.model,
+          toolCallId: result.toolCallId,
+          toolCallName: result.toolName,
+          toolName: result.toolName,
+          result: wireContent,
+          ...result.state !== void 0 && { state: result.state }
+        });
       }
-      chunks.push({
-        type: "TOOL_CALL_END",
-        timestamp: Date.now(),
-        model: finishEvent.model,
-        toolCallId: result.toolCallId,
-        toolCallName: result.toolName,
-        toolName: result.toolName,
-        result: wireContent,
-        ...result.state !== void 0 && { state: result.state }
-      });
       chunks.push({
         type: "TOOL_CALL_RESULT",
         timestamp: Date.now(),
@@ -1406,6 +1417,7 @@ class TextEngine {
   }
 }
 function chat(options) {
+  validateCapabilities(options.middleware ?? [], options.adapter);
   const { outputSchema, stream } = options;
   if (outputSchema && stream === true) {
     return runStreamingStructuredOutput({