@warlock.js/ai-openai 5.29.0 → 5.31.0

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../ai-openai/src/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  buildPromptCacheParams,\n  buildUsage,\n  inferStructuredOutput,\n  mapFinishReason,\n  planStructuredOutput,\n  toOpenAIMessages,\n  toOpenAIReasoningEffort,\n  toOpenAITools,\n  wrapOpenAIError,\n} from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * OpenAI-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `openai` SDK. Agents,\n * workflows, and supervisors never talk to OpenAI directly — they hold a\n * `ModelContract`, and this class is what makes that contract concrete for\n * any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, OpenRouter, local\n * gateways that speak the Chat Completions protocol).\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client + frozen `ModelConfig` (name,\n *   temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` and\n *   `ToolContract[]` into OpenAI wire shapes on the way out, and\n *   translating OpenAI's response (content, finish reason, tool calls,\n *   usage) back into the neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as a\n * class (see §4.2 of code-style.md — \"long-lived state across calls\").\n *\n * @example\n * import OpenAI from \"openai\";\n * const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n * const model = new OpenAIModel(client, { name: \"gpt-4o\", temperature: 0.3 });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class OpenAIModel 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    // `api: \"responses\"` belongs to `OpenAIResponsesModel`, which only the\n    // `OpenAISDK.model()` factory constructs. This class speaks Chat\n    // Completions only; accepting the flag here would silently ignore it.\n    if (config.api === \"responses\") {\n      throw new InvalidRequestError(\n        'OpenAIModel speaks Chat Completions only and cannot serve api: \"responses\". Use OpenAISDK.model({ api: \"responses\" }) on the direct OpenAI provider.',\n        { context: { provider, model: config.name, api: config.api } },\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      // o-series + gpt-5 / gpt-6 models surface a reasoning channel and accept\n      // the `reasoning_effort` param. Explicit config wins over the\n      // name-prefix inference.\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      // Direct OpenAI requests support automatic cache hits and explicit\n      // content-part breakpoints. Wrapper endpoints have distinct provider\n      // labels and may reject these OpenAI-only request fields.\n      promptCaching: this.provider === \"openai\",\n      // PDF + audio INPUT are off by default — OpenAI accepts `file`\n      // (PDF) and `input_audio` parts only on specific models, so the\n      // flags are conservative/honest and opt-in via config rather than\n      // name-inferred. When set, the agent admits the attachments and\n      // `toOpenAIMessages` maps them to the real wire parts.\n      pdf: config.pdf ?? false,\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Chat\n   * Completions endpoint, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    // Per-call request/response logs are hot-path in production agents\n    // — keep them at `debug` so `info` stays reserved for lifecycle\n    // events (agent starting/completed, etc.). Operators who need to\n    // audit every LLM call can raise log-level at runtime.\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: OpenAI.Chat.Completions.ChatCompletion;\n\n    try {\n      response = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n          ...this.buildSamplingParams(options),\n          ...buildPromptCacheParams(this.provider, this.config),\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    const [choice] = response.choices;\n\n    if (choice === undefined) {\n      // A 200 carrying an empty `choices` array. Every line below reads off\n      // `choice`, so without this the caller got a bare `TypeError: Cannot\n      // read properties of undefined` — no provider, no model, no hint that\n      // the response itself was the problem rather than our handling of it.\n      //\n      // `ProviderError` because that is what this is: the provider answered\n      // successfully and returned nothing usable. Callers already branch on\n      // `instanceof ProviderError` for any provider-side failure.\n      throw new ProviderError(\"OpenAI returned a response with no choices.\", {\n        context: { provider: \"openai\", model: this.name, choices: response.choices.length },\n      });\n    }\n\n    const finishReason = mapFinishReason(choice.finish_reason);\n    const usage = this.extractUsage(response.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: choice.message.content ?? \"\",\n      finishReason,\n      usage,\n      toolCalls: this.extractToolCalls(choice.message.tool_calls),\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s —\n   * `delta` for text tokens, `tool-call` when the model requests a tool,\n   * and a terminal `done` carrying the final finish reason + usage totals.\n   * Callers consume it with `for await`.\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 chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<typeof this.client.chat.completions.create>>;\n\n    try {\n      stream = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n          ...this.buildSamplingParams(options),\n          ...buildPromptCacheParams(this.provider, this.config),\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    let rawFinishReason: string = \"stop\";\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolCallAccum = new Map<number, { id: string; name: string; arguments: string }>();\n\n    try {\n      for await (const chunk of stream as AsyncIterable<OpenAI.Chat.Completions.ChatCompletionChunk>) {\n        const delta = chunk.choices[0]?.delta;\n        const finish = chunk.choices[0]?.finish_reason;\n\n        if (delta?.content) {\n          yield { type: \"delta\", content: delta.content };\n        }\n\n        if (delta?.tool_calls) {\n          for (const toolCall of delta.tool_calls) {\n            const idx = toolCall.index ?? 0;\n            if (!toolCallAccum.has(idx)) {\n              toolCallAccum.set(idx, { id: \"\", name: \"\", arguments: \"\" });\n            }\n            const acc = toolCallAccum.get(idx)!;\n            if (toolCall.id) acc.id = toolCall.id;\n            if (toolCall.function?.name) acc.name = toolCall.function.name;\n            if (toolCall.function?.arguments) acc.arguments += toolCall.function.arguments;\n          }\n        }\n\n        if (finish) {\n          rawFinishReason = finish;\n        }\n\n        if (chunk.usage) {\n          usage.input = chunk.usage.prompt_tokens ?? 0;\n          usage.output = chunk.usage.completion_tokens ?? 0;\n          usage.total = chunk.usage.total_tokens ?? 0;\n          const cached = chunk.usage.prompt_tokens_details?.cached_tokens;\n          if (cached !== undefined && cached > 0) {\n            usage.cachedTokens = cached;\n          }\n          const reasoning = chunk.usage.completion_tokens_details?.reasoning_tokens;\n          if (reasoning !== undefined && reasoning > 0) {\n            usage.reasoningTokens = reasoning;\n          }\n        }\n      }\n\n      for (const acc of toolCallAccum.values()) {\n        // Skip accumulators that never received a function name — those\n        // are partial fragments the model started but never identified\n        // (e.g. arguments-only deltas with no originating `id`/`name`).\n        // Yielding them produces nameless tool-calls the agent runtime\n        // can't dispatch and would mis-attribute as a registered tool.\n        if (!acc.name) continue;\n\n        yield {\n          type: \"tool-call\",\n          id: acc.id,\n          name: acc.name,\n          input: safeJsonParse<Record<string, unknown>>(acc.arguments, {}),\n        };\n      }\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    const finishReason = mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /** Map the neutral cache hint only for the direct OpenAI provider. */\n  private promptCacheBreakpoints(options: ModelCallOptions | undefined): number {\n    return this.provider === \"openai\" ? (options?.cacheControl?.breakpoints ?? 0) : 0;\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter. The mode decision (override handling,\n   * strict-compatibility downgrade to `json_object`) lives in\n   * `planStructuredOutput`, shared with the Responses adapter.\n   *\n   * Returns an empty spread when no schema was supplied, so the caller\n   * can unconditionally `...buildResponseFormat(...)` into the request.\n   */\n  private buildResponseFormat(responseSchema: Record<string, unknown> | undefined): {\n    response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"response_format\"];\n  } {\n    const plan = planStructuredOutput(responseSchema, this.config.responseFormat);\n\n    if (plan.mode === \"json_schema\") {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: { name: \"response\", schema: plan.schema, strict: true },\n        },\n      };\n    }\n\n    if (plan.mode === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    return {};\n  }\n\n  /**\n   * Normalize OpenAI's `usage` block (which may be absent on some responses\n   * or partials) into the neutral `Usage` shape. Missing usage collapses to\n   * zeros rather than propagating `undefined`, so downstream aggregation\n   * math stays safe.\n   *\n   * `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the\n   * subset of the prompt served from OpenAI's automatic prompt cache);\n   * `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`\n   * (the hidden reasoning channel on o-series / gpt-5 models, already\n   * counted within `output`). Both are emitted only when the provider\n   * reports a positive value, so non-reasoning / uncached calls keep the\n   * lean `{ input, output, total }` shape.\n   */\n  private extractUsage(raw: OpenAI.Completions.CompletionUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    return buildUsage({\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      cachedTokens: raw.prompt_tokens_details?.cached_tokens,\n      reasoningTokens: raw.completion_tokens_details?.reasoning_tokens,\n    });\n  }\n\n  /**\n   * Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's\n   * `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat\n   * Completions API exposes a discrete effort knob, not a token budget,\n   * so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no\n   * wire equivalent here and is silently ignored.\n   *\n   * The neutral `ReasoningEffort` (`\"low\" | \"medium\" | \"high\" | \"none\"`)\n   * is a subset of OpenAI's accepted values, so it forwards verbatim —\n   * `\"none\"` included. `\"none\"` is load-bearing: gpt-5 / o-series models\n   * **reject function tools** on Chat Completions while reasoning is\n   * active, and the endpoint accepts tools only when `reasoning_effort`\n   * is `\"none\"` (the alternative is the Responses API). This is why the\n   * param is EMITTED for `\"none\"` rather than omitted — omitting it\n   * leaves the model reasoning server-side by default, so tools would\n   * still be rejected.\n   *\n   * No-ops when the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`).\n   *\n   * When the caller gave no explicit `effort`: defaults to `\"none\"` IF\n   * `hasTools` is true, else omits the param (provider default — the\n   * pre-existing behavior for a tool-less call). The default exists\n   * because, on a reasoning-capable model, there is no working alternative\n   * to `\"none\"` when tools are attached — omitting `reasoning_effort`\n   * leaves reasoning on server-side, which Chat Completions rejects tools\n   * for (empty replies on some model generations, a hard 400 — \"Function\n   * tools with reasoning_effort are not supported ... in\n   * /v1/chat/completions\" — on newer ones). An explicit `effort` from the\n   * caller always wins over this default, in either direction.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  /**\n   * Output-length and sampling params, in current Chat Completions shapes only.\n   *\n   * - `max_completion_tokens`, never the legacy `max_tokens`: current\n   *   models (the gpt-5 family and later) answer `400 unsupported_parameter`\n   *   to `max_tokens`, and `max_completion_tokens` is accepted everywhere.\n   * - No `temperature` for a reasoning-capable model: those accept only the\n   *   default value and reject any other with a 400.\n   */\n  private buildSamplingParams(options: ModelCallOptions | undefined): {\n    max_completion_tokens?: number;\n    temperature?: number;\n  } {\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      max_completion_tokens: maxTokens,\n      ...(!this.capabilities.reasoning && { temperature }),\n    };\n  }\n\n  private buildReasoningParams(\n    reasoning: ModelCallOptions[\"reasoning\"],\n    hasTools: boolean,\n  ): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning) {\n      return {};\n    }\n\n    if (!reasoning?.effort) {\n      return hasTools ? { reasoning_effort: \"none\" } : {};\n    }\n\n    return { reasoning_effort: toOpenAIReasoningEffort(reasoning.effort) };\n  }\n\n  /**\n   * Reshape OpenAI's `tool_calls` array into the neutral\n   * `ModelToolCallRequest[]`. The raw `arguments` field is a JSON string\n   * per OpenAI's protocol — we parse it defensively via `safeJsonParse` so\n   * malformed or empty arguments yield an empty object instead of crashing\n   * the trip. Returns `undefined` when no tools were requested so callers\n   * can branch on presence.\n   */\n  private extractToolCalls(\n    rawToolCalls: OpenAI.Chat.Completions.ChatCompletionMessageToolCall[] | undefined,\n  ): ModelToolCallRequest[] | undefined {\n    if (!rawToolCalls || rawToolCalls.length === 0) {\n      return undefined;\n    }\n\n    return rawToolCalls.map((toolCall) => ({\n      id: toolCall.id,\n      name: (toolCall as any).function.name,\n      input: safeJsonParse<Record<string, unknown>>((toolCall as any).function.arguments, {}),\n    }));\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;AA+BA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCnB,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAMhC,IAAI,OAAO,QAAQ,aACjB,MAAM,IAAI,oBACR,4JACA,EAAE,SAAS;GAAE;GAAU,OAAO,OAAO;GAAM,KAAK,OAAO;EAAI,EAAE,CAC/D;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;GAI1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GAInE,eAAe,KAAK,aAAa;GAMjC,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,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,KAAK,YAAY,OAC5C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,CAAC,UAAU,SAAS;EAE1B,IAAI,WAAW,QASb,MAAM,IAAI,cAAc,+CAA+C,EACrE,SAAS;GAAE,UAAU;GAAU,OAAO,KAAK;GAAM,SAAS,SAAS,QAAQ;EAAO,EACpF,CAAC;EAGH,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,sCAAsC;GAC9E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,OAAO,QAAQ,WAAW;GACnC;GACA;GACA,WAAW,KAAK,iBAAiB,OAAO,QAAQ,UAAU;EAC5D;CACF;;;;;;;CAQA,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,KAAK,YAAY,OAC1C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,IAAI,kBAA0B;EAC9B,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,gCAAgB,IAAI,IAA6D;EAEvF,IAAI;GACF,WAAW,MAAM,SAAS,QAAsE;IAC9F,MAAM,QAAQ,MAAM,QAAQ,EAAE,EAAE;IAChC,MAAM,SAAS,MAAM,QAAQ,EAAE,EAAE;IAEjC,IAAI,OAAO,SACT,MAAM;KAAE,MAAM;KAAS,SAAS,MAAM;IAAQ;IAGhD,IAAI,OAAO,YACT,KAAK,MAAM,YAAY,MAAM,YAAY;KACvC,MAAM,MAAM,SAAS,SAAS;KAC9B,IAAI,CAAC,cAAc,IAAI,GAAG,GACxB,cAAc,IAAI,KAAK;MAAE,IAAI;MAAI,MAAM;MAAI,WAAW;KAAG,CAAC;KAE5D,MAAM,MAAM,cAAc,IAAI,GAAG;KACjC,IAAI,SAAS,IAAI,IAAI,KAAK,SAAS;KACnC,IAAI,SAAS,UAAU,MAAM,IAAI,OAAO,SAAS,SAAS;KAC1D,IAAI,SAAS,UAAU,WAAW,IAAI,aAAa,SAAS,SAAS;IACvE;IAGF,IAAI,QACF,kBAAkB;IAGpB,IAAI,MAAM,OAAO;KACf,MAAM,QAAQ,MAAM,MAAM,iBAAiB;KAC3C,MAAM,SAAS,MAAM,MAAM,qBAAqB;KAChD,MAAM,QAAQ,MAAM,MAAM,gBAAgB;KAC1C,MAAM,SAAS,MAAM,MAAM,uBAAuB;KAClD,IAAI,WAAW,UAAa,SAAS,GACnC,MAAM,eAAe;KAEvB,MAAM,YAAY,MAAM,MAAM,2BAA2B;KACzD,IAAI,cAAc,UAAa,YAAY,GACzC,MAAM,kBAAkB;IAE5B;GACF;GAEA,KAAK,MAAM,OAAO,cAAc,OAAO,GAAG;IAMxC,IAAI,CAAC,IAAI,MAAM;IAEf,MAAM;KACJ,MAAM;KACN,IAAI,IAAI;KACR,MAAM,IAAI;KACV,OAAO,cAAuC,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAM,YAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;CAGA,AAAQ,uBAAuB,SAA+C;EAC5E,OAAO,KAAK,aAAa,WAAY,SAAS,cAAc,eAAe,IAAK;CAClF;;;;;;;;;;CAWA,AAAQ,oBAAoB,gBAE1B;EACA,MAAM,OAAO,qBAAqB,gBAAgB,KAAK,OAAO,cAAc;EAE5E,IAAI,KAAK,SAAS,eAChB,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IAAE,MAAM;IAAY,QAAQ,KAAK;IAAQ,QAAQ;GAAK;EACrE,EACF;EAGF,IAAI,KAAK,SAAS,eAChB,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAGpD,OAAO,CAAC;CACV;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,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,uBAAuB;GACzC,iBAAiB,IAAI,2BAA2B;EAClD,CAAC;CACH;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8CA,AAAQ,oBAAoB,SAG1B;EACA,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,uBAAuB;GACvB,GAAI,CAAC,KAAK,aAAa,aAAa,EAAE,YAAY;EACpD;CACF;CAEA,AAAQ,qBACN,WACA,UAGA;EACA,IAAI,CAAC,KAAK,aAAa,WACrB,OAAO,CAAC;EAGV,IAAI,CAAC,WAAW,QACd,OAAO,WAAW,EAAE,kBAAkB,OAAO,IAAI,CAAC;EAGpD,OAAO,EAAE,kBAAkB,wBAAwB,UAAU,MAAM,EAAE;CACvE;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,OAAO,cAAwC,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF"}
+{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../../packages/ai-openai/src/model.ts"],"sourcesContent":["import {\n  InvalidRequestError,\n  ProviderError,\n  resolveStreamChunks,\n  safeJsonParse,\n  selectProviderTools,\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 { mergeProviderBody } from \"./provider-options\";\nimport type { OpenAIModelConfig } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport {\n  buildPromptCacheParams,\n  buildUsage,\n  inferStructuredOutput,\n  mapFinishReason,\n  planStructuredOutput,\n  toOpenAIMessages,\n  toOpenAIReasoningEffort,\n  toOpenAITools,\n  wrapOpenAIError,\n} from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * OpenAI-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `openai` SDK. Agents,\n * workflows, and supervisors never talk to OpenAI directly — they hold a\n * `ModelContract`, and this class is what makes that contract concrete for\n * any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, OpenRouter, local\n * gateways that speak the Chat Completions protocol).\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client + frozen `ModelConfig` (name,\n *   temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` and\n *   `ToolContract[]` into OpenAI wire shapes on the way out, and\n *   translating OpenAI's response (content, finish reason, tool calls,\n *   usage) back into the neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as a\n * class (see §4.2 of code-style.md — \"long-lived state across calls\").\n *\n * @example\n * import OpenAI from \"openai\";\n * const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n * const model = new OpenAIModel(client, { name: \"gpt-4o\", temperature: 0.3 });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class OpenAIModel 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    // `api: \"responses\"` belongs to `OpenAIResponsesModel`, which only the\n    // `OpenAISDK.model()` factory constructs. This class speaks Chat\n    // Completions only; accepting the flag here would silently ignore it.\n    if (config.api === \"responses\") {\n      throw new InvalidRequestError(\n        'OpenAIModel speaks Chat Completions only and cannot serve api: \"responses\". Use OpenAISDK.model({ api: \"responses\" }) on the direct OpenAI provider.',\n        { context: { provider, model: config.name, api: config.api } },\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      // o-series + gpt-5 / gpt-6 models surface a reasoning channel and accept\n      // the `reasoning_effort` param. Explicit config wins over the\n      // name-prefix inference.\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      // Direct OpenAI requests support automatic cache hits and explicit\n      // content-part breakpoints. Wrapper endpoints have distinct provider\n      // labels and may reject these OpenAI-only request fields.\n      promptCaching: this.provider === \"openai\",\n      // PDF + audio INPUT are off by default — OpenAI accepts `file`\n      // (PDF) and `input_audio` parts only on specific models, so the\n      // flags are conservative/honest and opt-in via config rather than\n      // name-inferred. When set, the agent admits the attachments and\n      // `toOpenAIMessages` maps them to the real wire parts.\n      pdf: config.pdf ?? false,\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Chat\n   * Completions endpoint, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    // Per-call request/response logs are hot-path in production agents\n    // — keep them at `debug` so `info` stays reserved for lifecycle\n    // events (agent starting/completed, etc.). Operators who need to\n    // audit every LLM call can raise log-level at runtime.\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: OpenAI.Chat.Completions.ChatCompletion;\n\n    try {\n      response = await this.client.chat.completions.create(\n        mergeProviderBody<OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming>(\n          {\n            model: this.name,\n            messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n            ...this.buildSamplingParams(options),\n            ...buildPromptCacheParams(this.provider, this.config),\n            ...this.buildTools(options),\n            ...this.buildResponseFormat(options?.responseSchema),\n            ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n          },\n          options,\n        ),\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    const [choice] = response.choices;\n\n    if (choice === undefined) {\n      // A 200 carrying an empty `choices` array. Every line below reads off\n      // `choice`, so without this the caller got a bare `TypeError: Cannot\n      // read properties of undefined` — no provider, no model, no hint that\n      // the response itself was the problem rather than our handling of it.\n      //\n      // `ProviderError` because that is what this is: the provider answered\n      // successfully and returned nothing usable. Callers already branch on\n      // `instanceof ProviderError` for any provider-side failure.\n      throw new ProviderError(\"OpenAI returned a response with no choices.\", {\n        context: { provider: \"openai\", model: this.name, choices: response.choices.length },\n      });\n    }\n\n    const finishReason = mapFinishReason(choice.finish_reason);\n    const usage = this.extractUsage(response.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: choice.message.content ?? \"\",\n      finishReason,\n      usage,\n      toolCalls: this.extractToolCalls(choice.message.tool_calls),\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s —\n   * `delta` for text tokens, `tool-call` when the model requests a tool,\n   * and a terminal `done` carrying the final finish reason + usage totals.\n   * Callers consume it with `for await`.\n   *\n   * With `options.streamChunks` it also yields `tool-args` for each\n   * `tool_calls[].function.arguments` fragment and `reasoning` for a\n   * vendor's `reasoning_content` / `reasoning` delta field. Usage arrives\n   * only on the final chunk here, so no `usage` chunk is emitted.\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 chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<typeof this.client.chat.completions.create>>;\n\n    try {\n      stream = await this.client.chat.completions.create(\n        mergeProviderBody<OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming>(\n          {\n            model: this.name,\n            messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n            ...this.buildSamplingParams(options),\n            ...buildPromptCacheParams(this.provider, this.config),\n            ...this.buildTools(options),\n            stream: true,\n            stream_options: { include_usage: true },\n            ...this.buildResponseFormat(options?.responseSchema),\n            ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n          },\n          options,\n        ),\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    let rawFinishReason: string = \"stop\";\n    let usage: Usage = { input: 0, output: 0, total: 0 };\n    const want = resolveStreamChunks(options);\n    const toolCallAccum = new Map<\n      number,\n      { id: string; name: string; arguments: string; streamed: number }\n    >();\n\n    try {\n      for await (const chunk of stream as AsyncIterable<OpenAI.Chat.Completions.ChatCompletionChunk>) {\n        const delta = chunk.choices[0]?.delta;\n        const finish = chunk.choices[0]?.finish_reason;\n\n        if (want.reasoning && delta) {\n          const reasoning = readReasoningDelta(delta);\n\n          if (reasoning) {\n            yield { type: \"reasoning\", delta: reasoning };\n          }\n        }\n\n        if (delta?.content) {\n          yield { type: \"delta\", content: delta.content };\n        }\n\n        if (delta?.tool_calls) {\n          for (const toolCall of delta.tool_calls) {\n            const idx = toolCall.index ?? 0;\n            if (!toolCallAccum.has(idx)) {\n              toolCallAccum.set(idx, { id: \"\", name: \"\", arguments: \"\", streamed: 0 });\n            }\n            const acc = toolCallAccum.get(idx)!;\n            if (toolCall.id) acc.id = toolCall.id;\n            if (toolCall.function?.name) acc.name = toolCall.function.name;\n            if (toolCall.function?.arguments) acc.arguments += toolCall.function.arguments;\n\n            // The first fragment of a call carries its id and name; later\n            // ones only `index` + argument text. Anything that arrived\n            // before the name is flushed once the call is identified.\n            if (want.toolArgs && acc.name && acc.arguments.length > acc.streamed) {\n              yield {\n                type: \"tool-args\",\n                callId: acc.id,\n                name: acc.name,\n                delta: acc.arguments.slice(acc.streamed),\n              };\n              acc.streamed = acc.arguments.length;\n            }\n          }\n        }\n\n        if (finish) {\n          rawFinishReason = finish;\n        }\n\n        if (chunk.usage) {\n          usage = this.extractUsage(chunk.usage);\n        }\n      }\n\n      for (const acc of toolCallAccum.values()) {\n        // Skip accumulators that never received a function name — those\n        // are partial fragments the model started but never identified\n        // (e.g. arguments-only deltas with no originating `id`/`name`).\n        // Yielding them produces nameless tool-calls the agent runtime\n        // can't dispatch and would mis-attribute as a registered tool.\n        if (!acc.name) continue;\n\n        yield {\n          type: \"tool-call\",\n          id: acc.id,\n          name: acc.name,\n          input: safeJsonParse<Record<string, unknown>>(acc.arguments, {}),\n        };\n      }\n    } catch (thrown) {\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      throw wrapped;\n    }\n\n    const finishReason = mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Spread-friendly `tools` fragment: the function tools, then the\n   * provider-defined entries whose `provider` equals this model's\n   * provider label (`\"openai\"`, or the vendor label an OpenAI-compatible\n   * wrapper sets, e.g. `\"groq\"`), sent verbatim. Chat Completions only\n   * accepts `function` and `custom` tool types from OpenAI itself; its\n   * hosted tools (`web_search`, `file_search`, …) need `api: \"responses\"`\n   * and OpenAI rejects them here with a 400.\n   */\n  private buildTools(options: ModelCallOptions | undefined): {\n    tools?: OpenAI.Chat.Completions.ChatCompletionTool[];\n  } {\n    const { tools, definitions } = selectProviderTools(options, {\n      accepts: [this.provider],\n      module: LOG_MODULE,\n      model: this.name,\n      logger: this.logger,\n    });\n    const mapped = [\n      ...(toOpenAITools(tools) ?? []),\n      ...(definitions as unknown as OpenAI.Chat.Completions.ChatCompletionTool[]),\n    ];\n\n    return mapped.length > 0 ? { tools: mapped } : {};\n  }\n\n  /** Map the neutral cache hint only for the direct OpenAI provider. */\n  private promptCacheBreakpoints(options: ModelCallOptions | undefined): number {\n    return this.provider === \"openai\" ? (options?.cacheControl?.breakpoints ?? 0) : 0;\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter. The mode decision (override handling,\n   * strict-compatibility downgrade to `json_object`) lives in\n   * `planStructuredOutput`, shared with the Responses adapter.\n   *\n   * Returns an empty spread when no schema was supplied, so the caller\n   * can unconditionally `...buildResponseFormat(...)` into the request.\n   */\n  private buildResponseFormat(responseSchema: Record<string, unknown> | undefined): {\n    response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"response_format\"];\n  } {\n    const plan = planStructuredOutput(responseSchema, this.config.responseFormat);\n\n    if (plan.mode === \"json_schema\") {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: { name: \"response\", schema: plan.schema, strict: true },\n        },\n      };\n    }\n\n    if (plan.mode === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    return {};\n  }\n\n  /**\n   * Normalize OpenAI's `usage` block (which may be absent on some responses\n   * or partials) into the neutral `Usage` shape. Missing usage collapses to\n   * zeros rather than propagating `undefined`, so downstream aggregation\n   * math stays safe.\n   *\n   * `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the\n   * subset of the prompt served from OpenAI's automatic prompt cache);\n   * `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`\n   * (the hidden reasoning channel on o-series / gpt-5 models, already\n   * counted within `output`). Both are emitted only when the provider\n   * reports a positive value, so non-reasoning / uncached calls keep the\n   * lean `{ input, output, total }` shape. `cacheWriteTokens` mirrors\n   * `prompt_tokens_details.cache_write_tokens` (billed writes on the\n   * models that charge for them). `cost` is attached when priced.\n   */\n  private extractUsage(raw: OpenAI.Completions.CompletionUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    return buildUsage(\n      {\n        input: raw.prompt_tokens ?? 0,\n        output: raw.completion_tokens ?? 0,\n        total: raw.total_tokens ?? 0,\n        cachedTokens: raw.prompt_tokens_details?.cached_tokens,\n        cacheWriteTokens: raw.prompt_tokens_details?.cache_write_tokens,\n        reasoningTokens: raw.completion_tokens_details?.reasoning_tokens,\n      },\n      this.pricing,\n    );\n  }\n\n  /**\n   * Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's\n   * `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat\n   * Completions API exposes a discrete effort knob, not a token budget,\n   * so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no\n   * wire equivalent here and is silently ignored.\n   *\n   * The neutral `ReasoningEffort` (`\"low\" | \"medium\" | \"high\" | \"none\"`)\n   * is a subset of OpenAI's accepted values, so it forwards verbatim —\n   * `\"none\"` included. `\"none\"` is load-bearing: gpt-5 / o-series models\n   * **reject function tools** on Chat Completions while reasoning is\n   * active, and the endpoint accepts tools only when `reasoning_effort`\n   * is `\"none\"` (the alternative is the Responses API). This is why the\n   * param is EMITTED for `\"none\"` rather than omitted — omitting it\n   * leaves the model reasoning server-side by default, so tools would\n   * still be rejected.\n   *\n   * No-ops when the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`).\n   *\n   * When the caller gave no explicit `effort`: defaults to `\"none\"` IF\n   * `hasTools` is true, else omits the param (provider default — the\n   * pre-existing behavior for a tool-less call). The default exists\n   * because, on a reasoning-capable model, there is no working alternative\n   * to `\"none\"` when tools are attached — omitting `reasoning_effort`\n   * leaves reasoning on server-side, which Chat Completions rejects tools\n   * for (empty replies on some model generations, a hard 400 — \"Function\n   * tools with reasoning_effort are not supported ... in\n   * /v1/chat/completions\" — on newer ones). An explicit `effort` from the\n   * caller always wins over this default, in either direction.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  /**\n   * Output-length and sampling params, in current Chat Completions shapes only.\n   *\n   * - `max_completion_tokens`, never the legacy `max_tokens`: current\n   *   models (the gpt-5 family and later) answer `400 unsupported_parameter`\n   *   to `max_tokens`, and `max_completion_tokens` is accepted everywhere.\n   * - No `temperature` for a reasoning-capable model: those accept only the\n   *   default value and reject any other with a 400.\n   */\n  private buildSamplingParams(options: ModelCallOptions | undefined): {\n    max_completion_tokens?: number;\n    temperature?: number;\n  } {\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      max_completion_tokens: maxTokens,\n      ...(!this.capabilities.reasoning && { temperature }),\n    };\n  }\n\n  private buildReasoningParams(\n    reasoning: ModelCallOptions[\"reasoning\"],\n    hasTools: boolean,\n  ): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning) {\n      return {};\n    }\n\n    if (!reasoning?.effort) {\n      return hasTools ? { reasoning_effort: \"none\" } : {};\n    }\n\n    return { reasoning_effort: toOpenAIReasoningEffort(reasoning.effort) };\n  }\n\n  /**\n   * Reshape OpenAI's `tool_calls` array into the neutral\n   * `ModelToolCallRequest[]`. The raw `arguments` field is a JSON string\n   * per OpenAI's protocol — we parse it defensively via `safeJsonParse` so\n   * malformed or empty arguments yield an empty object instead of crashing\n   * the trip. Returns `undefined` when no tools were requested so callers\n   * can branch on presence.\n   */\n  private extractToolCalls(\n    rawToolCalls: OpenAI.Chat.Completions.ChatCompletionMessageToolCall[] | undefined,\n  ): ModelToolCallRequest[] | undefined {\n    if (!rawToolCalls || rawToolCalls.length === 0) {\n      return undefined;\n    }\n\n    return rawToolCalls.map((toolCall) => ({\n      id: toolCall.id,\n      name: (toolCall as any).function.name,\n      input: safeJsonParse<Record<string, unknown>>((toolCall as any).function.arguments, {}),\n    }));\n  }\n}\n\n/**\n * Reasoning text an OpenAI-compatible vendor puts on a Chat Completions\n * stream delta. OpenAI itself streams none on this API. DeepSeek documents\n * `delta.reasoning_content` (API reference, chat completion chunk); Groq\n * documents the parsed reasoning under a `reasoning` field. Neither is in\n * the OpenAI SDK's delta type, so the delta is read as a plain record.\n */\nfunction readReasoningDelta(delta: object): string | undefined {\n  const record = delta as Record<string, unknown>;\n\n  for (const key of [\"reasoning_content\", \"reasoning\"]) {\n    const value = record[key];\n\n    if (typeof value === \"string\" && value !== \"\") {\n      return value;\n    }\n  }\n\n  return undefined;\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAkCA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCnB,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAMhC,IAAI,OAAO,QAAQ,aACjB,MAAM,IAAI,oBACR,4JACA,EAAE,SAAS;GAAE;GAAU,OAAO,OAAO;GAAM,KAAK,OAAO;EAAI,EAAE,CAC/D;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;GAI1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GAInE,eAAe,KAAK,aAAa;GAMjC,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,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,KAAK,YAAY,OAC5C,kBACE;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,GAAG,KAAK,WAAW,OAAO;IAC1B,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,OACF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,CAAC,UAAU,SAAS;EAE1B,IAAI,WAAW,QASb,MAAM,IAAI,cAAc,+CAA+C,EACrE,SAAS;GAAE,UAAU;GAAU,OAAO,KAAK;GAAM,SAAS,SAAS,QAAQ;EAAO,EACpF,CAAC;EAGH,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,sCAAsC;GAC9E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,OAAO,QAAQ,WAAW;GACnC;GACA;GACA,WAAW,KAAK,iBAAiB,OAAO,QAAQ,UAAU;EAC5D;CACF;;;;;;;;;;;;CAaA,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,KAAK,YAAY,OAC1C,kBACE;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,GAAG,KAAK,WAAW,OAAO;IAC1B,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,OACF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,IAAI,kBAA0B;EAC9B,IAAI,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACnD,MAAM,OAAO,oBAAoB,OAAO;EACxC,MAAM,gCAAgB,IAAI,IAGxB;EAEF,IAAI;GACF,WAAW,MAAM,SAAS,QAAsE;IAC9F,MAAM,QAAQ,MAAM,QAAQ,EAAE,EAAE;IAChC,MAAM,SAAS,MAAM,QAAQ,EAAE,EAAE;IAEjC,IAAI,KAAK,aAAa,OAAO;KAC3B,MAAM,YAAY,mBAAmB,KAAK;KAE1C,IAAI,WACF,MAAM;MAAE,MAAM;MAAa,OAAO;KAAU;IAEhD;IAEA,IAAI,OAAO,SACT,MAAM;KAAE,MAAM;KAAS,SAAS,MAAM;IAAQ;IAGhD,IAAI,OAAO,YACT,KAAK,MAAM,YAAY,MAAM,YAAY;KACvC,MAAM,MAAM,SAAS,SAAS;KAC9B,IAAI,CAAC,cAAc,IAAI,GAAG,GACxB,cAAc,IAAI,KAAK;MAAE,IAAI;MAAI,MAAM;MAAI,WAAW;MAAI,UAAU;KAAE,CAAC;KAEzE,MAAM,MAAM,cAAc,IAAI,GAAG;KACjC,IAAI,SAAS,IAAI,IAAI,KAAK,SAAS;KACnC,IAAI,SAAS,UAAU,MAAM,IAAI,OAAO,SAAS,SAAS;KAC1D,IAAI,SAAS,UAAU,WAAW,IAAI,aAAa,SAAS,SAAS;KAKrE,IAAI,KAAK,YAAY,IAAI,QAAQ,IAAI,UAAU,SAAS,IAAI,UAAU;MACpE,MAAM;OACJ,MAAM;OACN,QAAQ,IAAI;OACZ,MAAM,IAAI;OACV,OAAO,IAAI,UAAU,MAAM,IAAI,QAAQ;MACzC;MACA,IAAI,WAAW,IAAI,UAAU;KAC/B;IACF;IAGF,IAAI,QACF,kBAAkB;IAGpB,IAAI,MAAM,OACR,QAAQ,KAAK,aAAa,MAAM,KAAK;GAEzC;GAEA,KAAK,MAAM,OAAO,cAAc,OAAO,GAAG;IAMxC,IAAI,CAAC,IAAI,MAAM;IAEf,MAAM;KACJ,MAAM;KACN,IAAI,IAAI;KACR,MAAM,IAAI;KACV,OAAO,cAAuC,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAM,YAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;CAWA,AAAQ,WAAW,SAEjB;EACA,MAAM,EAAE,OAAO,gBAAgB,oBAAoB,SAAS;GAC1D,SAAS,CAAC,KAAK,QAAQ;GACvB,QAAQ;GACR,OAAO,KAAK;GACZ,QAAQ,KAAK;EACf,CAAC;EACD,MAAM,SAAS,CACb,GAAI,cAAc,KAAK,KAAK,CAAC,GAC7B,GAAI,WACN;EAEA,OAAO,OAAO,SAAS,IAAI,EAAE,OAAO,OAAO,IAAI,CAAC;CAClD;;CAGA,AAAQ,uBAAuB,SAA+C;EAC5E,OAAO,KAAK,aAAa,WAAY,SAAS,cAAc,eAAe,IAAK;CAClF;;;;;;;;;;CAWA,AAAQ,oBAAoB,gBAE1B;EACA,MAAM,OAAO,qBAAqB,gBAAgB,KAAK,OAAO,cAAc;EAE5E,IAAI,KAAK,SAAS,eAChB,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IAAE,MAAM;IAAY,QAAQ,KAAK;IAAQ,QAAQ;GAAK;EACrE,EACF;EAGF,IAAI,KAAK,SAAS,eAChB,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAGpD,OAAO,CAAC;CACV;;;;;;;;;;;;;;;;;CAkBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,OAAO,WACL;GACE,OAAO,IAAI,iBAAiB;GAC5B,QAAQ,IAAI,qBAAqB;GACjC,OAAO,IAAI,gBAAgB;GAC3B,cAAc,IAAI,uBAAuB;GACzC,kBAAkB,IAAI,uBAAuB;GAC7C,iBAAiB,IAAI,2BAA2B;EAClD,GACA,KAAK,OACP;CACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8CA,AAAQ,oBAAoB,SAG1B;EACA,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,uBAAuB;GACvB,GAAI,CAAC,KAAK,aAAa,aAAa,EAAE,YAAY;EACpD;CACF;CAEA,AAAQ,qBACN,WACA,UAGA;EACA,IAAI,CAAC,KAAK,aAAa,WACrB,OAAO,CAAC;EAGV,IAAI,CAAC,WAAW,QACd,OAAO,WAAW,EAAE,kBAAkB,OAAO,IAAI,CAAC;EAGpD,OAAO,EAAE,kBAAkB,wBAAwB,UAAU,MAAM,EAAE;CACvE;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,OAAO,cAAwC,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF;;;;;;;;AASA,SAAS,mBAAmB,OAAmC;CAC7D,MAAM,SAAS;CAEf,KAAK,MAAM,OAAO,CAAC,qBAAqB,WAAW,GAAG;EACpD,MAAM,QAAQ,OAAO;EAErB,IAAI,OAAO,UAAU,YAAY,UAAU,IACzC,OAAO;CAEX;AAGF"}

package/esm/provider-options.d.mts

@@ -0,0 +1,8 @@
+//#region ../../packages/ai-openai/src/provider-options.d.ts
+/** Adapter-only fields merged after the neutral request parameters. */
+type OpenAICallOptions = {
+  body?: Record<string, unknown>;
+};
+//#endregion
+export { OpenAICallOptions };
+//# sourceMappingURL=provider-options.d.mts.map

package/esm/provider-options.d.mts.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-options.d.mts","names":[],"sources":["../../../../../../../packages/ai-openai/src/provider-options.ts"],"mappings":";;KAGY,iBAAA;EAAsB,IAAA,GAAO,MAAM;AAAA"}

package/esm/provider-options.mjs

@@ -0,0 +1,49 @@
+import { InvalidRequestError } from "@warlock.js/ai";
+
+//#region ../../packages/ai-openai/src/provider-options.ts
+const RESERVED_FIELDS = new Set([
+	"model",
+	"messages",
+	"input",
+	"instructions",
+	"tools",
+	"stream",
+	"stream_options",
+	"previous_response_id",
+	"conversation",
+	"store",
+	"include",
+	"functions",
+	"function_call",
+	"prompt",
+	"background"
+]);
+function mergeProviderBody(built, options) {
+	const body = (options?.providerOptions?.openai)?.body ?? {};
+	const reserved = Object.keys(body).filter((key) => RESERVED_FIELDS.has(key));
+	if (reserved.length > 0) throw new InvalidRequestError(`providerOptions.openai.body cannot set ${reserved.join(", ")}: use the call's messages, tools and model options.`, { context: { fields: reserved } });
+	const defaults = built;
+	const overrides = Object.fromEntries(Object.entries(body).map(([key, value]) => {
+		const previous = defaults[key];
+		if (!isPlainObject(previous) || !isPlainObject(value)) return [key, value];
+		if (typeof previous.type === "string" && typeof value.type === "string" && previous.type !== value.type) return [key, value];
+		return [key, {
+			...previous,
+			...value
+		}];
+	}));
+	return {
+		...built,
+		...overrides
+	};
+}
+/** Merge only plain objects; arrays, null and provider scalar values replace the field. */
+function isPlainObject(value) {
+	if (value === null || typeof value !== "object") return false;
+	const prototype = Object.getPrototypeOf(value);
+	return prototype === Object.prototype || prototype === null;
+}
+
+//#endregion
+export { mergeProviderBody };
+//# sourceMappingURL=provider-options.mjs.map

package/esm/provider-options.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-options.mjs","names":[],"sources":["../../../../../../../packages/ai-openai/src/provider-options.ts"],"sourcesContent":["import { InvalidRequestError, type ModelCallOptions } from \"@warlock.js/ai\";\n\n/** Adapter-only fields merged after the neutral request parameters. */\nexport type OpenAICallOptions = { body?: Record<string, unknown> };\n\nconst RESERVED_FIELDS = new Set([\n  \"model\",\n  \"messages\",\n  \"input\",\n  \"instructions\",\n  \"tools\",\n  \"stream\",\n  \"stream_options\",\n  \"previous_response_id\",\n  \"conversation\",\n  \"store\",\n  \"include\",\n  \"functions\",\n  \"function_call\",\n  \"prompt\",\n  \"background\",\n]);\n\nexport function mergeProviderBody<T extends object>(\n  built: T,\n  options: ModelCallOptions | undefined,\n): T {\n  const provider = options?.providerOptions?.openai as OpenAICallOptions | undefined;\n  const body = provider?.body ?? {};\n  const reserved = Object.keys(body).filter((key) => RESERVED_FIELDS.has(key));\n  if (reserved.length > 0) {\n    throw new InvalidRequestError(\n      `providerOptions.openai.body cannot set ${reserved.join(\", \")}: use the call's messages, tools and model options.`,\n      { context: { fields: reserved } },\n    );\n  }\n  const defaults = built as Record<string, unknown>;\n  const overrides = Object.fromEntries(\n    Object.entries(body).map(([key, value]) => {\n      const previous = defaults[key];\n      if (!isPlainObject(previous) || !isPlainObject(value)) return [key, value];\n      // A different discriminator selects a whole SDK variant. Keeping fields\n      // from the previous variant can make an otherwise valid override invalid.\n      if (\n        typeof previous.type === \"string\" &&\n        typeof value.type === \"string\" &&\n        previous.type !== value.type\n      )\n        return [key, value];\n      return [key, { ...previous, ...value }];\n    }),\n  );\n  return { ...built, ...overrides };\n}\n\n/** Merge only plain objects; arrays, null and provider scalar values replace the field. */\nfunction isPlainObject(value: unknown): value is Record<string, unknown> {\n  if (value === null || typeof value !== \"object\") return false;\n  const prototype = Object.getPrototypeOf(value);\n  return prototype === Object.prototype || prototype === null;\n}\n"],"mappings":";;;AAKA,MAAM,kBAAkB,IAAI,IAAI;CAC9B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,SAAgB,kBACd,OACA,SACG;CAEH,MAAM,QADW,SAAS,iBAAiB,OACtB,EAAE,QAAQ,CAAC;CAChC,MAAM,WAAW,OAAO,KAAK,IAAI,CAAC,CAAC,QAAQ,QAAQ,gBAAgB,IAAI,GAAG,CAAC;CAC3E,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,oBACR,0CAA0C,SAAS,KAAK,IAAI,EAAE,sDAC9D,EAAE,SAAS,EAAE,QAAQ,SAAS,EAAE,CAClC;CAEF,MAAM,WAAW;CACjB,MAAM,YAAY,OAAO,YACvB,OAAO,QAAQ,IAAI,CAAC,CAAC,KAAK,CAAC,KAAK,WAAW;EACzC,MAAM,WAAW,SAAS;EAC1B,IAAI,CAAC,cAAc,QAAQ,KAAK,CAAC,cAAc,KAAK,GAAG,OAAO,CAAC,KAAK,KAAK;EAGzE,IACE,OAAO,SAAS,SAAS,YACzB,OAAO,MAAM,SAAS,YACtB,SAAS,SAAS,MAAM,MAExB,OAAO,CAAC,KAAK,KAAK;EACpB,OAAO,CAAC,KAAK;GAAE,GAAG;GAAU,GAAG;EAAM,CAAC;CACxC,CAAC,CACH;CACA,OAAO;EAAE,GAAG;EAAO,GAAG;CAAU;AAClC;;AAGA,SAAS,cAAc,OAAkD;CACvE,IAAI,UAAU,QAAQ,OAAO,UAAU,UAAU,OAAO;CACxD,MAAM,YAAY,OAAO,eAAe,KAAK;CAC7C,OAAO,cAAc,OAAO,aAAa,cAAc;AACzD"}

package/esm/responses-model.mjs

@@ -4,18 +4,32 @@ import { mapResponsesFinishReason, toResponseFailure } from "./utils/map-respons
 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 { attachReasoningReplay, pickTurnReasoning } 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 { mergeProviderBody } from "./provider-options.mjs";
 import { inferReasoningCapability } from "./known-reasoning-models.mjs";
 import { inferVisionCapability } from "./known-vision-models.mjs";
-import { InvalidRequestError, ProviderError, safeJsonParse } from "@warlock.js/ai";
+import { InvalidRequestError, ProviderError, resolveStreamChunks, safeJsonParse, selectProviderTools } from "@warlock.js/ai";
 import { log } from "@warlock.js/logger";
 
-//#region ../ai-openai/src/responses-model.ts
+//#region ../../packages/ai-openai/src/responses-model.ts
 const LOG_MODULE = "ai.openai";
 /**
+* Responses output item types of built-in tools the CALLER must execute
+* (installed SDK `ResponseOutputItem` union). The adapter maps only
+* `function_call` to a neutral tool call, so these fail closed.
+*/
+const CLIENT_RUN_ITEM_TYPES = new Set([
+	"computer_call",
+	"custom_tool_call",
+	"local_shell_call",
+	"shell_call",
+	"apply_patch_call",
+	"mcp_approval_request"
+]);
+/**
 * OpenAI Responses API implementation of `ModelContract` (`api: "responses"`).
 *
 * **Role.** A sibling of `OpenAIModel` that speaks `POST /v1/responses`
@@ -42,7 +56,16 @@ const LOG_MODULE = "ai.openai";
 *
 * **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`.
+* `tool-call`s, then one `done`. With `streamChunks` it also yields
+* `reasoning` (reasoning summary / reasoning text deltas) and `tool-args`
+* (`response.function_call_arguments.delta`).
+*
+* **Provider-defined tools.** `ai.providerTool({ provider: "openai" })`
+* definitions are appended to `tools` verbatim. Hosted tools run inside the
+* response; their output items (`web_search_call`, …) are not replayed
+* (with `store: false` the next turn only needs the answer text), and a
+* reasoning item that led into one is not replayed either (see
+* `pickTurnReasoning`).
 */
 var OpenAIResponsesModel = class {
 	constructor(client, config, provider = "openai") {
@@ -105,6 +128,26 @@ var OpenAIResponsesModel = class {
 			...toolCalls ? { toolCalls } : {}
 		};
 	}
+	/** Count native Responses input tokens using the same converted input and tools. */
+	async countTokens(messages, options) {
+		const request = this.buildRequest(messages, options);
+		const body = {
+			model: request.model,
+			input: request.input,
+			instructions: request.instructions,
+			tools: request.tools,
+			tool_choice: request.tool_choice,
+			reasoning: request.reasoning,
+			text: request.text,
+			parallel_tool_calls: request.parallel_tool_calls,
+			truncation: request.truncation ?? void 0
+		};
+		try {
+			return (await this.client.responses.inputTokens.count(body, options?.signal ? { signal: options.signal } : void 0)).input_tokens;
+		} catch (thrown) {
+			throw this.logAndWrap(thrown);
+		}
+	}
 	/**
 	* Incremental streaming completion via `responses.create({ stream: true })`
 	* (the SDK's `Stream<ResponseStreamEvent>`, `responses.d.ts:55`). Yields the
@@ -139,7 +182,9 @@ var OpenAIResponsesModel = class {
 			throw this.logAndWrap(thrown);
 		}
 		const toolCalls = [];
-		const reasoningItems = [];
+		const outputItems = [];
+		const want = resolveStreamChunks(options);
+		const streamingCalls = /* @__PURE__ */ new Map();
 		let terminal;
 		try {
 			for await (const event of events) {
@@ -150,10 +195,36 @@ var OpenAIResponsesModel = class {
 					};
 					continue;
 				}
+				if (event.type === "response.reasoning_summary_text.delta" || event.type === "response.reasoning_text.delta") {
+					if (want.reasoning && event.delta) yield {
+						type: "reasoning",
+						delta: event.delta
+					};
+					continue;
+				}
+				if (event.type === "response.output_item.added") {
+					const { item } = event;
+					if (item.type === "function_call" && item.id) streamingCalls.set(item.id, {
+						callId: item.call_id,
+						name: item.name
+					});
+					continue;
+				}
+				if (event.type === "response.function_call_arguments.delta") {
+					const call = streamingCalls.get(event.item_id);
+					if (want.toolArgs && call && event.delta) yield {
+						type: "tool-args",
+						callId: call.callId,
+						name: call.name,
+						delta: 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));
+					this.assertServerRunItem(item);
+					outputItems.push(item);
+					if (item.type === "function_call" && item.status !== "incomplete") toolCalls.push(this.toToolCall(item));
 					continue;
 				}
 				if (event.type === "response.completed" || event.type === "response.incomplete") {
@@ -174,7 +245,7 @@ var OpenAIResponsesModel = class {
 		} catch (thrown) {
 			throw this.logAndWrap(thrown);
 		}
-		for (const call of attachReasoningReplay(toolCalls, this.name, reasoningItems)) yield {
+		for (const call of attachReasoningReplay(toolCalls, this.name, pickTurnReasoning(outputItems))) yield {
 			type: "tool-call",
 			id: call.id,
 			name: call.name,
@@ -212,11 +283,11 @@ var OpenAIResponsesModel = class {
 	*/
 	buildRequest(messages, options) {
 		const { instructions, input } = toResponsesInput(messages, this.name);
-		const tools = toResponsesTools(options?.tools);
+		const tools = this.buildTools(options);
 		const toolChoice = tools ? toResponsesToolChoice(options?.toolChoice) : void 0;
 		const maxTokens = options?.maxTokens ?? this.config.maxTokens;
 		const temperature = options?.temperature ?? this.config.temperature;
-		return {
+		return mergeProviderBody({
 			model: this.name,
 			...instructions !== void 0 ? { instructions } : {},
 			input,
@@ -229,7 +300,24 @@ var OpenAIResponsesModel = class {
 			...this.buildTextFormat(options?.responseSchema),
 			...this.buildReasoningParams(options?.reasoning),
 			...buildPromptCacheParams(this.provider, this.config)
-		};
+		}, options);
+	}
+	/**
+	* The request `tools`: function tools (flat, `strict: false`), then the
+	* `ai.providerTool({ provider: "openai" })` definitions verbatim — the
+	* Responses built-in tools (`web_search`, `file_search`,
+	* `code_interpreter`, `image_generation`, `mcp`, …). Other providers'
+	* definitions are skipped with a warning. `undefined` when there are none.
+	*/
+	buildTools(options) {
+		const { tools, definitions } = selectProviderTools(options, {
+			accepts: [this.provider],
+			module: LOG_MODULE,
+			model: this.name,
+			logger: this.logger
+		});
+		const mapped = [...toResponsesTools(tools) ?? [], ...definitions];
+		return mapped.length > 0 ? mapped : void 0;
 	}
 	/**
 	* Structured output on Responses lives in `text.format`, and `name` /
@@ -266,12 +354,8 @@ var OpenAIResponsesModel = class {
 	readOutput(output) {
 		let content = "";
 		const toolCalls = [];
-		const reasoningItems = [];
 		for (const item of output) {
-			if (item.type === "reasoning") {
-				reasoningItems.push(item);
-				continue;
-			}
+			this.assertServerRunItem(item);
 			if (item.type === "message") {
 				for (const part of item.content) content += part.type === "refusal" ? part.refusal : part.text;
 				continue;
@@ -280,9 +364,24 @@ var OpenAIResponsesModel = class {
 		}
 		return {
 			content,
-			toolCalls: toolCalls.length > 0 ? attachReasoningReplay(toolCalls, this.name, reasoningItems) : void 0
+			toolCalls: toolCalls.length > 0 ? attachReasoningReplay(toolCalls, this.name, pickTurnReasoning(output)) : void 0
 		};
 	}
+	/**
+	* Fail closed on a built-in tool call that YOUR code would have to run
+	* (`shell_call`, `apply_patch_call`, `computer_call`, …, from a
+	* provider-defined client tool). These items have their own output item
+	* types that this adapter does not map, so the turn would otherwise end
+	* silently with the call dropped. Hosted (server-run) tool items pass.
+	*/
+	assertServerRunItem(item) {
+		if (!CLIENT_RUN_ITEM_TYPES.has(item.type)) return;
+		throw new InvalidRequestError(`OpenAI returned a "${item.type}" item: that built-in tool is executed by the caller, which the Responses adapter does not support. Use OpenAI's server-run built-in tools (web_search, file_search, code_interpreter, image_generation, mcp without approval) with ai.providerTool(), or a function tool from ai.tool().`, { context: {
+			provider: this.provider,
+			model: this.name,
+			itemType: item.type
+		} });
+	}
 	/** One completed `function_call` item as a neutral tool call (`call_id` -> `id`). */
 	toToolCall(item) {
 		return {
@@ -311,7 +410,7 @@ var OpenAIResponsesModel = class {
 			cachedTokens: raw.input_tokens_details?.cached_tokens,
 			cacheWriteTokens: raw.input_tokens_details?.cache_write_tokens,
 			reasoningTokens: raw.output_tokens_details?.reasoning_tokens
-		});
+		}, this.pricing);
 	}
 	/** Wrap (idempotently), log and return a failure for the caller to throw. */
 	logAndWrap(thrown) {

package/esm/responses-model.mjs.map

@@ -1 +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"}
+{"version":3,"file":"responses-model.mjs","names":[],"sources":["../../../../../../../packages/ai-openai/src/responses-model.ts"],"sourcesContent":["import {\n  InvalidRequestError,\n  ProviderError,\n  resolveStreamChunks,\n  safeJsonParse,\n  selectProviderTools,\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 { mergeProviderBody } from \"./provider-options\";\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  pickTurnReasoning,\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 * Responses output item types of built-in tools the CALLER must execute\n * (installed SDK `ResponseOutputItem` union). The adapter maps only\n * `function_call` to a neutral tool call, so these fail closed.\n */\nconst CLIENT_RUN_ITEM_TYPES: ReadonlySet<string> = new Set([\n  \"computer_call\",\n  \"custom_tool_call\",\n  \"local_shell_call\",\n  \"shell_call\",\n  \"apply_patch_call\",\n  \"mcp_approval_request\",\n]);\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`. With `streamChunks` it also yields\n * `reasoning` (reasoning summary / reasoning text deltas) and `tool-args`\n * (`response.function_call_arguments.delta`).\n *\n * **Provider-defined tools.** `ai.providerTool({ provider: \"openai\" })`\n * definitions are appended to `tools` verbatim. Hosted tools run inside the\n * response; their output items (`web_search_call`, …) are not replayed\n * (with `store: false` the next turn only needs the answer text), and a\n * reasoning item that led into one is not replayed either (see\n * `pickTurnReasoning`).\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  /** Count native Responses input tokens using the same converted input and tools. */\n  public async countTokens(messages: Message[], options?: ModelCallOptions): Promise<number> {\n    const request = this.buildRequest(messages, options);\n    const body: OpenAI.Responses.InputTokenCountParams = {\n      model: request.model,\n      input: request.input,\n      instructions: request.instructions,\n      tools: request.tools,\n      tool_choice: request.tool_choice,\n      reasoning: request.reasoning,\n      text: request.text,\n      parallel_tool_calls: request.parallel_tool_calls,\n      truncation: request.truncation ?? undefined,\n    };\n    try {\n      const response = await this.client.responses.inputTokens.count(\n        body,\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n      return response.input_tokens;\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\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 outputItems: OpenAI.Responses.ResponseOutputItem[] = [];\n    const want = resolveStreamChunks(options);\n    // `function_call_arguments.delta` names its item by `item_id` only; the\n    // `call_id` and `name` come from that item's `output_item.added`.\n    const streamingCalls = new Map<string, { callId: string; name: string }>();\n    let terminal: OpenAI.Responses.Response | undefined;\n\n    try {\n      for await (const event of events) {\n        if (\n          event.type === \"response.output_text.delta\" ||\n          event.type === \"response.refusal.delta\"\n        ) {\n          if (event.delta) {\n            yield { type: \"delta\", content: event.delta };\n          }\n\n          continue;\n        }\n\n        if (\n          event.type === \"response.reasoning_summary_text.delta\" ||\n          event.type === \"response.reasoning_text.delta\"\n        ) {\n          if (want.reasoning && event.delta) {\n            yield { type: \"reasoning\", delta: event.delta };\n          }\n\n          continue;\n        }\n\n        if (event.type === \"response.output_item.added\") {\n          const { item } = event;\n\n          if (item.type === \"function_call\" && item.id) {\n            streamingCalls.set(item.id, { callId: item.call_id, name: item.name });\n          }\n\n          continue;\n        }\n\n        if (event.type === \"response.function_call_arguments.delta\") {\n          const call = streamingCalls.get(event.item_id);\n\n          if (want.toolArgs && call && event.delta) {\n            yield { type: \"tool-args\", callId: call.callId, name: call.name, delta: event.delta };\n          }\n\n          continue;\n        }\n\n        if (event.type === \"response.output_item.done\") {\n          const { item } = event;\n\n          this.assertServerRunItem(item);\n          outputItems.push(item);\n\n          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(\n      toolCalls,\n      this.name,\n      pickTurnReasoning(outputItems),\n    )) {\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 = this.buildTools(options);\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 mergeProviderBody<OpenAI.Responses.ResponseCreateParamsNonStreaming>(\n      {\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      options,\n    );\n  }\n\n  /**\n   * The request `tools`: function tools (flat, `strict: false`), then the\n   * `ai.providerTool({ provider: \"openai\" })` definitions verbatim — the\n   * Responses built-in tools (`web_search`, `file_search`,\n   * `code_interpreter`, `image_generation`, `mcp`, …). Other providers'\n   * definitions are skipped with a warning. `undefined` when there are none.\n   */\n  private buildTools(options: ModelCallOptions | undefined): OpenAI.Responses.Tool[] | undefined {\n    const { tools, definitions } = selectProviderTools(options, {\n      accepts: [this.provider],\n      module: LOG_MODULE,\n      model: this.name,\n      logger: this.logger,\n    });\n    const mapped: OpenAI.Responses.Tool[] = [\n      ...(toResponsesTools(tools) ?? []),\n      ...(definitions as unknown as OpenAI.Responses.Tool[]),\n    ];\n\n    return mapped.length > 0 ? mapped : undefined;\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: {\n          format: { type: \"json_schema\", name: \"response\", schema: plan.schema, strict: true },\n        },\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\n    for (const item of output) {\n      this.assertServerRunItem(item);\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, pickTurnReasoning(output))\n          : undefined,\n    };\n  }\n\n  /**\n   * Fail closed on a built-in tool call that YOUR code would have to run\n   * (`shell_call`, `apply_patch_call`, `computer_call`, …, from a\n   * provider-defined client tool). These items have their own output item\n   * types that this adapter does not map, so the turn would otherwise end\n   * silently with the call dropped. Hosted (server-run) tool items pass.\n   */\n  private assertServerRunItem(item: { type: string }): void {\n    if (!CLIENT_RUN_ITEM_TYPES.has(item.type)) {\n      return;\n    }\n\n    throw new InvalidRequestError(\n      `OpenAI returned a \"${item.type}\" item: that built-in tool is executed by the caller, ` +\n        \"which the Responses adapter does not support. Use OpenAI's server-run built-in tools \" +\n        \"(web_search, file_search, code_interpreter, image_generation, mcp without approval) \" +\n        \"with ai.providerTool(), or a function tool from ai.tool().\",\n      { context: { provider: this.provider, model: this.name, itemType: item.type } },\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      {\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      this.pricing,\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":";;;;;;;;;;;;;;;;;AAsCA,MAAM,aAAa;;;;;;AAOnB,MAAM,wBAA6C,IAAI,IAAI;CACzD;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwCD,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;;CAGA,MAAa,YAAY,UAAqB,SAA6C;EACzF,MAAM,UAAU,KAAK,aAAa,UAAU,OAAO;EACnD,MAAM,OAA+C;GACnD,OAAO,QAAQ;GACf,OAAO,QAAQ;GACf,cAAc,QAAQ;GACtB,OAAO,QAAQ;GACf,aAAa,QAAQ;GACrB,WAAW,QAAQ;GACnB,MAAM,QAAQ;GACd,qBAAqB,QAAQ;GAC7B,YAAY,QAAQ,cAAc;EACpC;EACA,IAAI;GAKF,QAAO,MAJgB,KAAK,OAAO,UAAU,YAAY,MACvD,MACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD,EACe,CAAC;EAClB,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;CACF;;;;;;;;;;;;;;;;;;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,cAAqD,CAAC;EAC5D,MAAM,OAAO,oBAAoB,OAAO;EAGxC,MAAM,iCAAiB,IAAI,IAA8C;EACzE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,SAAS,QAAQ;IAChC,IACE,MAAM,SAAS,gCACf,MAAM,SAAS,0BACf;KACA,IAAI,MAAM,OACR,MAAM;MAAE,MAAM;MAAS,SAAS,MAAM;KAAM;KAG9C;IACF;IAEA,IACE,MAAM,SAAS,2CACf,MAAM,SAAS,iCACf;KACA,IAAI,KAAK,aAAa,MAAM,OAC1B,MAAM;MAAE,MAAM;MAAa,OAAO,MAAM;KAAM;KAGhD;IACF;IAEA,IAAI,MAAM,SAAS,8BAA8B;KAC/C,MAAM,EAAE,SAAS;KAEjB,IAAI,KAAK,SAAS,mBAAmB,KAAK,IACxC,eAAe,IAAI,KAAK,IAAI;MAAE,QAAQ,KAAK;MAAS,MAAM,KAAK;KAAK,CAAC;KAGvE;IACF;IAEA,IAAI,MAAM,SAAS,0CAA0C;KAC3D,MAAM,OAAO,eAAe,IAAI,MAAM,OAAO;KAE7C,IAAI,KAAK,YAAY,QAAQ,MAAM,OACjC,MAAM;MAAE,MAAM;MAAa,QAAQ,KAAK;MAAQ,MAAM,KAAK;MAAM,OAAO,MAAM;KAAM;KAGtF;IACF;IAEA,IAAI,MAAM,SAAS,6BAA6B;KAC9C,MAAM,EAAE,SAAS;KAEjB,KAAK,oBAAoB,IAAI;KAC7B,YAAY,KAAK,IAAI;KAErB,IAAI,KAAK,SAAS,mBAAmB,KAAK,WAAW,cACnD,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,sBACjB,WACA,KAAK,MACL,kBAAkB,WAAW,CAC/B,GACE,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,KAAK,WAAW,OAAO;EACrC,MAAM,aAAa,QAAQ,sBAAsB,SAAS,UAAU,IAAI;EACxE,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO,kBACL;GACE,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,GACA,OACF;CACF;;;;;;;;CASA,AAAQ,WAAW,SAA4E;EAC7F,MAAM,EAAE,OAAO,gBAAgB,oBAAoB,SAAS;GAC1D,SAAS,CAAC,KAAK,QAAQ;GACvB,QAAQ;GACR,OAAO,KAAK;GACZ,QAAQ,KAAK;EACf,CAAC;EACD,MAAM,SAAkC,CACtC,GAAI,iBAAiB,KAAK,KAAK,CAAC,GAChC,GAAI,WACN;EAEA,OAAO,OAAO,SAAS,IAAI,SAAS;CACtC;;;;;;;CAQA,AAAQ,gBAAgB,gBAEtB;EACA,MAAM,OAAO,qBAAqB,gBAAgB,KAAK,OAAO,cAAc;EAE5E,IAAI,KAAK,SAAS,eAChB,OAAO,EACL,MAAM,EACJ,QAAQ;GAAE,MAAM;GAAe,MAAM;GAAY,QAAQ,KAAK;GAAQ,QAAQ;EAAK,EACrF,EACF;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;EAE3C,KAAK,MAAM,QAAQ,QAAQ;GACzB,KAAK,oBAAoB,IAAI;GAE7B,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,kBAAkB,MAAM,CAAC,IACrE;EACR;CACF;;;;;;;;CASA,AAAQ,oBAAoB,MAA8B;EACxD,IAAI,CAAC,sBAAsB,IAAI,KAAK,IAAI,GACtC;EAGF,MAAM,IAAI,oBACR,sBAAsB,KAAK,KAAK,4RAIhC,EAAE,SAAS;GAAE,UAAU,KAAK;GAAU,OAAO,KAAK;GAAM,UAAU,KAAK;EAAK,EAAE,CAChF;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,WACL;GACE,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,cAAc,IAAI,sBAAsB;GACxC,kBAAkB,IAAI,sBAAsB;GAC5C,iBAAiB,IAAI,uBAAuB;EAC9C,GACA,KAAK,OACP;CACF;;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

@@ -1,7 +1,7 @@
 import { OpenAIEmbedderConfig, OpenAIImageConfig, OpenAIModelConfig, OpenAISDKConfig, OpenAISpeechConfig, OpenAITranscriptionConfig } from "./config.type.mjs";
 import { EmbedderContract, ImageModelContract, ModelContract, SDKAdapterContract, SpeechModelContract, TranscriptionModelContract } from "@warlock.js/ai";
 
-//#region ../ai-openai/src/sdk.d.ts
+//#region ../../packages/ai-openai/src/sdk.d.ts
 /**
  * OpenAI-backed implementation of `SDKAdapterContract`.
  *
@@ -50,8 +50,9 @@ declare class OpenAISDK implements SDKAdapterContract {
    * coming from the same upstream.
    *
    * Pricing resolution: per-model `config.pricing` wins; otherwise the
-   * SDK-level registry entry keyed by `config.name`; otherwise
-   * `undefined` (no cost computed).
+   * SDK-level registry entry keyed by `config.name`; otherwise, when the
+   * provider label is `"openai"`, the built-in list price
+   * (`OPENAI_MODEL_PRICING`); otherwise `undefined` (no cost computed).
    *
    * `config.api` selects the wire API. Omitted or `"chat"` builds the Chat
    * Completions `OpenAIModel`. `"responses"` builds `OpenAIResponsesModel`

package/esm/sdk.d.mts.map

@@ -1 +1 @@
-{"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"}
+{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../../packages/ai-openai/src/sdk.ts"],"mappings":";;;;;;AA6DA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,kBAAA;EAAA,iBACf,MAAA;EAAA,iBACA,QAAA;EAAA,iBACA,OAAA;cAEE,MAAA,EAAQ,eAAA;EAyGpB;;;;;;;;;;;;;AAkCyE;;;;;;;EA3GzE,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,aAAA;;;;;;;;EAsC5B,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

@@ -1,3 +1,4 @@
+import { openAIPricingFor } from "./known-pricing.mjs";
 import { OpenAIEmbedder } from "./embedder.mjs";
 import { OpenAIImageModel } from "./image.mjs";
 import { OpenAIModel } from "./model.mjs";
@@ -7,7 +8,7 @@ import { OpenAITranscriptionModel } from "./transcription.mjs";
 import OpenAI from "openai";
 import { InvalidRequestError, approximateTokenCount } from "@warlock.js/ai";
 
-//#region ../ai-openai/src/sdk.ts
+//#region ../../packages/ai-openai/src/sdk.ts
 /**
 * OpenAI-backed implementation of `SDKAdapterContract`.
 *
@@ -58,8 +59,9 @@ var OpenAISDK = class {
 	* coming from the same upstream.
 	*
 	* Pricing resolution: per-model `config.pricing` wins; otherwise the
-	* SDK-level registry entry keyed by `config.name`; otherwise
-	* `undefined` (no cost computed).
+	* SDK-level registry entry keyed by `config.name`; otherwise, when the
+	* provider label is `"openai"`, the built-in list price
+	* (`OPENAI_MODEL_PRICING`); otherwise `undefined` (no cost computed).
 	*
 	* `config.api` selects the wire API. Omitted or `"chat"` builds the Chat
 	* Completions `OpenAIModel`. `"responses"` builds `OpenAIResponsesModel`
@@ -80,7 +82,7 @@ var OpenAISDK = class {
 			model: config.name,
 			api
 		} });
-		const resolvedPricing = config.pricing ?? this.pricing?.[config.name];
+		const resolvedPricing = config.pricing ?? this.pricing?.[config.name] ?? (this.provider === "openai" ? openAIPricingFor(config.name) : void 0);
 		const resolvedConfig = resolvedPricing === config.pricing ? config : {
 			...config,
 			pricing: resolvedPricing