Class: OpenAIResponsesProvider

Defined in: providers/openai-responses.ts:417

OpenAI Responses-API provider.

Example

const provider = createOpenAIResponsesProvider({ model: 'gpt-6-sol' });

Implements

LLMProvider

Constructors

Constructor

new OpenAIResponsesProvider(config?): OpenAIResponsesProvider;

Defined in: providers/openai-responses.ts:433

Parameters

Parameter	Type
`config`	`OpenAIResponsesProviderConfig`

Returns

OpenAIResponsesProvider

Properties

name

readonly name: "openai" = 'openai';

Defined in: providers/openai-responses.ts:423

⚠️ 'openai', not 'openai-responses'. This string keys error classification, llm_retry event payloads and ProviderError.provider across the SDK, CLI and Desktop. It names the vendor, not the transport.

Implementation of

LLMProvider.name

Methods

buildRequestBody()

buildRequestBody(messages, options?): ResponsesRequestBody;

Defined in: providers/openai-responses.ts:483

Build the request body.

What is deliberately NEVER sent, and why:

messages — this endpoint takes input.
max_tokens — measured “Unknown parameter”; the cap is max_output_tokens.
temperature / top_p — measured “‘temperature’ is not supported with this model”. Dropped silently, exactly as ClaudeProvider drops them for the models that reject them.
stop — Chat-Completions-shaped, no caller sets it, unmeasured here.
stream_options: {include_usage: true} — Chat-Completions only; usage arrives on response.completed.

Parameters

Parameter	Type
`messages`	`Message`[]
`options?`	`ChatOptions`

Returns

ResponsesRequestBody

chat()

chat(messages, options?): AsyncIterable<StreamChunk>;

Defined in: providers/openai-responses.ts:515

Stream a response.

⚠️ EVERY tool-call chunk carries toolCallId. This endpoint announces BOTH calls of a parallel turn (output_item.added ×2) before either has any arguments. Untagged, the agent’s accumulator closes whichever call it saw last: the two argument streams merge into one buffer, JSON.parse fails, and BOTH calls are dropped while the turn is reported as output-limit truncation — the user sees the agent say what it will do, then do nothing. With the id set, the accumulator keys each call separately and we can stream both live.

Parameters

Parameter	Type
`messages`	`Message`[]
`options?`	`ChatOptions`

Returns

AsyncIterable<StreamChunk>

Implementation of

LLMProvider.chat

countTokens()

countTokens(messages): Promise<number>;

Defined in: providers/openai-responses.ts:459

Count tokens in messages (optional, provider-specific)

Parameters

Parameter	Type
`messages`	`Message`[]

Returns

Promise<number>

Implementation of

LLMProvider.countTokens

getModel()

getModel(): string;

Defined in: providers/openai-responses.ts:451

Get the current default model ID.

Returns

string

Implementation of

LLMProvider.getModel

setModel()

setModel(modelId): void;

Defined in: providers/openai-responses.ts:455

Change the default model for subsequent calls. Same provider only. Takes effect on the next chat() call, not mid-stream.

Parameters

Parameter	Type	Description
`modelId`	`string`	The new model ID (e.g., ‘claude-opus-4-20250514’)

Returns

void

Implementation of

LLMProvider.setModel