TextAdapter

Defined in: packages/ai/src/activities/chat/adapter.ts:75

Text adapter interface with pre-resolved generics.

An adapter is created by a provider function: provider('model') → adapter All type resolution happens at the provider call site, not in this interface.

Generic parameters:

• TModel: The specific model name (e.g., 'gpt-4o')
• TProviderOptions: Provider-specific options for this model (already resolved)
• TInputModalities: Supported input modalities for this model (already resolved)
• TMessageMetadata: Metadata types for content parts (already resolved)
• TToolCapabilities: Tuple of tool-kind strings supported by this model, resolved from supports.tools
• TToolCallMetadata: Metadata type that round-trips with tool calls (e.g. Gemini's thoughtSignature)
• TSystemPromptMetadata: Provider-typed metadata accepted on each systemPrompts[i] entry (e.g. Anthropic cache_control). Defaults to never — adapters without per-prompt metadata reject the metadata field at the call site.

Type Parameters#

TModel#

TModel extends string

TProviderOptions#

TProviderOptions extends Record<string, any>

TInputModalities#

TInputModalities extends ReadonlyArray<Modality>

TMessageMetadataByModality#

TMessageMetadataByModality extends DefaultMessageMetadataByModality

TToolCapabilities#

TToolCapabilities extends ReadonlyArray<string> = ReadonlyArray<string>

TToolCallMetadata#

TToolCallMetadata = unknown

TSystemPromptMetadata#

TSystemPromptMetadata = never

TReasoning#

TReasoning extends ReasoningCapability = never

Properties#

~types#

~types: object;

Defined in: packages/ai/src/activities/chat/adapter.ts:133

Internal

Type-only properties for inference. Not assigned at runtime.

inputModalities#

inputModalities: TInputModalities;

messageMetadataByModality#

messageMetadataByModality: TMessageMetadataByModality;

providerOptions#

providerOptions: TProviderOptions;

reasoning?#

optional reasoning?: TReasoning;

The model's reasoning levels, for chat({ reasoning }). never: the model does not reason (or the adapter does not map reasoning), and the option is a type error. Optional, so an adapter written before this key existed still type-checks (and gets no reasoning option).

systemPromptMetadata#

systemPromptMetadata: TSystemPromptMetadata;

toolCallMetadata#

toolCallMetadata: TToolCallMetadata;

toolCapabilities#

toolCapabilities: TToolCapabilities;

api?#

readonly optional api?: string;

Defined in: packages/ai/src/activities/chat/adapter.ts:90

Wire API identity. Defaults to the adapter kind when omitted.

chatStream#

chatStream: (options) => AsyncIterable<AdapterYieldChunk>;

Defined in: packages/ai/src/activities/chat/adapter.ts:152

Stream text completions from the model

Parameters#

options#

TextOptions<TProviderOptions>

Returns#

AsyncIterable<AdapterYieldChunk>

combinedStructuredOutputSource?#

optional combinedStructuredOutputSource?: (modelOptions?) => "text" | "event";

Defined in: packages/ai/src/activities/chat/adapter.ts:227

Where native-combined structured output is taken from.

• 'text' (default when omitted): the agent loop's accumulated assistant text is schema JSON. The engine parses it after the loop. HTTP adapters use this.
• 'event': the adapter emits structured-output.complete during chatStream. The engine must not parse accumulated prose. Harness adapters use this.

Parameters#

modelOptions?#

TProviderOptions

Returns#

"text" | "event"

compact?#

optional compact?: (options) => Promise<ModelMessage<
  | string
  | ContentPart<unknown, unknown, unknown, unknown, unknown>[]
| null>[]>;

Defined in: packages/ai/src/activities/chat/adapter.ts:192

Compact the history with the provider's own compaction endpoint. Resolves to the history to send from the next call on. It can hold provider items that only this provider reads. Optional: adapters without native compaction omit it.

Parameters#

options#

TextCompactOptions

Returns#

Promise<ModelMessage< | string | ContentPart<unknown, unknown, unknown, unknown, unknown>[] | null>[]>

inputModalities?#

readonly optional inputModalities?: readonly ("text" | "image" | "audio" | "video" | "document")[];

Defined in: packages/ai/src/activities/chat/adapter.ts:120

The input kinds the model reads, at runtime, from the provider's model metadata. undefined means not known. This mirrors the type-level '~types'.inputModalities, which has no runtime value. | undefined keeps model lookups assignable under exactOptionalPropertyTypes.

kind#

readonly kind: "text";

Defined in: packages/ai/src/activities/chat/adapter.ts:86

Discriminator for adapter kind

midConversationChannels?#

readonly optional midConversationChannels?: MidConversationChannels;

Defined in: packages/ai/src/activities/chat/adapter.ts:128

The mid-conversation channels of the model: tools and system prompts added between model calls can go out with no change to the cached prompt prefix. chat() passes TextOptions.midConversationChanges only when a channel is on. undefined means no channels.

model#

readonly model: TModel;

Defined in: packages/ai/src/activities/chat/adapter.ts:94

The model this adapter is configured for

name#

readonly name: string;

Defined in: packages/ai/src/activities/chat/adapter.ts:88

Provider name identifier (e.g., 'openai', 'anthropic')

provider?#

readonly optional provider?: string;

Defined in: packages/ai/src/activities/chat/adapter.ts:92

Provider identity. Defaults to name when omitted.

requires?#

readonly optional requires?: readonly CapabilityHandle[];

Defined in: packages/ai/src/activities/chat/adapter.ts:103

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.

structuredOutput#

structuredOutput: (options) => Promise<StructuredOutputResult<unknown>>;

Defined in: packages/ai/src/activities/chat/adapter.ts:164

Generate structured output using the provider's native structured output API. This method uses stream: false and sends the JSON schema to the provider to ensure the response conforms to the expected structure.

Parameters#

options#

StructuredOutputOptions<TProviderOptions>

Structured output options containing chat options and JSON schema

Returns#

Promise<StructuredOutputResult<unknown>>

Promise with the raw data (validation is done in the chat function)

structuredOutputStream?#

optional structuredOutputStream?: (options) => AsyncIterable<AdapterYieldChunk>;

Defined in: packages/ai/src/activities/chat/adapter.ts:182

Stream structured output using the provider's native streaming structured output API (stream + response_format json_schema in a single request).

Optional — adapters without native streaming JSON omit this method and the activity layer synthesizes a stream around the non-streaming structuredOutput call.

Implementations must emit standard AG-UI lifecycle events (RUN_STARTED, TEXT_MESSAGE_*, RUN_FINISHED) carrying raw JSON text deltas, plus a final CUSTOM event named structured-output.complete whose value is { object, raw, reasoning? }. Events must be timestamped when emitted so their timestamps follow stream order.

Parameters#

options#

StructuredOutputOptions<TProviderOptions>

Returns#

AsyncIterable<AdapterYieldChunk>

supportsCombinedToolsAndSchema?#

optional supportsCombinedToolsAndSchema?: (modelOptions?) => boolean;

Defined in: packages/ai/src/activities/chat/adapter.ts:213

Declares whether the adapter supports combining tools and a schema-constrained final answer in a single streaming request.

When true, the engine wires outputSchema into the regular chatStream() call and skips the separate runStructuredFinalization round-trip. The model's natural final turn carries the schema-constrained JSON text and the engine harvests it from the agent loop's accumulated content.

When false, undefined, or the method is omitted, the engine runs the agent loop without outputSchema and then issues a separate structuredOutput / structuredOutputStream call against the JSON schema for finalization (the legacy path).

The method receives the per-call modelOptions so providers whose support depends on the resolved upstream model (e.g. OpenRouter) can answer per-request. Most adapters can return a constant.

Parameters#

modelOptions?#

TProviderOptions

Returns#

boolean

supportsFileSources?#

readonly optional supportsFileSources?: boolean;

Defined in: packages/ai/src/activities/chat/adapter.ts:112

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.