@warlock.js/ai-openai 5.26.0 → 5.28.0

package/esm/responses-model.mjs

@@ -0,0 +1,329 @@
+import { buildUsage } from "./utils/build-usage.mjs";
+import { wrapOpenAIError } from "./utils/wrap-openai-error.mjs";
+import { mapResponsesFinishReason, toResponseFailure } from "./utils/map-responses-status.mjs";
+import { buildPromptCacheParams } from "./utils/prompt-cache-params.mjs";
+import { toOpenAIReasoningEffort } from "./utils/reasoning-effort.mjs";
+import { inferStructuredOutput, planStructuredOutput } from "./utils/structured-output.mjs";
+import { attachReasoningReplay } from "./utils/reasoning-replay.mjs";
+import { toResponsesInput } from "./utils/to-responses-input.mjs";
+import { toResponsesToolChoice, toResponsesTools } from "./utils/to-responses-tools.mjs";
+import "./utils/index.mjs";
+import { inferReasoningCapability } from "./known-reasoning-models.mjs";
+import { inferVisionCapability } from "./known-vision-models.mjs";
+import { InvalidRequestError, ProviderError, safeJsonParse } from "@warlock.js/ai";
+import { log } from "@warlock.js/logger";
+
+//#region ../ai-openai/src/responses-model.ts
+const LOG_MODULE = "ai.openai";
+/**
+* OpenAI Responses API implementation of `ModelContract` (`api: "responses"`).
+*
+* **Role.** A sibling of `OpenAIModel` that speaks `POST /v1/responses`
+* instead of Chat Completions. Its point is letting a tool-using agent run a
+* reasoning model with reasoning ON: the Chat path has to force
+* `reasoning_effort: "none"` whenever tools are attached, Responses does not.
+*
+* **Why a separate class.** Only `OpenAISDK.model()` constructs it, and only
+* for the direct `"openai"` provider. The wrapper packages (deepseek, groq,
+* xai, mistral) build `OpenAIModel` and cannot reach this class, so an
+* endpoint that does not implement `/v1/responses` is never sent there. The
+* constructor re-checks the provider label as a second lock.
+*
+* **Stateless.** Every request sends `store: false`; nothing is kept on the
+* server and no `previous_response_id` is used. Full history is re-sent each
+* call, exactly like the Chat path.
+*
+* **Reasoning replay.** Every request asks for
+* `include: ["reasoning.encrypted_content"]`. The `reasoning` items of a
+* tool-calling turn are stored on the FIRST tool call as
+* `providerMetadata.openaiResponses = { model, reasoningItems }` and sent back
+* before that turn's `function_call` items on the next request (only when the
+* stored model equals the current one). Text-only turns are not replayed.
+*
+* **Streaming.** `stream()` uses `responses.create({ stream: true })` and yields
+* the same chunk sequence as the Chat path: text `delta`s, then the
+* `tool-call`s, then one `done`.
+*/
+var OpenAIResponsesModel = class {
+	constructor(client, config, provider = "openai") {
+		this.logger = log;
+		if (provider !== "openai") throw new InvalidRequestError(`api: "responses" is supported only on the direct "openai" provider; got provider "${provider}".`, { context: {
+			provider,
+			model: config.name,
+			api: "responses"
+		} });
+		this.client = client;
+		this.config = config;
+		this.name = config.name;
+		this.provider = provider;
+		this.pricing = config.pricing;
+		this.capabilities = {
+			structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),
+			vision: config.vision ?? inferVisionCapability(config.name),
+			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
+			promptCaching: true,
+			pdf: false,
+			audio: false
+		};
+	}
+	/**
+	* Single-shot completion against `client.responses.create`. Sends the full
+	* history as Responses `input` items, walks `response.output`, and reshapes
+	* it into a vendor-neutral `ModelResponse`. Per-call `options` override the
+	* instance defaults for this call only.
+	*/
+	async complete(messages, options) {
+		this.logger.debug(LOG_MODULE, "request", "Starting call to responses.create", {
+			model: this.name,
+			messageCount: messages.length,
+			streaming: false,
+			toolCount: options?.tools?.length ?? 0
+		});
+		let response;
+		try {
+			response = await this.client.responses.create(this.buildRequest(messages, options), options?.signal ? { signal: options.signal } : void 0);
+		} catch (thrown) {
+			throw this.logAndWrap(thrown);
+		}
+		if (response.status === "failed") throw this.logAndWrap(toResponseFailure(response, this.name));
+		const output = Array.isArray(response.output) ? response.output : [];
+		const { content, toolCalls } = this.readOutput(output);
+		const finishReason = mapResponsesFinishReason({
+			status: response.status,
+			incomplete_details: response.incomplete_details,
+			hasToolCalls: toolCalls !== void 0
+		});
+		const usage = this.extractUsage(response.usage);
+		this.logger.debug(LOG_MODULE, "response", "call to responses.create succeeded", {
+			finishReason,
+			usage
+		});
+		return {
+			content,
+			finishReason,
+			usage,
+			...toolCalls ? { toolCalls } : {}
+		};
+	}
+	/**
+	* Incremental streaming completion via `responses.create({ stream: true })`
+	* (the SDK's `Stream<ResponseStreamEvent>`, `responses.d.ts:55`). Yields the
+	* same neutral `ModelStreamChunk` sequence as the Chat path:
+	*
+	* - `response.output_text.delta` / `response.refusal.delta` -> `delta`.
+	* - `response.output_item.done` with a completed `function_call` item ->
+	*   buffered; `reasoning` items are captured for replay. The `done` copy is
+	*   used because the `added` copy may be incomplete (`responses.d.ts:5711`).
+	* - After the loop: the buffered `tool-call`s (the first carries the
+	*   replay metadata), then a single `done` with the finish reason and usage
+	*   from `response.completed` / `response.incomplete`.
+	* - `response.failed` and `error` events throw a typed `AIError`; a stream
+	*   that ends with no terminal event throws `ProviderError`.
+	* - Abort goes through the request `signal`; the SDK rejects the read and
+	*   the error is wrapped like any other.
+	*/
+	async *stream(messages, options) {
+		this.logger.debug(LOG_MODULE, "request", "Starting streaming call to responses.create", {
+			model: this.name,
+			messageCount: messages.length,
+			streaming: true,
+			toolCount: options?.tools?.length ?? 0
+		});
+		let events;
+		try {
+			events = await this.client.responses.create({
+				...this.buildRequest(messages, options),
+				stream: true
+			}, options?.signal ? { signal: options.signal } : void 0);
+		} catch (thrown) {
+			throw this.logAndWrap(thrown);
+		}
+		const toolCalls = [];
+		const reasoningItems = [];
+		let terminal;
+		try {
+			for await (const event of events) {
+				if (event.type === "response.output_text.delta" || event.type === "response.refusal.delta") {
+					if (event.delta) yield {
+						type: "delta",
+						content: event.delta
+					};
+					continue;
+				}
+				if (event.type === "response.output_item.done") {
+					const { item } = event;
+					if (item.type === "reasoning") reasoningItems.push(item);
+					else if (item.type === "function_call" && item.status !== "incomplete") toolCalls.push(this.toToolCall(item));
+					continue;
+				}
+				if (event.type === "response.completed" || event.type === "response.incomplete") {
+					terminal = event.response;
+					continue;
+				}
+				if (event.type === "response.failed") throw toResponseFailure(event.response, this.name);
+				if (event.type === "error") throw wrapOpenAIError({
+					code: event.code,
+					message: event.message,
+					param: event.param
+				});
+			}
+			if (terminal === void 0) throw new ProviderError("OpenAI Responses stream ended before response.completed.", { context: {
+				provider: this.provider,
+				model: this.name
+			} });
+		} catch (thrown) {
+			throw this.logAndWrap(thrown);
+		}
+		for (const call of attachReasoningReplay(toolCalls, this.name, reasoningItems)) yield {
+			type: "tool-call",
+			id: call.id,
+			name: call.name,
+			input: call.input,
+			...call.providerMetadata ? { providerMetadata: call.providerMetadata } : {}
+		};
+		const finishReason = mapResponsesFinishReason({
+			status: terminal.status,
+			incomplete_details: terminal.incomplete_details,
+			hasToolCalls: toolCalls.length > 0
+		});
+		const usage = this.extractUsage(terminal.usage);
+		this.logger.debug(LOG_MODULE, "response", "Streaming call to responses.create succeeded", {
+			finishReason,
+			usage
+		});
+		yield {
+			type: "done",
+			finishReason,
+			usage
+		};
+	}
+	/**
+	* Assemble the `responses.create` body.
+	*
+	* - System prompt -> `instructions`; history -> `input` items.
+	* - `store: false` always (the API default is `true` when omitted), plus
+	*   `include: ["reasoning.encrypted_content"]` for stateless reasoning replay.
+	* - `max_output_tokens` caps reasoning and visible output together.
+	* - `temperature` only for non-reasoning models (same rule as Chat).
+	* - `tools` flat with `strict: false`; `tool_choice` only with tools.
+	* - `reasoning.effort` only when the caller gave one, through the same
+	*   clamp table as Chat. With no explicit effort the provider default
+	*   applies EVEN with tools attached - that is the point of this API.
+	*/
+	buildRequest(messages, options) {
+		const { instructions, input } = toResponsesInput(messages, this.name);
+		const tools = toResponsesTools(options?.tools);
+		const toolChoice = tools ? toResponsesToolChoice(options?.toolChoice) : void 0;
+		const maxTokens = options?.maxTokens ?? this.config.maxTokens;
+		const temperature = options?.temperature ?? this.config.temperature;
+		return {
+			model: this.name,
+			...instructions !== void 0 ? { instructions } : {},
+			input,
+			store: false,
+			include: ["reasoning.encrypted_content"],
+			...maxTokens !== void 0 ? { max_output_tokens: maxTokens } : {},
+			...!this.capabilities.reasoning && temperature !== void 0 ? { temperature } : {},
+			...tools ? { tools } : {},
+			...toolChoice !== void 0 ? { tool_choice: toolChoice } : {},
+			...this.buildTextFormat(options?.responseSchema),
+			...this.buildReasoningParams(options?.reasoning),
+			...buildPromptCacheParams(this.provider, this.config)
+		};
+	}
+	/**
+	* Structured output on Responses lives in `text.format`, and `name` /
+	* `schema` / `strict` sit directly on the format object (Chat nests them
+	* under `json_schema`). The strict-or-downgrade decision is the shared
+	* `planStructuredOutput`.
+	*/
+	buildTextFormat(responseSchema) {
+		const plan = planStructuredOutput(responseSchema, this.config.responseFormat);
+		if (plan.mode === "json_schema") return { text: { format: {
+			type: "json_schema",
+			name: "response",
+			schema: plan.schema,
+			strict: true
+		} } };
+		if (plan.mode === "json_object") return { text: { format: { type: "json_object" } } };
+		return {};
+	}
+	/** `reasoning.effort` for reasoning-capable models, only when the caller set one. */
+	buildReasoningParams(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning?.effort) return {};
+		return { reasoning: { effort: toOpenAIReasoningEffort(reasoning.effort) } };
+	}
+	/**
+	* Walk `response.output`: `message` items contribute their `output_text`
+	* text and any `refusal` text (surfaced as ordinary content, decision 5);
+	* each completed `function_call` item becomes one tool call with
+	* `id = call_id`. `reasoning` items carry no visible content and are
+	* not content. A `function_call` the API marked `incomplete` (truncated
+	* arguments) is skipped rather than dispatched with empty input.
+	* `reasoning` items are captured and attached to the first tool call for
+	* replay (see `attachReasoningReplay`).
+	*/
+	readOutput(output) {
+		let content = "";
+		const toolCalls = [];
+		const reasoningItems = [];
+		for (const item of output) {
+			if (item.type === "reasoning") {
+				reasoningItems.push(item);
+				continue;
+			}
+			if (item.type === "message") {
+				for (const part of item.content) content += part.type === "refusal" ? part.refusal : part.text;
+				continue;
+			}
+			if (item.type === "function_call" && item.status !== "incomplete") toolCalls.push(this.toToolCall(item));
+		}
+		return {
+			content,
+			toolCalls: toolCalls.length > 0 ? attachReasoningReplay(toolCalls, this.name, reasoningItems) : void 0
+		};
+	}
+	/** One completed `function_call` item as a neutral tool call (`call_id` -> `id`). */
+	toToolCall(item) {
+		return {
+			id: item.call_id,
+			name: item.name,
+			input: safeJsonParse(item.arguments, {})
+		};
+	}
+	/**
+	* Normalize `usage` (absent on some partials) into the neutral `Usage`.
+	* `input_tokens_details.cached_tokens` -> `cachedTokens`,
+	* `cache_write_tokens` -> `cacheWriteTokens`,
+	* `output_tokens_details.reasoning_tokens` -> `reasoningTokens`. The
+	* detail blocks are typed non-optional but read defensively.
+	*/
+	extractUsage(raw) {
+		if (!raw) return {
+			input: 0,
+			output: 0,
+			total: 0
+		};
+		return buildUsage({
+			input: raw.input_tokens,
+			output: raw.output_tokens,
+			total: raw.total_tokens,
+			cachedTokens: raw.input_tokens_details?.cached_tokens,
+			cacheWriteTokens: raw.input_tokens_details?.cache_write_tokens,
+			reasoningTokens: raw.output_tokens_details?.reasoning_tokens
+		});
+	}
+	/** Wrap (idempotently), log and return a failure for the caller to throw. */
+	logAndWrap(thrown) {
+		const wrapped = wrapOpenAIError(thrown);
+		this.logger.error(LOG_MODULE, "error", wrapped.message, {
+			code: wrapped.code,
+			context: wrapped.context
+		});
+		return wrapped;
+	}
+};
+
+//#endregion
+export { OpenAIResponsesModel };
+//# sourceMappingURL=responses-model.mjs.map

package/esm/responses-model.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"responses-model.mjs","names":[],"sources":["../../../../../../ai-openai/src/responses-model.ts"],"sourcesContent":["import {\n  InvalidRequestError,\n  ProviderError,\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIModelConfig } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport {\n  attachReasoningReplay,\n  buildPromptCacheParams,\n  buildUsage,\n  inferStructuredOutput,\n  mapResponsesFinishReason,\n  planStructuredOutput,\n  toOpenAIReasoningEffort,\n  toResponseFailure,\n  toResponsesInput,\n  toResponsesToolChoice,\n  toResponsesTools,\n  wrapOpenAIError,\n} from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * OpenAI Responses API implementation of `ModelContract` (`api: \"responses\"`).\n *\n * **Role.** A sibling of `OpenAIModel` that speaks `POST /v1/responses`\n * instead of Chat Completions. Its point is letting a tool-using agent run a\n * reasoning model with reasoning ON: the Chat path has to force\n * `reasoning_effort: \"none\"` whenever tools are attached, Responses does not.\n *\n * **Why a separate class.** Only `OpenAISDK.model()` constructs it, and only\n * for the direct `\"openai\"` provider. The wrapper packages (deepseek, groq,\n * xai, mistral) build `OpenAIModel` and cannot reach this class, so an\n * endpoint that does not implement `/v1/responses` is never sent there. The\n * constructor re-checks the provider label as a second lock.\n *\n * **Stateless.** Every request sends `store: false`; nothing is kept on the\n * server and no `previous_response_id` is used. Full history is re-sent each\n * call, exactly like the Chat path.\n *\n * **Reasoning replay.** Every request asks for\n * `include: [\"reasoning.encrypted_content\"]`. The `reasoning` items of a\n * tool-calling turn are stored on the FIRST tool call as\n * `providerMetadata.openaiResponses = { model, reasoningItems }` and sent back\n * before that turn's `function_call` items on the next request (only when the\n * stored model equals the current one). Text-only turns are not replayed.\n *\n * **Streaming.** `stream()` uses `responses.create({ stream: true })` and yields\n * the same chunk sequence as the Chat path: text `delta`s, then the\n * `tool-call`s, then one `done`.\n */\nexport class OpenAIResponsesModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly config: OpenAIModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIModelConfig, provider: string = \"openai\") {\n    if (provider !== \"openai\") {\n      throw new InvalidRequestError(\n        `api: \"responses\" is supported only on the direct \"openai\" provider; got provider \"${provider}\".`,\n        { context: { provider, model: config.name, api: \"responses\" } },\n      );\n    }\n\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),\n      vision: config.vision ?? inferVisionCapability(config.name),\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      // Read-side cache accounting (`cachedTokens`, `cacheWriteTokens`) works\n      // and `prompt_cache_key` is forwarded. Explicit write breakpoints are not\n      // mapped onto Responses yet, so a `cacheControl.breakpoints` hint is\n      // ignored rather than rejected.\n      promptCaching: true,\n      // PDF and audio input are off under Responses in this adapter: the\n      // `input_file` / `input_audio` field-level mapping is unverified. The\n      // `pdf` / `audio` config flags are deliberately NOT honoured here.\n      pdf: false,\n      audio: false,\n    };\n  }\n\n  /**\n   * Single-shot completion against `client.responses.create`. Sends the full\n   * history as Responses `input` items, walks `response.output`, and reshapes\n   * it into a vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to responses.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: OpenAI.Responses.Response;\n\n    try {\n      response = await this.client.responses.create(\n        this.buildRequest(messages, options),\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    if (response.status === \"failed\") {\n      throw this.logAndWrap(toResponseFailure(response, this.name));\n    }\n\n    const output = Array.isArray(response.output) ? response.output : [];\n    const { content, toolCalls } = this.readOutput(output);\n    const finishReason = mapResponsesFinishReason({\n      status: response.status,\n      incomplete_details: response.incomplete_details,\n      hasToolCalls: toolCalls !== undefined,\n    });\n    const usage = this.extractUsage(response.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to responses.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return { content, finishReason, usage, ...(toolCalls ? { toolCalls } : {}) };\n  }\n\n  /**\n   * Incremental streaming completion via `responses.create({ stream: true })`\n   * (the SDK's `Stream<ResponseStreamEvent>`, `responses.d.ts:55`). Yields the\n   * same neutral `ModelStreamChunk` sequence as the Chat path:\n   *\n   * - `response.output_text.delta` / `response.refusal.delta` -> `delta`.\n   * - `response.output_item.done` with a completed `function_call` item ->\n   *   buffered; `reasoning` items are captured for replay. The `done` copy is\n   *   used because the `added` copy may be incomplete (`responses.d.ts:5711`).\n   * - After the loop: the buffered `tool-call`s (the first carries the\n   *   replay metadata), then a single `done` with the finish reason and usage\n   *   from `response.completed` / `response.incomplete`.\n   * - `response.failed` and `error` events throw a typed `AIError`; a stream\n   *   that ends with no terminal event throws `ProviderError`.\n   * - Abort goes through the request `signal`; the SDK rejects the read and\n   *   the error is wrapped like any other.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to responses.create\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let events: AsyncIterable<OpenAI.Responses.ResponseStreamEvent>;\n\n    try {\n      events = await this.client.responses.create(\n        { ...this.buildRequest(messages, options), stream: true },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const toolCalls: ModelToolCallRequest[] = [];\n    const reasoningItems: OpenAI.Responses.ResponseReasoningItem[] = [];\n    let terminal: OpenAI.Responses.Response | undefined;\n\n    try {\n      for await (const event of events) {\n        if (event.type === \"response.output_text.delta\" || event.type === \"response.refusal.delta\") {\n          if (event.delta) {\n            yield { type: \"delta\", content: event.delta };\n          }\n\n          continue;\n        }\n\n        if (event.type === \"response.output_item.done\") {\n          const { item } = event;\n\n          if (item.type === \"reasoning\") {\n            reasoningItems.push(item);\n          } else if (item.type === \"function_call\" && item.status !== \"incomplete\") {\n            toolCalls.push(this.toToolCall(item));\n          }\n\n          continue;\n        }\n\n        if (event.type === \"response.completed\" || event.type === \"response.incomplete\") {\n          terminal = event.response;\n          continue;\n        }\n\n        if (event.type === \"response.failed\") {\n          throw toResponseFailure(event.response, this.name);\n        }\n\n        if (event.type === \"error\") {\n          throw wrapOpenAIError({ code: event.code, message: event.message, param: event.param });\n        }\n      }\n\n      if (terminal === undefined) {\n        throw new ProviderError(\"OpenAI Responses stream ended before response.completed.\", {\n          context: { provider: this.provider, model: this.name },\n        });\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    for (const call of attachReasoningReplay(toolCalls, this.name, reasoningItems)) {\n      yield {\n        type: \"tool-call\",\n        id: call.id,\n        name: call.name,\n        input: call.input,\n        ...(call.providerMetadata ? { providerMetadata: call.providerMetadata } : {}),\n      };\n    }\n\n    const finishReason = mapResponsesFinishReason({\n      status: terminal.status,\n      incomplete_details: terminal.incomplete_details,\n      hasToolCalls: toolCalls.length > 0,\n    });\n    const usage = this.extractUsage(terminal.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to responses.create succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the `responses.create` body.\n   *\n   * - System prompt -> `instructions`; history -> `input` items.\n   * - `store: false` always (the API default is `true` when omitted), plus\n   *   `include: [\"reasoning.encrypted_content\"]` for stateless reasoning replay.\n   * - `max_output_tokens` caps reasoning and visible output together.\n   * - `temperature` only for non-reasoning models (same rule as Chat).\n   * - `tools` flat with `strict: false`; `tool_choice` only with tools.\n   * - `reasoning.effort` only when the caller gave one, through the same\n   *   clamp table as Chat. With no explicit effort the provider default\n   *   applies EVEN with tools attached - that is the point of this API.\n   */\n  private buildRequest(\n    messages: Message[],\n    options: ModelCallOptions | undefined,\n  ): OpenAI.Responses.ResponseCreateParamsNonStreaming {\n    const { instructions, input } = toResponsesInput(messages, this.name);\n    const tools = toResponsesTools(options?.tools);\n    const toolChoice = tools ? toResponsesToolChoice(options?.toolChoice) : undefined;\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      model: this.name,\n      ...(instructions !== undefined ? { instructions } : {}),\n      input,\n      store: false,\n      // Stateless replay: ask for the encrypted reasoning blobs so they can be\n      // sent back on the next tool turn (decision 3).\n      include: [\"reasoning.encrypted_content\"],\n      ...(maxTokens !== undefined ? { max_output_tokens: maxTokens } : {}),\n      ...(!this.capabilities.reasoning && temperature !== undefined ? { temperature } : {}),\n      ...(tools ? { tools } : {}),\n      ...(toolChoice !== undefined ? { tool_choice: toolChoice } : {}),\n      ...this.buildTextFormat(options?.responseSchema),\n      ...this.buildReasoningParams(options?.reasoning),\n      ...buildPromptCacheParams(this.provider, this.config),\n    };\n  }\n\n  /**\n   * Structured output on Responses lives in `text.format`, and `name` /\n   * `schema` / `strict` sit directly on the format object (Chat nests them\n   * under `json_schema`). The strict-or-downgrade decision is the shared\n   * `planStructuredOutput`.\n   */\n  private buildTextFormat(responseSchema: Record<string, unknown> | undefined): {\n    text?: OpenAI.Responses.ResponseTextConfig;\n  } {\n    const plan = planStructuredOutput(responseSchema, this.config.responseFormat);\n\n    if (plan.mode === \"json_schema\") {\n      return {\n        text: { format: { type: \"json_schema\", name: \"response\", schema: plan.schema, strict: true } },\n      };\n    }\n\n    if (plan.mode === \"json_object\") {\n      return { text: { format: { type: \"json_object\" } } };\n    }\n\n    return {};\n  }\n\n  /** `reasoning.effort` for reasoning-capable models, only when the caller set one. */\n  private buildReasoningParams(reasoning: ModelCallOptions[\"reasoning\"]): {\n    reasoning?: { effort: ReturnType<typeof toOpenAIReasoningEffort> };\n  } {\n    if (!this.capabilities.reasoning || !reasoning?.effort) {\n      return {};\n    }\n\n    return { reasoning: { effort: toOpenAIReasoningEffort(reasoning.effort) } };\n  }\n\n  /**\n   * Walk `response.output`: `message` items contribute their `output_text`\n   * text and any `refusal` text (surfaced as ordinary content, decision 5);\n   * each completed `function_call` item becomes one tool call with\n   * `id = call_id`. `reasoning` items carry no visible content and are\n   * not content. A `function_call` the API marked `incomplete` (truncated\n   * arguments) is skipped rather than dispatched with empty input.\n   * `reasoning` items are captured and attached to the first tool call for\n   * replay (see `attachReasoningReplay`).\n   */\n  private readOutput(output: OpenAI.Responses.ResponseOutputItem[]): {\n    content: string;\n    toolCalls: ModelToolCallRequest[] | undefined;\n  } {\n    let content = \"\";\n    const toolCalls: ModelToolCallRequest[] = [];\n    const reasoningItems: OpenAI.Responses.ResponseReasoningItem[] = [];\n\n    for (const item of output) {\n      if (item.type === \"reasoning\") {\n        reasoningItems.push(item);\n        continue;\n      }\n\n      if (item.type === \"message\") {\n        for (const part of item.content) {\n          content += part.type === \"refusal\" ? part.refusal : part.text;\n        }\n\n        continue;\n      }\n\n      if (item.type === \"function_call\" && item.status !== \"incomplete\") {\n        toolCalls.push(this.toToolCall(item));\n      }\n    }\n\n    return {\n      content,\n      toolCalls:\n        toolCalls.length > 0\n          ? attachReasoningReplay(toolCalls, this.name, reasoningItems)\n          : undefined,\n    };\n  }\n\n  /** One completed `function_call` item as a neutral tool call (`call_id` -> `id`). */\n  private toToolCall(item: OpenAI.Responses.ResponseFunctionToolCall): ModelToolCallRequest {\n    return {\n      id: item.call_id,\n      name: item.name,\n      input: safeJsonParse<Record<string, unknown>>(item.arguments, {}),\n    };\n  }\n\n  /**\n   * Normalize `usage` (absent on some partials) into the neutral `Usage`.\n   * `input_tokens_details.cached_tokens` -> `cachedTokens`,\n   * `cache_write_tokens` -> `cacheWriteTokens`,\n   * `output_tokens_details.reasoning_tokens` -> `reasoningTokens`. The\n   * detail blocks are typed non-optional but read defensively.\n   */\n  private extractUsage(raw: OpenAI.Responses.ResponseUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    return buildUsage({\n      input: raw.input_tokens,\n      output: raw.output_tokens,\n      total: raw.total_tokens,\n      cachedTokens: raw.input_tokens_details?.cached_tokens,\n      cacheWriteTokens: raw.input_tokens_details?.cache_write_tokens,\n      reasoningTokens: raw.output_tokens_details?.reasoning_tokens,\n    });\n  }\n\n  /** Wrap (idempotently), log and return a failure for the caller to throw. */\n  private logAndWrap(thrown: unknown): ReturnType<typeof wrapOpenAIError> {\n    const wrapped = wrapOpenAIError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAkCA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BnB,IAAa,uBAAb,MAA2D;CAUzD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAGhC,IAAI,aAAa,UACf,MAAM,IAAI,oBACR,qFAAqF,SAAS,KAC9F,EAAE,SAAS;GAAE;GAAU,OAAO,OAAO;GAAM,KAAK;EAAY,EAAE,CAChE;EAGF,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB,sBAAsB,OAAO,cAAc;GACxF,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;GAC1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GAKnE,eAAe;GAIf,KAAK;GACL,OAAO;EACT;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,qCAAqC;GAC5E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,UAAU,OACrC,KAAK,aAAa,UAAU,OAAO,GACnC,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI,SAAS,WAAW,UACtB,MAAM,KAAK,WAAW,kBAAkB,UAAU,KAAK,IAAI,CAAC;EAG9D,MAAM,SAAS,MAAM,QAAQ,SAAS,MAAM,IAAI,SAAS,SAAS,CAAC;EACnE,MAAM,EAAE,SAAS,cAAc,KAAK,WAAW,MAAM;EACrD,MAAM,eAAe,yBAAyB;GAC5C,QAAQ,SAAS;GACjB,oBAAoB,SAAS;GAC7B,cAAc,cAAc;EAC9B,CAAC;EACD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,sCAAsC;GAC9E;GACA;EACF,CAAC;EAED,OAAO;GAAE;GAAS;GAAc;GAAO,GAAI,YAAY,EAAE,UAAU,IAAI,CAAC;EAAG;CAC7E;;;;;;;;;;;;;;;;;;CAmBA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,+CAA+C;GACtF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,UAAU,OACnC;IAAE,GAAG,KAAK,aAAa,UAAU,OAAO;IAAG,QAAQ;GAAK,GACxD,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,YAAoC,CAAC;EAC3C,MAAM,iBAA2D,CAAC;EAClE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,SAAS,QAAQ;IAChC,IAAI,MAAM,SAAS,gCAAgC,MAAM,SAAS,0BAA0B;KAC1F,IAAI,MAAM,OACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAM;KAG9C;IACF;IAEA,IAAI,MAAM,SAAS,6BAA6B;KAC9C,MAAM,EAAE,SAAS;KAEjB,IAAI,KAAK,SAAS,aAChB,eAAe,KAAK,IAAI;UACnB,IAAI,KAAK,SAAS,mBAAmB,KAAK,WAAW,cAC1D,UAAU,KAAK,KAAK,WAAW,IAAI,CAAC;KAGtC;IACF;IAEA,IAAI,MAAM,SAAS,wBAAwB,MAAM,SAAS,uBAAuB;KAC/E,WAAW,MAAM;KACjB;IACF;IAEA,IAAI,MAAM,SAAS,mBACjB,MAAM,kBAAkB,MAAM,UAAU,KAAK,IAAI;IAGnD,IAAI,MAAM,SAAS,SACjB,MAAM,gBAAgB;KAAE,MAAM,MAAM;KAAM,SAAS,MAAM;KAAS,OAAO,MAAM;IAAM,CAAC;GAE1F;GAEA,IAAI,aAAa,QACf,MAAM,IAAI,cAAc,4DAA4D,EAClF,SAAS;IAAE,UAAU,KAAK;IAAU,OAAO,KAAK;GAAK,EACvD,CAAC;EAEL,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,KAAK,MAAM,QAAQ,sBAAsB,WAAW,KAAK,MAAM,cAAc,GAC3E,MAAM;GACJ,MAAM;GACN,IAAI,KAAK;GACT,MAAM,KAAK;GACX,OAAO,KAAK;GACZ,GAAI,KAAK,mBAAmB,EAAE,kBAAkB,KAAK,iBAAiB,IAAI,CAAC;EAC7E;EAGF,MAAM,eAAe,yBAAyB;GAC5C,QAAQ,SAAS;GACjB,oBAAoB,SAAS;GAC7B,cAAc,UAAU,SAAS;EACnC,CAAC;EACD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;;;CAeA,AAAQ,aACN,UACA,SACmD;EACnD,MAAM,EAAE,cAAc,UAAU,iBAAiB,UAAU,KAAK,IAAI;EACpE,MAAM,QAAQ,iBAAiB,SAAS,KAAK;EAC7C,MAAM,aAAa,QAAQ,sBAAsB,SAAS,UAAU,IAAI;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,OAAO,KAAK;GACZ,GAAI,iBAAiB,SAAY,EAAE,aAAa,IAAI,CAAC;GACrD;GACA,OAAO;GAGP,SAAS,CAAC,6BAA6B;GACvC,GAAI,cAAc,SAAY,EAAE,mBAAmB,UAAU,IAAI,CAAC;GAClE,GAAI,CAAC,KAAK,aAAa,aAAa,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACnF,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;GACzB,GAAI,eAAe,SAAY,EAAE,aAAa,WAAW,IAAI,CAAC;GAC9D,GAAG,KAAK,gBAAgB,SAAS,cAAc;GAC/C,GAAG,KAAK,qBAAqB,SAAS,SAAS;GAC/C,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;EACtD;CACF;;;;;;;CAQA,AAAQ,gBAAgB,gBAEtB;EACA,MAAM,OAAO,qBAAqB,gBAAgB,KAAK,OAAO,cAAc;EAE5E,IAAI,KAAK,SAAS,eAChB,OAAO,EACL,MAAM,EAAE,QAAQ;GAAE,MAAM;GAAe,MAAM;GAAY,QAAQ,KAAK;GAAQ,QAAQ;EAAK,EAAE,EAC/F;EAGF,IAAI,KAAK,SAAS,eAChB,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,cAAc,EAAE,EAAE;EAGrD,OAAO,CAAC;CACV;;CAGA,AAAQ,qBAAqB,WAE3B;EACA,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WAAW,QAC9C,OAAO,CAAC;EAGV,OAAO,EAAE,WAAW,EAAE,QAAQ,wBAAwB,UAAU,MAAM,EAAE,EAAE;CAC5E;;;;;;;;;;;CAYA,AAAQ,WAAW,QAGjB;EACA,IAAI,UAAU;EACd,MAAM,YAAoC,CAAC;EAC3C,MAAM,iBAA2D,CAAC;EAElE,KAAK,MAAM,QAAQ,QAAQ;GACzB,IAAI,KAAK,SAAS,aAAa;IAC7B,eAAe,KAAK,IAAI;IACxB;GACF;GAEA,IAAI,KAAK,SAAS,WAAW;IAC3B,KAAK,MAAM,QAAQ,KAAK,SACtB,WAAW,KAAK,SAAS,YAAY,KAAK,UAAU,KAAK;IAG3D;GACF;GAEA,IAAI,KAAK,SAAS,mBAAmB,KAAK,WAAW,cACnD,UAAU,KAAK,KAAK,WAAW,IAAI,CAAC;EAExC;EAEA,OAAO;GACL;GACA,WACE,UAAU,SAAS,IACf,sBAAsB,WAAW,KAAK,MAAM,cAAc,IAC1D;EACR;CACF;;CAGA,AAAQ,WAAW,MAAuE;EACxF,OAAO;GACL,IAAI,KAAK;GACT,MAAM,KAAK;GACX,OAAO,cAAuC,KAAK,WAAW,CAAC,CAAC;EAClE;CACF;;;;;;;;CASA,AAAQ,aAAa,KAAwD;EAC3E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,OAAO,WAAW;GAChB,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,cAAc,IAAI,sBAAsB;GACxC,kBAAkB,IAAI,sBAAsB;GAC5C,iBAAiB,IAAI,uBAAuB;EAC9C,CAAC;CACH;;CAGA,AAAQ,WAAW,QAAqD;EACtE,MAAM,UAAU,gBAAgB,MAAM;EAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF"}

package/esm/sdk.d.mts

@@ -52,6 +52,13 @@ declare class OpenAISDK implements SDKAdapterContract {
    * Pricing resolution: per-model `config.pricing` wins; otherwise the
    * SDK-level registry entry keyed by `config.name`; otherwise
    * `undefined` (no cost computed).
+   *
+   * `config.api` selects the wire API. Omitted or `"chat"` builds the Chat
+   * Completions `OpenAIModel`. `"responses"` builds `OpenAIResponsesModel`
+   * and is accepted ONLY when this SDK's provider label is `"openai"`: any
+   * other label (the wrapper packages deepseek / groq / xai / mistral, or a
+   * custom gateway label) throws, so an endpoint not known to implement
+   * `/v1/responses` is never sent there.
    */
   model(config: OpenAIModelConfig): ModelContract;
   /**

package/esm/sdk.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../ai-openai/src/sdk.ts"],"mappings":";;;;;;AA2DA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,kBAAA;EAAA,iBACf,MAAA;EAAA,iBACA,QAAA;EAAA,iBACA,OAAA;cAEE,MAAA,EAAQ,eAAA;EA0EpB;;;;;;;;;;;;EAlDA,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,aAAA;EAoFuC;;;;;;;EArEnE,KAAA,CAAM,IAAA,UAAc,MAAA,YAAkB,OAAA;;;;;;;;;;;EAc5C,QAAA,CAAS,MAAA,EAAQ,oBAAA,GAAuB,gBAAA;;;;;;;;;;;;;;;;;;EAqBxC,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,kBAAA;;;;;;;;;;;;EAmBlC,MAAA,CAAO,MAAA,EAAQ,kBAAA,GAAqB,mBAAA;;;;;;;;;;;;EAepC,UAAA,CAAW,MAAA,EAAQ,yBAAA,GAA4B,0BAAA;AAAA"}
+{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../ai-openai/src/sdk.ts"],"mappings":";;;;;;AA4DA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,kBAAA;EAAA,iBACf,MAAA;EAAA,iBACA,QAAA;EAAA,iBACA,OAAA;cAEE,MAAA,EAAQ,eAAA;EAqGpB;;;;;;;;;;;;;AAkCyE;;;;;;EAxGzE,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,aAAA;;;;;;;;EAmC5B,KAAA,CAAM,IAAA,UAAc,MAAA,YAAkB,OAAA;;;;;;;;;;;EAc5C,QAAA,CAAS,MAAA,EAAQ,oBAAA,GAAuB,gBAAA;;;;;;;;;;;;;;;;;;EAqBxC,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,kBAAA;;;;;;;;;;;;EAmBlC,MAAA,CAAO,MAAA,EAAQ,kBAAA,GAAqB,mBAAA;;;;;;;;;;;;EAepC,UAAA,CAAW,MAAA,EAAQ,yBAAA,GAA4B,0BAAA;AAAA"}

package/esm/sdk.mjs

@@ -1,10 +1,11 @@
 import { OpenAIEmbedder } from "./embedder.mjs";
 import { OpenAIImageModel } from "./image.mjs";
 import { OpenAIModel } from "./model.mjs";
+import { OpenAIResponsesModel } from "./responses-model.mjs";
 import { OpenAISpeechModel } from "./speech.mjs";
 import { OpenAITranscriptionModel } from "./transcription.mjs";
 import OpenAI from "openai";
-import { approximateTokenCount } from "@warlock.js/ai";
+import { InvalidRequestError, approximateTokenCount } from "@warlock.js/ai";
 
 //#region ../ai-openai/src/sdk.ts
 /**
@@ -59,13 +60,32 @@ var OpenAISDK = class {
 	* Pricing resolution: per-model `config.pricing` wins; otherwise the
 	* SDK-level registry entry keyed by `config.name`; otherwise
 	* `undefined` (no cost computed).
+	*
+	* `config.api` selects the wire API. Omitted or `"chat"` builds the Chat
+	* Completions `OpenAIModel`. `"responses"` builds `OpenAIResponsesModel`
+	* and is accepted ONLY when this SDK's provider label is `"openai"`: any
+	* other label (the wrapper packages deepseek / groq / xai / mistral, or a
+	* custom gateway label) throws, so an endpoint not known to implement
+	* `/v1/responses` is never sent there.
 	*/
 	model(config) {
+		const api = config.api;
+		if (api !== void 0 && api !== "chat" && api !== "responses") throw new InvalidRequestError(`Unknown OpenAI api "${String(api)}"; expected "chat" or "responses".`, { context: {
+			provider: this.provider,
+			model: config.name,
+			api
+		} });
+		if (api === "responses" && this.provider !== "openai") throw new InvalidRequestError(`api: "responses" is supported only on the direct "openai" provider; this SDK is labeled "${this.provider}". Use the default api: "chat".`, { context: {
+			provider: this.provider,
+			model: config.name,
+			api
+		} });
 		const resolvedPricing = config.pricing ?? this.pricing?.[config.name];
 		const resolvedConfig = resolvedPricing === config.pricing ? config : {
 			...config,
 			pricing: resolvedPricing
 		};
+		if (api === "responses") return new OpenAIResponsesModel(this.client, resolvedConfig, this.provider);
 		return new OpenAIModel(this.client, resolvedConfig, this.provider);
 	}
 	/**

package/esm/sdk.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../ai-openai/src/sdk.ts"],"sourcesContent":["import OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n  SpeechModelContract,\n  TranscriptionModelContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  OpenAIEmbedderConfig,\n  OpenAIImageConfig,\n  OpenAIModelConfig,\n  OpenAISDKConfig,\n  OpenAISpeechConfig,\n  OpenAITranscriptionConfig,\n} from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIImageModel } from \"./image\";\nimport { OpenAIModel } from \"./model\";\nimport { OpenAISpeechModel } from \"./speech\";\nimport { OpenAITranscriptionModel } from \"./transcription\";\n\n/**\n * OpenAI-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any OpenAI-compatible provider\n * (OpenAI, Azure OpenAI, OpenRouter, local gateways speaking the Chat\n * Completions protocol). A single `OpenAISDK` instance holds one live\n * `OpenAI` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per provider/account and reuse it\n * across all agents, workflows, and supervisors that target that\n * provider.\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client (authentication, base URL) and\n *   its lifetime scope. Factory for `OpenAIModel` instances — each\n *   model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `OpenAIModel` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `OpenAI` client is heavy to construct and\n * designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the PascalCase naming convention readers\n * expect from a constructor.\n *\n * @example\n * const openai = new OpenAISDK({ apiKey: process.env.OPENAI_API_KEY! });\n * const model = openai.model({ name: \"gpt-4o\", temperature: 0.7 });\n * const tokens = await openai.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.openai` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, persona, instruction, openai: new OpenAISDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.openai.model({ name: \"gpt-4o-mini\" }) });\n */\nexport class OpenAISDK implements SDKAdapterContract {\n  private readonly client: OpenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: OpenAISDKConfig) {\n    // Peel off the framework-only keys and forward every other upstream\n    // `ClientOptions` (timeout, maxRetries, defaultHeaders, fetch,\n    // organization, project, …) verbatim — they type-check, so dropping them\n    // is a silent footgun. Mirrors the Bedrock/Google/Ollama adapters.\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.client = new OpenAI(clientOptions);\n    this.provider = provider ?? \"openai\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build an `OpenAIModel` bound to this SDK's client. Each call returns\n   * a fresh model instance, but all instances share the underlying\n   * `OpenAI` client — connection pools, rate limits, and authentication\n   * state stay unified across every model produced here. The SDK's\n   * `provider` label is forwarded so every model self-identifies as\n   * coming from the same upstream.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: OpenAIModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses a\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Accepts an optional model id for future per-model tokenizer\n   * dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build an `OpenAIEmbedder` bound to this SDK's client. Each call\n   * returns a fresh embedder instance sharing the same underlying\n   * `OpenAI` client — connection pools and authentication stay unified\n   * across every embedder produced here.\n   *\n   * @example\n   * const embedder = openai.embedder({ name: \"text-embedding-3-small\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: OpenAIEmbedderConfig): EmbedderContract {\n    return new OpenAIEmbedder(this.client, config);\n  }\n\n  /**\n   * Build an `OpenAIImageModel` bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. `config.name` is passed through to\n   * `client.images.generate` as given — no id is rejected locally, so\n   * an unsupported model fails at OpenAI, not here. Metering follows\n   * the family: `gpt-image-*` is token-metered, `dall-e-*` per-image.\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined` (no cost computed). A token-priced\n   * `gpt-image-1` entry can live in the same SDK registry as the chat\n   * models.\n   *\n   * @example\n   * const model = openai.image({ name: \"gpt-image-1\" });\n   * const { data } = await ai.image({ model, prompt: \"a red bicycle\" });\n   */\n  public image(config: OpenAIImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIImageModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Build an `OpenAISpeechModel` (text-to-speech) bound to this SDK's\n   * client, for use with `ai.speech({ model, text })`. `config.name` is\n   * passed through to `audio.speech.create` as given — no id is\n   * rejected locally, so an unsupported model fails at OpenAI, not\n   * here.\n   *\n   * @example\n   * const tts = openai.speech({ name: \"tts-1\", voice: \"alloy\" });\n   * const { data } = await ai.speech({ model: tts, text: \"Hello\" });\n   */\n  public speech(config: OpenAISpeechConfig): SpeechModelContract {\n    return new OpenAISpeechModel(this.client, config, this.provider);\n  }\n\n  /**\n   * Build an `OpenAITranscriptionModel` (speech-to-text) bound to this\n   * SDK's client, for use with `ai.transcribe({ model, audio })`.\n   * `config.name` is passed through to `audio.transcriptions.create`\n   * as given — no id is rejected locally, so an unsupported model\n   * fails at OpenAI, not here.\n   *\n   * @example\n   * const stt = openai.transcribe({ name: \"whisper-1\" });\n   * const { data } = await ai.transcribe({ model: stt, audio });\n   */\n  public transcribe(config: OpenAITranscriptionConfig): TranscriptionModelContract {\n    return new OpenAITranscriptionModel(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2DA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAK1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,SAAS,IAAI,OAAO,aAAa;EACtC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;;;CAcA,AAAO,MAAM,QAA0C;EACrD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,YAAY,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACnE;;;;;;;;CASA,MAAa,MAAM,MAAc,QAAkC;EACjE,OAAO,sBAAsB,IAAI;CACnC;;;;;;;;;;;CAYA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,QAAQ,MAAM;CAC/C;;;;;;;;;;;;;;;;;;CAmBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,iBAAiB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACxE;;;;;;;;;;;;CAaA,AAAO,OAAO,QAAiD;EAC7D,OAAO,IAAI,kBAAkB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACjE;;;;;;;;;;;;CAaA,AAAO,WAAW,QAA+D;EAC/E,OAAO,IAAI,yBAAyB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACxE;AACF"}
+{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../ai-openai/src/sdk.ts"],"sourcesContent":["import OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n  SpeechModelContract,\n  TranscriptionModelContract,\n} from \"@warlock.js/ai\";\nimport { InvalidRequestError, approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  OpenAIEmbedderConfig,\n  OpenAIImageConfig,\n  OpenAIModelConfig,\n  OpenAISDKConfig,\n  OpenAISpeechConfig,\n  OpenAITranscriptionConfig,\n} from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIImageModel } from \"./image\";\nimport { OpenAIModel } from \"./model\";\nimport { OpenAIResponsesModel } from \"./responses-model\";\nimport { OpenAISpeechModel } from \"./speech\";\nimport { OpenAITranscriptionModel } from \"./transcription\";\n\n/**\n * OpenAI-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any OpenAI-compatible provider\n * (OpenAI, Azure OpenAI, OpenRouter, local gateways speaking the Chat\n * Completions protocol). A single `OpenAISDK` instance holds one live\n * `OpenAI` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per provider/account and reuse it\n * across all agents, workflows, and supervisors that target that\n * provider.\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client (authentication, base URL) and\n *   its lifetime scope. Factory for `OpenAIModel` instances — each\n *   model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `OpenAIModel` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `OpenAI` client is heavy to construct and\n * designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the PascalCase naming convention readers\n * expect from a constructor.\n *\n * @example\n * const openai = new OpenAISDK({ apiKey: process.env.OPENAI_API_KEY! });\n * const model = openai.model({ name: \"gpt-4o\", temperature: 0.7 });\n * const tokens = await openai.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.openai` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, persona, instruction, openai: new OpenAISDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.openai.model({ name: \"gpt-4o-mini\" }) });\n */\nexport class OpenAISDK implements SDKAdapterContract {\n  private readonly client: OpenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: OpenAISDKConfig) {\n    // Peel off the framework-only keys and forward every other upstream\n    // `ClientOptions` (timeout, maxRetries, defaultHeaders, fetch,\n    // organization, project, …) verbatim — they type-check, so dropping them\n    // is a silent footgun. Mirrors the Bedrock/Google/Ollama adapters.\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.client = new OpenAI(clientOptions);\n    this.provider = provider ?? \"openai\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build an `OpenAIModel` bound to this SDK's client. Each call returns\n   * a fresh model instance, but all instances share the underlying\n   * `OpenAI` client — connection pools, rate limits, and authentication\n   * state stay unified across every model produced here. The SDK's\n   * `provider` label is forwarded so every model self-identifies as\n   * coming from the same upstream.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   *\n   * `config.api` selects the wire API. Omitted or `\"chat\"` builds the Chat\n   * Completions `OpenAIModel`. `\"responses\"` builds `OpenAIResponsesModel`\n   * and is accepted ONLY when this SDK's provider label is `\"openai\"`: any\n   * other label (the wrapper packages deepseek / groq / xai / mistral, or a\n   * custom gateway label) throws, so an endpoint not known to implement\n   * `/v1/responses` is never sent there.\n   */\n  public model(config: OpenAIModelConfig): ModelContract {\n    const api: unknown = config.api;\n\n    if (api !== undefined && api !== \"chat\" && api !== \"responses\") {\n      throw new InvalidRequestError(\n        `Unknown OpenAI api \"${String(api)}\"; expected \"chat\" or \"responses\".`,\n        { context: { provider: this.provider, model: config.name, api } },\n      );\n    }\n\n    if (api === \"responses\" && this.provider !== \"openai\") {\n      throw new InvalidRequestError(\n        `api: \"responses\" is supported only on the direct \"openai\" provider; this SDK is labeled \"${this.provider}\". Use the default api: \"chat\".`,\n        { context: { provider: this.provider, model: config.name, api } },\n      );\n    }\n\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    if (api === \"responses\") {\n      return new OpenAIResponsesModel(this.client, resolvedConfig, this.provider);\n    }\n\n    return new OpenAIModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses a\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Accepts an optional model id for future per-model tokenizer\n   * dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build an `OpenAIEmbedder` bound to this SDK's client. Each call\n   * returns a fresh embedder instance sharing the same underlying\n   * `OpenAI` client — connection pools and authentication stay unified\n   * across every embedder produced here.\n   *\n   * @example\n   * const embedder = openai.embedder({ name: \"text-embedding-3-small\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: OpenAIEmbedderConfig): EmbedderContract {\n    return new OpenAIEmbedder(this.client, config);\n  }\n\n  /**\n   * Build an `OpenAIImageModel` bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. `config.name` is passed through to\n   * `client.images.generate` as given — no id is rejected locally, so\n   * an unsupported model fails at OpenAI, not here. Metering follows\n   * the family: `gpt-image-*` is token-metered, `dall-e-*` per-image.\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined` (no cost computed). A token-priced\n   * `gpt-image-1` entry can live in the same SDK registry as the chat\n   * models.\n   *\n   * @example\n   * const model = openai.image({ name: \"gpt-image-1\" });\n   * const { data } = await ai.image({ model, prompt: \"a red bicycle\" });\n   */\n  public image(config: OpenAIImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIImageModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Build an `OpenAISpeechModel` (text-to-speech) bound to this SDK's\n   * client, for use with `ai.speech({ model, text })`. `config.name` is\n   * passed through to `audio.speech.create` as given — no id is\n   * rejected locally, so an unsupported model fails at OpenAI, not\n   * here.\n   *\n   * @example\n   * const tts = openai.speech({ name: \"tts-1\", voice: \"alloy\" });\n   * const { data } = await ai.speech({ model: tts, text: \"Hello\" });\n   */\n  public speech(config: OpenAISpeechConfig): SpeechModelContract {\n    return new OpenAISpeechModel(this.client, config, this.provider);\n  }\n\n  /**\n   * Build an `OpenAITranscriptionModel` (speech-to-text) bound to this\n   * SDK's client, for use with `ai.transcribe({ model, audio })`.\n   * `config.name` is passed through to `audio.transcriptions.create`\n   * as given — no id is rejected locally, so an unsupported model\n   * fails at OpenAI, not here.\n   *\n   * @example\n   * const stt = openai.transcribe({ name: \"whisper-1\" });\n   * const { data } = await ai.transcribe({ model: stt, audio });\n   */\n  public transcribe(config: OpenAITranscriptionConfig): TranscriptionModelContract {\n    return new OpenAITranscriptionModel(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4DA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAK1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,SAAS,IAAI,OAAO,aAAa;EACtC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;;;;;;;;;;CAqBA,AAAO,MAAM,QAA0C;EACrD,MAAM,MAAe,OAAO;EAE5B,IAAI,QAAQ,UAAa,QAAQ,UAAU,QAAQ,aACjD,MAAM,IAAI,oBACR,uBAAuB,OAAO,GAAG,EAAE,qCACnC,EAAE,SAAS;GAAE,UAAU,KAAK;GAAU,OAAO,OAAO;GAAM;EAAI,EAAE,CAClE;EAGF,IAAI,QAAQ,eAAe,KAAK,aAAa,UAC3C,MAAM,IAAI,oBACR,4FAA4F,KAAK,SAAS,kCAC1G,EAAE,SAAS;GAAE,UAAU,KAAK;GAAU,OAAO,OAAO;GAAM;EAAI,EAAE,CAClE;EAGF,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,IAAI,QAAQ,aACV,OAAO,IAAI,qBAAqB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;EAG5E,OAAO,IAAI,YAAY,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACnE;;;;;;;;CASA,MAAa,MAAM,MAAc,QAAkC;EACjE,OAAO,sBAAsB,IAAI;CACnC;;;;;;;;;;;CAYA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,QAAQ,MAAM;CAC/C;;;;;;;;;;;;;;;;;;CAmBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,iBAAiB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACxE;;;;;;;;;;;;CAaA,AAAO,OAAO,QAAiD;EAC7D,OAAO,IAAI,kBAAkB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACjE;;;;;;;;;;;;CAaA,AAAO,WAAW,QAA+D;EAC/E,OAAO,IAAI,yBAAyB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACxE;AACF"}

package/esm/utils/build-usage.mjs

@@ -0,0 +1,23 @@
+//#region ../ai-openai/src/utils/build-usage.ts
+/**
+* Build the neutral `Usage` from extracted counts. The optional detail
+* counters are emitted only when the provider reports a positive value, so
+* non-reasoning / uncached calls keep the lean `{ input, output, total }`
+* shape. Shared by the Chat Completions and Responses adapters so both
+* apply the same rule.
+*/
+function buildUsage(counts) {
+	const { input, output, total, cachedTokens, reasoningTokens, cacheWriteTokens } = counts;
+	return {
+		input,
+		output,
+		total,
+		...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {},
+		...reasoningTokens !== void 0 && reasoningTokens > 0 ? { reasoningTokens } : {},
+		...cacheWriteTokens !== void 0 && cacheWriteTokens > 0 ? { cacheWriteTokens } : {}
+	};
+}
+
+//#endregion
+export { buildUsage };
+//# sourceMappingURL=build-usage.mjs.map

package/esm/utils/build-usage.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"build-usage.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/build-usage.ts"],"sourcesContent":["import type { Usage } from \"@warlock.js/ai\";\n\n/** The already-extracted token counts of one provider usage block. */\nexport type UsageCounts = {\n  input: number;\n  output: number;\n  total: number;\n  /** Subset of `input` served from the prompt cache. */\n  cachedTokens?: number;\n  /** Hidden reasoning channel, already counted within `output`. */\n  reasoningTokens?: number;\n  /** Subset of `input` written to the prompt cache. */\n  cacheWriteTokens?: number;\n};\n\n/**\n * Build the neutral `Usage` from extracted counts. The optional detail\n * counters are emitted only when the provider reports a positive value, so\n * non-reasoning / uncached calls keep the lean `{ input, output, total }`\n * shape. Shared by the Chat Completions and Responses adapters so both\n * apply the same rule.\n */\nexport function buildUsage(counts: UsageCounts): Usage {\n  const { input, output, total, cachedTokens, reasoningTokens, cacheWriteTokens } = counts;\n\n  return {\n    input,\n    output,\n    total,\n    ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n    ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\n    ...(cacheWriteTokens !== undefined && cacheWriteTokens > 0 ? { cacheWriteTokens } : {}),\n  };\n}\n"],"mappings":";;;;;;;;AAsBA,SAAgB,WAAW,QAA4B;CACrD,MAAM,EAAE,OAAO,QAAQ,OAAO,cAAc,iBAAiB,qBAAqB;CAElF,OAAO;EACL;EACA;EACA;EACA,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;EACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EAClF,GAAI,qBAAqB,UAAa,mBAAmB,IAAI,EAAE,iBAAiB,IAAI,CAAC;CACvF;AACF"}

package/esm/utils/index.mjs

@@ -1,6 +1,14 @@
+import { buildUsage } from "./build-usage.mjs";
+import { wrapOpenAIError } from "./wrap-openai-error.mjs";
+import { mapResponsesFinishReason, toResponseFailure } from "./map-responses-status.mjs";
 import { mapFinishReason } from "./map-finish-reason.mjs";
+import { buildPromptCacheParams } from "./prompt-cache-params.mjs";
+import { toOpenAIReasoningEffort } from "./reasoning-effort.mjs";
+import { inferStructuredOutput, planStructuredOutput } from "./structured-output.mjs";
+import { RESPONSES_METADATA_KEY, attachReasoningReplay, pickReplayableReasoning, readReasoningReplay, toReplayReasoningItem } from "./reasoning-replay.mjs";
 import { toOpenAIMessages } from "./to-openai-messages.mjs";
 import { toOpenAITools } from "./to-openai-tools.mjs";
-import { wrapOpenAIError } from "./wrap-openai-error.mjs";
+import { toResponsesInput } from "./to-responses-input.mjs";
+import { toResponsesToolChoice, toResponsesTools } from "./to-responses-tools.mjs";
 
 export {  };

package/esm/utils/map-responses-status.mjs

@@ -0,0 +1,54 @@
+import { wrapOpenAIError } from "./wrap-openai-error.mjs";
+
+//#region ../ai-openai/src/utils/map-responses-status.ts
+/**
+* Derive our `FinishReason` from a Responses terminal state. Responses has
+* no `finish_reason`, so it comes from `status`, `incomplete_details` and the
+* output items (design section 2, decision 5):
+*
+* - `completed` with a function call -> `tool_calls`; without -> `stop`
+*   (a refusal is ordinary text and also `stop`).
+* - `incomplete` / `max_output_tokens` and `max_messages` -> `length`.
+* - `incomplete` / `content_filter` and `steered` (or an unknown / missing
+*   reason) -> `error`. There is no neutral `content_filter` value.
+* - `cancelled`, `queued`, `in_progress` at a terminal read -> `error`.
+* - `failed` is not mapped here: callers throw it via {@link toResponseFailure}.
+*
+* A missing `status` is treated as `completed` (the SDK types it optional).
+*/
+function mapResponsesFinishReason(state) {
+	const status = state.status ?? "completed";
+	if (status === "completed") return state.hasToolCalls ? "tool_calls" : "stop";
+	if (status === "incomplete") {
+		const reason = state.incomplete_details?.reason;
+		return reason === "max_output_tokens" || reason === "max_messages" ? "length" : "error";
+	}
+	return "error";
+}
+/**
+* Turn a `failed` Responses payload into the same typed `AIError` the Chat
+* path throws for a provider failure, by running it through
+* `wrapOpenAIError` (so `rate_limit_exceeded` becomes a
+* `ProviderRateLimitError`, everything else a `ProviderError`, with `code`
+* and request id kept on `context`). The response id and model are added so
+* the failure is traceable.
+*/
+function toResponseFailure(response, model) {
+	const error = response.error;
+	const requestId = response._request_id;
+	const wrapped = wrapOpenAIError({
+		code: error?.code,
+		message: error?.message ?? "OpenAI response failed without an error object.",
+		...requestId ? { request_id: requestId } : {}
+	});
+	if (wrapped.context) Object.assign(wrapped.context, {
+		provider: "openai",
+		model,
+		responseId: response.id
+	});
+	return wrapped;
+}
+
+//#endregion
+export { mapResponsesFinishReason, toResponseFailure };
+//# sourceMappingURL=map-responses-status.mjs.map

package/esm/utils/map-responses-status.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"map-responses-status.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/map-responses-status.ts"],"sourcesContent":["import type { AIError, FinishReason } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\nimport { wrapOpenAIError } from \"./wrap-openai-error\";\n\n/** The slice of a Responses `Response` the status mapper reads. */\nexport type ResponsesTerminalState = Pick<\n  OpenAI.Responses.Response,\n  \"status\" | \"incomplete_details\"\n> & {\n  /** True when `output` holds at least one `function_call` item the model completed. */\n  hasToolCalls: boolean;\n};\n\n/**\n * Derive our `FinishReason` from a Responses terminal state. Responses has\n * no `finish_reason`, so it comes from `status`, `incomplete_details` and the\n * output items (design section 2, decision 5):\n *\n * - `completed` with a function call -> `tool_calls`; without -> `stop`\n *   (a refusal is ordinary text and also `stop`).\n * - `incomplete` / `max_output_tokens` and `max_messages` -> `length`.\n * - `incomplete` / `content_filter` and `steered` (or an unknown / missing\n *   reason) -> `error`. There is no neutral `content_filter` value.\n * - `cancelled`, `queued`, `in_progress` at a terminal read -> `error`.\n * - `failed` is not mapped here: callers throw it via {@link toResponseFailure}.\n *\n * A missing `status` is treated as `completed` (the SDK types it optional).\n */\nexport function mapResponsesFinishReason(state: ResponsesTerminalState): FinishReason {\n  const status = state.status ?? \"completed\";\n\n  if (status === \"completed\") {\n    return state.hasToolCalls ? \"tool_calls\" : \"stop\";\n  }\n\n  if (status === \"incomplete\") {\n    const reason = state.incomplete_details?.reason;\n\n    return reason === \"max_output_tokens\" || reason === \"max_messages\" ? \"length\" : \"error\";\n  }\n\n  return \"error\";\n}\n\n/**\n * Turn a `failed` Responses payload into the same typed `AIError` the Chat\n * path throws for a provider failure, by running it through\n * `wrapOpenAIError` (so `rate_limit_exceeded` becomes a\n * `ProviderRateLimitError`, everything else a `ProviderError`, with `code`\n * and request id kept on `context`). The response id and model are added so\n * the failure is traceable.\n */\nexport function toResponseFailure(\n  response: OpenAI.Responses.Response,\n  model: string,\n): AIError {\n  const error = response.error;\n  const requestId = (response as { _request_id?: string | null })._request_id;\n\n  const wrapped = wrapOpenAIError({\n    code: error?.code,\n    message: error?.message ?? \"OpenAI response failed without an error object.\",\n    ...(requestId ? { request_id: requestId } : {}),\n  });\n\n  if (wrapped.context) {\n    Object.assign(wrapped.context, { provider: \"openai\", model, responseId: response.id });\n  }\n\n  return wrapped;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AA4BA,SAAgB,yBAAyB,OAA6C;CACpF,MAAM,SAAS,MAAM,UAAU;CAE/B,IAAI,WAAW,aACb,OAAO,MAAM,eAAe,eAAe;CAG7C,IAAI,WAAW,cAAc;EAC3B,MAAM,SAAS,MAAM,oBAAoB;EAEzC,OAAO,WAAW,uBAAuB,WAAW,iBAAiB,WAAW;CAClF;CAEA,OAAO;AACT;;;;;;;;;AAUA,SAAgB,kBACd,UACA,OACS;CACT,MAAM,QAAQ,SAAS;CACvB,MAAM,YAAa,SAA6C;CAEhE,MAAM,UAAU,gBAAgB;EAC9B,MAAM,OAAO;EACb,SAAS,OAAO,WAAW;EAC3B,GAAI,YAAY,EAAE,YAAY,UAAU,IAAI,CAAC;CAC/C,CAAC;CAED,IAAI,QAAQ,SACV,OAAO,OAAO,QAAQ,SAAS;EAAE,UAAU;EAAU;EAAO,YAAY,SAAS;CAAG,CAAC;CAGvF,OAAO;AACT"}

package/esm/utils/prompt-cache-params.mjs

@@ -0,0 +1,17 @@
+//#region ../ai-openai/src/utils/prompt-cache-params.ts
+/**
+* Build the OpenAI-only prompt-cache request fields from the model config.
+* Returns nothing for any provider label other than `"openai"`: wrapper
+* endpoints use the same transport but may reject these OpenAI-only fields.
+*/
+function buildPromptCacheParams(provider, config) {
+	if (provider !== "openai") return {};
+	return {
+		...config.promptCacheKey !== void 0 ? { prompt_cache_key: config.promptCacheKey } : {},
+		...config.promptCacheRetention !== void 0 ? { prompt_cache_retention: config.promptCacheRetention } : {}
+	};
+}
+
+//#endregion
+export { buildPromptCacheParams };
+//# sourceMappingURL=prompt-cache-params.mjs.map

package/esm/utils/prompt-cache-params.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt-cache-params.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/prompt-cache-params.ts"],"sourcesContent":["import type { OpenAIModelConfig } from \"../config.type\";\n\n/** The OpenAI-only prompt-cache request fields common to Chat Completions and Responses. */\nexport type PromptCacheParams = {\n  prompt_cache_key?: string;\n  prompt_cache_retention?: NonNullable<OpenAIModelConfig[\"promptCacheRetention\"]>;\n};\n\n/**\n * Build the OpenAI-only prompt-cache request fields from the model config.\n * Returns nothing for any provider label other than `\"openai\"`: wrapper\n * endpoints use the same transport but may reject these OpenAI-only fields.\n */\nexport function buildPromptCacheParams(\n  provider: string,\n  config: Pick<OpenAIModelConfig, \"promptCacheKey\" | \"promptCacheRetention\">,\n): PromptCacheParams {\n  if (provider !== \"openai\") {\n    return {};\n  }\n\n  return {\n    ...(config.promptCacheKey !== undefined ? { prompt_cache_key: config.promptCacheKey } : {}),\n    ...(config.promptCacheRetention !== undefined\n      ? { prompt_cache_retention: config.promptCacheRetention }\n      : {}),\n  };\n}\n"],"mappings":";;;;;;AAaA,SAAgB,uBACd,UACA,QACmB;CACnB,IAAI,aAAa,UACf,OAAO,CAAC;CAGV,OAAO;EACL,GAAI,OAAO,mBAAmB,SAAY,EAAE,kBAAkB,OAAO,eAAe,IAAI,CAAC;EACzF,GAAI,OAAO,yBAAyB,SAChC,EAAE,wBAAwB,OAAO,qBAAqB,IACtD,CAAC;CACP;AACF"}