@warlock.js/ai-openai 4.7.0 → 4.8.0

package/CHANGELOG.md

@@ -4,6 +4,12 @@ All notable changes to `@warlock.js/ai-openai` are documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). `@warlock.js/*` packages are released in lockstep — every package shares the same version number, so a version below may list only the changes that affected this package.
 
+## 4.8.0 - 2026-07-19
+
+### Added
+
+- **`reasoning: { effort: "none" }`** now emits `reasoning_effort: "none"` on the wire — unblocks function tools on gpt-5 / o-series reasoning models, which otherwise reject tools on Chat Completions while reasoning is active and return empty replies.
+
 ## 4.6.0
 
 ### Added

package/cjs/index.cjs

@@ -1016,12 +1016,20 @@ var OpenAIModel = class {
 	* so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no
 	* wire equivalent here and is silently ignored.
 	*
+	* The neutral `ReasoningEffort` (`"low" | "medium" | "high" | "none"`)
+	* is a subset of OpenAI's accepted values, so it forwards verbatim —
+	* `"none"` included. `"none"` is load-bearing: gpt-5 / o-series models
+	* **reject function tools** on Chat Completions while reasoning is
+	* active, and the endpoint accepts tools only when `reasoning_effort`
+	* is `"none"` (the alternative is the Responses API). This is why the
+	* param is EMITTED for `"none"` rather than omitted — omitting it
+	* leaves the model reasoning server-side by default, so tools would
+	* still be rejected.
+	*
 	* No-ops in two cases so the adapter never forwards an unsupported
 	* param: (1) the model is not reasoning-capable
-	* (`capabilities.reasoning` is false — e.g. `gpt-4o`), or (2) the caller
-	* supplied no `effort`. The neutral `ReasoningEffort`
-	* (`"low" | "medium" | "high"`) is a strict subset of OpenAI's accepted
-	* values, so it forwards verbatim.
+	* (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any
+	* `reasoning_effort`), or (2) the caller supplied no `effort`.
 	*
 	* Returns an empty spread when nothing applies, so the caller can
 	* unconditionally `...buildReasoningParams(...)` into the request.

package/cjs/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["InvalidRequestError","AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","ContentFilterError","InvalidRequestError","ProviderError","OpenAI","LOG_MODULE","log","LOG_MODULE","log","InvalidRequestError","ProviderError","LOG_MODULE","log","LOG_MODULE","log","InvalidRequestError","log","InvalidRequestError","OpenAI"],"sources":["../../../../../../@warlock.js/ai-openai/src/utils/map-finish-reason.ts","../../../../../../@warlock.js/ai-openai/src/utils/to-openai-messages.ts","../../../../../../@warlock.js/ai-openai/src/utils/to-openai-tools.ts","../../../../../../@warlock.js/ai-openai/src/utils/wrap-openai-error.ts","../../../../../../@warlock.js/ai-openai/src/embedder.ts","../../../../../../@warlock.js/ai-openai/src/known-image-models.ts","../../../../../../@warlock.js/ai-openai/src/image.ts","../../../../../../@warlock.js/ai-openai/src/known-reasoning-models.ts","../../../../../../@warlock.js/ai-openai/src/known-vision-models.ts","../../../../../../@warlock.js/ai-openai/src/model.ts","../../../../../../@warlock.js/ai-openai/src/speech.ts","../../../../../../@warlock.js/ai-openai/src/transcription.ts","../../../../../../@warlock.js/ai-openai/src/sdk.ts"],"sourcesContent":["import type { FinishReason } from \"@warlock.js/ai\";\n\nconst finishReasonMap: Record<string, FinishReason> = {\n  stop: \"stop\",\n  tool_calls: \"tool_calls\",\n  length: \"length\",\n};\n\n/**\n * Map the raw OpenAI `finish_reason` string to the normalized FinishReason union.\n * Unknown/unexpected values fall through to \"error\".\n *\n * @example\n * mapFinishReason(\"stop\");        // \"stop\"\n * mapFinishReason(\"tool_calls\");  // \"tool_calls\"\n * mapFinishReason(null);          // \"error\"\n */\nexport function mapFinishReason(raw: string | null | undefined): FinishReason {\n  return finishReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral Message[] to OpenAI's chat message shape.\n * Handles the `tool` role (requires `tool_call_id`) and assistant messages\n * that carry `toolCalls` from a prior model response.\n *\n * Multipart `content` (a `ContentPart[]`) is mapped into OpenAI's user-message\n * content-parts shape: text becomes `{ type: \"text\", text }`, images become\n * `{ type: \"image_url\", image_url: { url } }` — with base64 sources rendered\n * as `data:` URLs inline.\n *\n * @example\n * const openaiMessages = toOpenAIMessages([\n *   { role: \"user\", content: \"Hi\" },\n *   { role: \"tool\", toolCallId: \"call_1\", content: '{\"ok\":true}' },\n * ]);\n *\n * @example\n * toOpenAIMessages([\n *   { role: \"user\", content: [\n *     { type: \"text\", text: \"What is this?\" },\n *     { type: \"image\", source: { url: \"https://example.com/cat.jpg\" } },\n *   ]},\n * ]);\n */\nexport function toOpenAIMessages(\n  messages: Message[],\n): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {\n  return messages.map((m) => {\n    if (m.role === \"tool\") {\n      return {\n        role: \"tool\",\n        content: stringifyContent(m.content),\n        tool_call_id: m.toolCallId ?? \"\",\n      };\n    }\n    if (m.role === \"assistant\" && m.toolCalls && m.toolCalls.length > 0) {\n      return {\n        role: \"assistant\",\n        content: stringifyContent(m.content),\n        tool_calls: m.toolCalls.map((tc) => ({\n          id: tc.id,\n          type: \"function\" as const,\n          function: { name: tc.name, arguments: JSON.stringify(tc.input ?? {}) },\n        })),\n      };\n    }\n\n    if (m.role === \"user\" && Array.isArray(m.content)) {\n      return {\n        role: \"user\",\n        content: m.content.map(toOpenAIContentPart),\n      };\n    }\n\n    return { role: m.role, content: stringifyContent(m.content) } as\n      | OpenAI.Chat.Completions.ChatCompletionUserMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionSystemMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionAssistantMessageParam;\n  });\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role (system / assistant text / tool), collapse a `ContentPart[]` to\n * its concatenated text so OpenAI's wire format stays valid. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to an OpenAI chat content part — one\n * branch per modality, each to its real wire shape:\n *\n * - `text` → `{ type: \"text\" }`.\n * - `image` → `{ type: \"image_url\" }` (remote URL, or a `data:` URL for\n *   inlined base64 bytes).\n * - `pdf` → `{ type: \"file\", file: { file_data } }` (OpenAI document\n *   input; base64 only — there is no remote-URL file source).\n * - `audio` → `{ type: \"input_audio\", input_audio: { data, format } }`\n *   (base64 only; `wav` / `mp3` are the only formats OpenAI accepts).\n *\n * PDF and audio reach this point ONLY when the model declared the\n * matching capability (`openai.model({ name, pdf: true })` /\n * `{ audio: true }`) — the agent's modality gate throws upfront\n * otherwise, so capability and behavior stay in lockstep. A remote-URL\n * pdf/audio source raises a typed `InvalidRequestError` here rather\n * than a downstream provider fault.\n */\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (part.type === \"image\") {\n    const url =\n      \"url\" in part.source\n        ? part.source.url\n        : `data:${part.source.mediaType};base64,${part.source.base64}`;\n\n    return { type: \"image_url\", image_url: { url } };\n  }\n\n  if (part.type === \"pdf\") {\n    if (\"url\" in part.source) {\n      throw new InvalidRequestError(\n        \"OpenAI chat completions cannot fetch a remote-URL PDF; supply base64 document bytes instead.\",\n      );\n    }\n\n    return {\n      type: \"file\",\n      file: {\n        filename: \"document.pdf\",\n        file_data: `data:${part.source.mediaType};base64,${part.source.base64}`,\n      },\n    };\n  }\n\n  // Audio — the remaining `ContentPart` variant.\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"OpenAI chat completions cannot fetch remote-URL audio; supply base64 audio bytes instead.\",\n    );\n  }\n\n  return {\n    type: \"input_audio\",\n    input_audio: {\n      data: part.source.base64,\n      format: toOpenAIAudioFormat(part.source.mediaType),\n    },\n  };\n}\n\n/**\n * Narrow a neutral audio media type to the two formats OpenAI's\n * `input_audio` accepts (`wav` / `mp3`). An unsupported type raises a\n * typed `InvalidRequestError` up front rather than a provider 400.\n */\nfunction toOpenAIAudioFormat(mediaType: string): \"wav\" | \"mp3\" {\n  if (mediaType === \"audio/wav\" || mediaType === \"audio/x-wav\" || mediaType === \"audio/wave\") {\n    return \"wav\";\n  }\n\n  if (mediaType === \"audio/mp3\" || mediaType === \"audio/mpeg\" || mediaType === \"audio/mpga\") {\n    return \"mp3\";\n  }\n\n  throw new InvalidRequestError(\n    `OpenAI input_audio supports only \"wav\" and \"mp3\"; got \"${mediaType}\".`,\n  );\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral ToolConfig[] to OpenAI's tools array.\n * Uses the shared `extractJsonSchema` helper; falls back to an empty-object\n * schema when extraction fails so the tool still registers with the provider.\n *\n * @example\n * const tools = toOpenAITools([weatherTool, calculatorTool]);\n * await client.chat.completions.create({ model, messages, tools });\n */\nexport function toOpenAITools(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): OpenAI.Chat.Completions.ChatCompletionTool[] | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return tools.map((tool) => ({\n    type: \"function\",\n    function: {\n      name: tool.name,\n      description: tool.description,\n      parameters: toParameters(tool.input),\n    },\n  }));\n}\n\n/**\n * Resolve a tool's input schema to a JSON-Schema object. OpenAI's\n * function `parameters` expects an object root; anything else (or a\n * failed extraction) degrades to an empty-object schema so the tool\n * still registers and the model simply sees no parameters.\n */\nfunction toParameters(input: ToolConfig<unknown, unknown>[\"input\"]): Record<string, unknown> {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema;\n  }\n\n  return { type: \"object\", properties: {} };\n}\n","import {\n  AIError,\n  ContentFilterError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n  QuotaExceededError,\n} from \"@warlock.js/ai\";\nimport OpenAI from \"openai\";\n\n/**\n * Raw-error fields the wrapper reads off an OpenAI SDK error.\n *\n * `APIError` exposes `status`, `code`, `message`, `type`, `headers` —\n * we duck-type because wrapped retries, proxied errors, and custom\n * error subclasses sometimes lose the `instanceof` relationship.\n */\ntype OpenAIErrorShape = {\n  status?: number;\n  code?: string | null;\n  message?: string;\n  type?: string | null;\n  headers?: Record<string, string> | undefined;\n  name?: string;\n};\n\n/**\n * Wrap any thrown value caught inside the OpenAI adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Prefers `APIError.code` when present (stable\n * machine identifier across SDK versions), falls back to `status` when\n * `code` is missing (common with proxied deployments that strip the\n * field). Name-based detection (`APIConnectionTimeoutError`) catches\n * transport-layer errors that never produced an HTTP response.\n *\n * `AIError` instances are returned unchanged — callers can pass the\n * error through `try/catch/throw wrap(e)` pipelines without accidental\n * double-wrapping.\n *\n * @example\n * try {\n *   return await this.client.chat.completions.create(...);\n * } catch (thrown) {\n *   throw wrapOpenAIError(thrown);\n * }\n */\nexport function wrapOpenAIError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(thrown, shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n\n  if (isTimeout(thrown, shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 401 || shape.code === \"invalid_api_key\") {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.code === \"insufficient_quota\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 429 || shape.code === \"rate_limit_exceeded\") {\n    return new ProviderRateLimitError(message, {\n      cause: thrown,\n      context,\n      retryAfter: parseRetryAfter(shape.headers),\n    });\n  }\n\n  if (shape.code === \"context_length_exceeded\") {\n    return new ContextLengthExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.code === \"content_filter\") {\n    return new ContentFilterError(message, {\n      cause: thrown,\n      context,\n      reason: message,\n    });\n  }\n\n  if (typeof shape.status === \"number\" && shape.status >= 400 && shape.status < 500) {\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  return new ProviderError(message, { cause: thrown, context });\n}\n\n/**\n * Read the raw error shape without depending on `instanceof APIError`\n * — some consumers wrap the SDK, and proxies sometimes strip the\n * prototype chain. Duck-typing on the visible fields is resilient to\n * both.\n */\nfunction toShape(thrown: unknown): OpenAIErrorShape {\n  if (thrown instanceof OpenAI.APIError) {\n    return {\n      status: thrown.status,\n      code: thrown.code,\n      message: thrown.message,\n      type: thrown.type,\n      headers: thrown.headers as Record<string, string> | undefined,\n      name: thrown.name,\n    };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      code: typeof raw.code === \"string\" ? raw.code : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      type: typeof raw.type === \"string\" ? raw.type : undefined,\n      headers:\n        typeof raw.headers === \"object\" && raw.headers !== null\n          ? (raw.headers as Record<string, string>)\n          : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the thrown value represents a timeout. OpenAI's SDK\n * throws `APIConnectionTimeoutError` for transport-level timeouts, and\n * Node surfaces `ETIMEDOUT` / `ECONNABORTED` on the lower socket\n * layer. Either signal counts.\n */\nfunction isTimeout(thrown: unknown, shape: OpenAIErrorShape): boolean {\n  if (thrown instanceof OpenAI.APIConnectionTimeoutError) {\n    return true;\n  }\n\n  if (shape.name === \"APIConnectionTimeoutError\") {\n    return true;\n  }\n\n  if (shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\") {\n    return true;\n  }\n\n  return false;\n}\n\n/**\n * Attach the raw diagnostic fields to `error.context` so consumers\n * have everything the provider surfaced without each subclass having\n * to redeclare them. Never includes `cause` — that lives on\n * `error.cause`.\n */\nfunction buildContext(\n  thrown: unknown,\n  shape: OpenAIErrorShape,\n): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.code) {\n    context.code = shape.code;\n  }\n\n  if (shape.type) {\n    context.type = shape.type;\n  }\n\n  const requestId = readRequestId(thrown);\n\n  if (requestId) {\n    context.requestId = requestId;\n  }\n\n  return context;\n}\n\n/**\n * OpenAI puts the request id on `APIError.request_id`. Extract\n * defensively — both camel and snake keys exist across SDK versions.\n */\nfunction readRequestId(thrown: unknown): string | undefined {\n  if (typeof thrown !== \"object\" || thrown === null) {\n    return undefined;\n  }\n\n  const raw = thrown as Record<string, unknown>;\n\n  if (typeof raw.request_id === \"string\") {\n    return raw.request_id;\n  }\n\n  if (typeof raw.requestId === \"string\") {\n    return raw.requestId;\n  }\n\n  return undefined;\n}\n\n/**\n * Parse the `Retry-After` response header (seconds per HTTP spec)\n * into milliseconds so consumers can feed it straight to `setTimeout`.\n * Returns `undefined` when missing or unparseable.\n */\nfunction parseRetryAfter(headers: Record<string, string> | undefined): number | undefined {\n  if (!headers) {\n    return undefined;\n  }\n\n  const raw = headers[\"retry-after\"] ?? headers[\"Retry-After\"];\n\n  if (!raw) {\n    return undefined;\n  }\n\n  const seconds = Number(raw);\n\n  if (!Number.isFinite(seconds) || seconds < 0) {\n    return undefined;\n  }\n\n  return Math.round(seconds * 1000);\n}\n","import {\n  type EmbeddingBatchResult,\n  type EmbeddingResult,\n  type EmbeddingUsage,\n  type EmbedderContract,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIEmbedderConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\ntype EmbeddingsResponse = Awaited<\n  ReturnType<OpenAI[\"embeddings\"][\"create\"]>\n>;\n\n/**\n * OpenAI-backed implementation of `EmbedderContract`.\n *\n * **Role.** Converts text (or a batch of texts) into floating-point\n * vectors via OpenAI's Embeddings API. Standalone primitive — no\n * relationship to chat completions, tools, or the agent loop.\n *\n * **Dimensions.** When no `dimensions` override is supplied in config,\n * `this.dimensions` starts at `0` and is populated from the first\n * response's vector length, then cached for all subsequent calls —\n * even if a later response were to return a different length, the\n * first value wins so batches stay dimensionally consistent. Passing\n * `dimensions` in config both forwards the truncation hint to the API\n * (for models like `text-embedding-3-*`) and sets the initial value.\n *\n * **Error handling.** Raw OpenAI SDK errors are wrapped into the\n * typed `@warlock.js/ai` `AIError` hierarchy via `wrapOpenAIError` —\n * callers catch `AIError` subclasses (`ProviderRateLimitError`,\n * `ProviderAuthError`, etc.) instead of OpenAI's own classes.\n *\n * @example\n * const embedder = new OpenAIEmbedder(client, { name: \"text-embedding-3-small\" });\n * const { vector, dimensions, usage } = await embedder.embed(\"Hello world\");\n * const { vectors } = await embedder.embedMany([\"doc 1\", \"doc 2\"]);\n */\nexport class OpenAIEmbedder implements EmbedderContract {\n  public readonly name: string;\n  public readonly provider = \"openai\";\n  public dimensions: number;\n\n  private readonly client: OpenAI;\n\n  /**\n   * User-specified truncation hint, or `undefined` if omitted.\n   * Forwarded to the API on every call so OpenAI can truncate the\n   * embedding server-side for models that support it.\n   */\n  private readonly configuredDimensions: number | undefined;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIEmbedderConfig) {\n    this.client = client;\n    this.name = config.name;\n    this.configuredDimensions = config.dimensions;\n    this.dimensions = config.dimensions ?? 0;\n  }\n\n  public async embed(input: string): Promise<EmbeddingResult> {\n    const { response, usage } = await this.request(input);\n\n    return {\n      vector: response.data[0].embedding,\n      dimensions: this.dimensions,\n      usage,\n    };\n  }\n\n  public async embedMany(inputs: string[]): Promise<EmbeddingBatchResult> {\n    const { response, usage } = await this.request(inputs);\n\n    return {\n      vectors: response.data.map((d) => d.embedding),\n      dimensions: this.dimensions,\n      usage,\n    };\n  }\n\n  /**\n   * Shared transport for both `embed()` and `embedMany()` — issues the\n   * `embeddings.create` call, wraps provider errors, caches dimensions\n   * on the first successful response, and returns the raw response\n   * plus a camelCase usage object for the caller to shape.\n   */\n  private async request(input: string | string[]): Promise<{\n    response: EmbeddingsResponse;\n    usage: EmbeddingUsage;\n  }> {\n    this.logger.debug(LOG_MODULE, \"embedder.request\", \"embeddings.create\", {\n      model: this.name,\n      batch: Array.isArray(input),\n      count: Array.isArray(input) ? input.length : 1,\n    });\n\n    let response: EmbeddingsResponse;\n\n    try {\n      response = await this.client.embeddings.create({\n        model: this.name,\n        input,\n        ...(this.configuredDimensions !== undefined\n          ? { dimensions: this.configuredDimensions }\n          : {}),\n      });\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"embedder.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    this.logger.debug(LOG_MODULE, \"embedder.response\", \"embeddings.create returned\", {\n      dimensions: response.data[0]?.embedding.length,\n      usage: {\n        promptTokens: response.usage.prompt_tokens,\n        totalTokens: response.usage.total_tokens,\n      },\n    });\n\n    // Cache dimensions on the first response. Once set, stays set —\n    // we trust the first call to define the shape for this embedder.\n    if (this.dimensions === 0) {\n      this.dimensions = response.data[0].embedding.length;\n    }\n\n    const usage: EmbeddingUsage = {\n      promptTokens: response.usage.prompt_tokens,\n      totalTokens: response.usage.total_tokens,\n    };\n\n    return { response, usage };\n  }\n}\n","/**\n * Model-id prefixes OpenAI exposes through the **Images** API\n * (`client.images.generate`). The two live families:\n *\n * - `gpt-image-*` — token-metered, always returns base64 bytes (no\n *   `response_format` knob), supports `output_format` + `background`.\n * - `dall-e-*` — per-image-metered, returns a URL or base64 via\n *   `response_format`.\n *\n * Used by {@link isOpenAIImageModel} for the construction-time guard so\n * `openai.image({ name: \"gpt-4o\" })` fails fast with a curated error\n * instead of a downstream 400 — mirroring the embedder/vision guards.\n */\nexport const OPENAI_IMAGE_MODEL_PREFIXES = [\"gpt-image\", \"dall-e\"] as const;\n\n/**\n * True when `name` is a recognized OpenAI image-generation model. A\n * prefix match (not an exact list) so dated snapshots\n * (`gpt-image-1-mini`, `dall-e-3`) are covered without a maintenance\n * burden every time OpenAI ships a point release.\n *\n * @example\n * isOpenAIImageModel(\"gpt-image-1\"); // true\n * isOpenAIImageModel(\"dall-e-3\");    // true\n * isOpenAIImageModel(\"gpt-4o\");      // false\n */\nexport function isOpenAIImageModel(name: string): boolean {\n  return OPENAI_IMAGE_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n","import {\n  InvalidRequestError,\n  ProviderError,\n  type GeneratedImage,\n  type ImageGenerationOptions,\n  type ImageGenerationResponse,\n  type ImageModelContract,\n  type ImageModelPricing,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIImageConfig } from \"./config.type\";\nimport { isOpenAIImageModel } from \"./known-image-models\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Map a neutral output container to its IANA media type. */\nfunction mediaTypeFor(format: string | undefined): string {\n  switch (format) {\n    case \"jpeg\":\n    case \"jpg\":\n      return \"image/jpeg\";\n    case \"webp\":\n      return \"image/webp\";\n    default:\n      return \"image/png\";\n  }\n}\n\n/**\n * OpenAI-backed implementation of `ImageModelContract`.\n *\n * **Role.** Bridges the vendor-neutral `ai.image()` verb to OpenAI's\n * **Images** API for the two image families OpenAI ships: the\n * token-metered `gpt-image-*` models (always return base64 bytes) and\n * the per-image-metered `dall-e-*` models (URL or base64). Like\n * `OpenAIEmbedder`, it's a standalone primitive — no relationship to\n * chat completions, tools, or the agent loop.\n *\n * **Capability guard.** The constructor rejects a non-image model id\n * up front (`gpt-4o` → typed `InvalidRequestError`) so the mistake\n * surfaces at wiring time, not as a downstream provider 400 — the\n * \"fail fast at construction\" rule shared with the embedder/vision\n * guards.\n *\n * **Error handling.** Raw OpenAI SDK errors are wrapped into the typed\n * `@warlock.js/ai` `AIError` hierarchy via `wrapOpenAIError`, so a\n * caller catches `ProviderRateLimitError` / `ContentFilterError` /\n * `ProviderAuthError` rather than OpenAI's own classes. `ai.image()`\n * turns those throws into `result.error`.\n *\n * @example\n * const model = new OpenAIImageModel(client, { name: \"gpt-image-1\" }, \"openai\");\n * const { images, usage } = await model.generate(\"a teal ceramic mug, studio light\");\n */\nexport class OpenAIImageModel implements ImageModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: ImageModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIImageConfig, provider: string = \"openai\") {\n    if (!isOpenAIImageModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI image-generation model. ` +\n          \"Use a `gpt-image-*` or `dall-e-*` model with openai.image({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n  }\n\n  public async generate(\n    prompt: string,\n    options?: ImageGenerationOptions,\n  ): Promise<ImageGenerationResponse> {\n    const isGptImage = this.name.startsWith(\"gpt-image\");\n    // gpt-image always returns base64 bytes (no `response_format` knob);\n    // DALL·E defaults to self-contained base64 here (URLs expire in ~60\n    // min), but the caller can ask for a URL via `options.responseFormat`.\n    const responseFormat =\n      (options?.responseFormat as \"url\" | \"b64_json\" | undefined) ??\n      (isGptImage ? undefined : \"b64_json\");\n\n    const body: OpenAI.Images.ImageGenerateParamsNonStreaming = {\n      model: this.name,\n      prompt,\n    };\n\n    if (options?.count !== undefined) body.n = options.count;\n    if (options?.size !== undefined) body.size = options.size;\n    if (options?.quality !== undefined) {\n      body.quality = options.quality as OpenAI.Images.ImageGenerateParamsBase[\"quality\"];\n    }\n    if (!isGptImage && responseFormat) body.response_format = responseFormat;\n    if (isGptImage && options?.format !== undefined) {\n      body.output_format = options.format as OpenAI.Images.ImageGenerateParamsBase[\"output_format\"];\n    }\n    if (options?.background !== undefined) {\n      body.background = options.background as OpenAI.Images.ImageGenerateParamsBase[\"background\"];\n    }\n\n    this.logger.debug(LOG_MODULE, \"image.request\", \"images.generate\", {\n      model: this.name,\n      count: options?.count ?? 1,\n    });\n\n    let response: OpenAI.Images.ImagesResponse;\n\n    try {\n      response = await this.client.images.generate(\n        body,\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"image.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const images = (response.data ?? []).map((image) =>\n      this.toGeneratedImage(image, options?.format),\n    );\n\n    const usage = response.usage\n      ? {\n          input: response.usage.input_tokens,\n          output: response.usage.output_tokens,\n          total: response.usage.total_tokens,\n        }\n      : { input: 0, output: 0, total: 0 };\n\n    this.logger.debug(LOG_MODULE, \"image.response\", \"images.generate succeeded\", {\n      images: images.length,\n      usage,\n    });\n\n    return { images, usage };\n  }\n\n  /**\n   * Normalize one OpenAI `Image` into the neutral discriminated shape.\n   * Base64 wins when present (gpt-image, and DALL·E in b64 mode);\n   * otherwise a hosted URL. A response carrying neither is a provider\n   * contract violation — surface it as a typed `ProviderError` rather\n   * than emitting a half-formed part.\n   */\n  private toGeneratedImage(image: OpenAI.Images.Image, format: string | undefined): GeneratedImage {\n    if (image.b64_json) {\n      return {\n        type: \"base64\",\n        base64: image.b64_json,\n        mediaType: mediaTypeFor(format),\n        ...(image.revised_prompt ? { revisedPrompt: image.revised_prompt } : {}),\n      };\n    }\n\n    if (image.url) {\n      return {\n        type: \"url\",\n        url: image.url,\n        ...(image.revised_prompt ? { revisedPrompt: image.revised_prompt } : {}),\n      };\n    }\n\n    throw new ProviderError(\"OpenAI image response contained neither base64 bytes nor a URL.\");\n  }\n}\n","/**\n * Model-name prefixes for OpenAI families that expose internal\n * reasoning / thinking tokens and accept the `reasoning_effort`\n * request parameter on the Chat Completions API.\n *\n * Matched as a prefix so dated variants (`o3-2025-04-16`) and\n * `-mini` / `-pro` suffixes (`o4-mini`, `gpt-5-pro`) are covered\n * without listing every release tag explicitly.\n *\n * Maintenance: append a new prefix when OpenAI ships a reasoning\n * model family that doesn't already match. Devs can always override\n * per-model via `openai.model({ name, reasoning: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst REASONING_CAPABLE_PREFIXES = [\"o1\", \"o3\", \"o4\", \"gpt-5\"];\n\n/**\n * Infer whether a given OpenAI model name is a reasoning model (o-series\n * and the gpt-5 family) based on the known-prefix list. Unknown models\n * default to `false` so the adapter never forwards an unsupported\n * `reasoning_effort` param to a non-reasoning model (which would 400).\n *\n * @example\n * inferReasoningCapability(\"o3-mini\");          // → true\n * inferReasoningCapability(\"o4-mini\");          // → true\n * inferReasoningCapability(\"gpt-5-pro\");        // → true\n * inferReasoningCapability(\"gpt-4o\");           // → false\n * inferReasoningCapability(\"custom-llm\");       // → false\n */\nexport function inferReasoningCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return REASONING_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","/**\n * Model-name prefixes for OpenAI families that support vision input\n * (image attachments) on the Chat Completions API.\n *\n * Matched as a prefix so dated variants (`gpt-4o-2024-08-06`) and\n * `-mini` / `-preview` suffixes (`gpt-4o-mini`, `gpt-4-turbo-preview`)\n * are covered without listing every release tag explicitly.\n *\n * Maintenance: append a new prefix when OpenAI ships a vision-capable\n * model family that doesn't already match. Devs can always override\n * per-model via `openai.model({ name, vision: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst VISION_CAPABLE_PREFIXES = [\n  \"gpt-4o\",\n  \"gpt-4-turbo\",\n  \"gpt-4.1\",\n  \"o1\",\n  \"o3\",\n  \"chatgpt-4o\",\n];\n\n/**\n * Infer whether a given OpenAI model name supports vision based on the\n * known-prefix list. Unknown models default to `false` so that passing\n * an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque OpenAI 400.\n *\n * @example\n * inferVisionCapability(\"gpt-4o-mini\");          // → true\n * inferVisionCapability(\"gpt-4o-2024-08-06\");    // → true\n * inferVisionCapability(\"gpt-3.5-turbo\");        // → false\n * inferVisionCapability(\"custom-llm\");           // → false\n */\nexport function inferVisionCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return VISION_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","import {\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, OpenAIResponseFormat } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toOpenAIMessages, toOpenAITools, wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * Map an explicit `responseFormat` override to the default\n * `structuredOutput` capability. Loose wire modes (`\"json_object\"`,\n * `\"text\"`) don't enforce shape, so the agent needs to see the soft\n * schema hint in the system prompt — that only happens when the\n * capability is `false`. Default (no override) stays `true` to\n * preserve the prior assumption that OpenAI models support strict\n * structured output.\n */\nfunction inferStructuredOutput(responseFormat: OpenAIResponseFormat | undefined): boolean {\n  if (responseFormat === \"json_object\" || responseFormat === \"text\") {\n    return false;\n  }\n\n  return true;\n}\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    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 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      // OpenAI prompt caching is automatic on the Chat Completions API\n      // (no caller-supplied breakpoints — the platform caches long\n      // prompt prefixes server-side and reports the hit count via\n      // `prompt_tokens_details.cached_tokens`). We therefore advertise\n      // the read-side accounting capability as always available while\n      // treating `ModelCallOptions.cacheControl` write breakpoints as a\n      // no-op (see `buildReasoningParams` siblings — there is no cache\n      // param to emit).\n      promptCaching: true,\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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[0];\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter.\n   *\n   * When `config.responseFormat` is set, it wins: `\"text\"` emits no\n   * `response_format` at all, `\"json_object\"` always picks the loose\n   * mode, and `\"json_schema\"` picks strict mode (with the same\n   * `isStrictCompatible` safety check — a malformed schema still\n   * degrades to `json_object` rather than 400). The override exists\n   * because some targets (older OpenAI models, OpenRouter routes,\n   * Ollama OpenAI-compat) reject strict `json_schema` outright.\n   *\n   * When the override is omitted, uses strict `json_schema` mode\n   * (token-level enforcement) only when the schema is a proper\n   * root-object JSON Schema (`{ type: \"object\", properties: ... }`).\n   * For anything else — malformed extractor output, non-object\n   * schemas, or future shapes we haven't tested — falls back to loose\n   * `json_object` mode, which guarantees *some* valid JSON without\n   * enforcing shape. The agent's soft instruction already embeds the\n   * schema text in the system prompt when the model declares no\n   * native structured-output capability, so shape validation still\n   * runs client-side via the Standard Schema `validate()` call.\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    if (!responseSchema) {\n      return {};\n    }\n\n    const override = this.config.responseFormat;\n\n    if (override === \"text\") {\n      return {};\n    }\n\n    if (override === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    // Either auto-select (no override) or explicit `\"json_schema\"`.\n    // The strict-compat check still applies in the explicit case —\n    // a malformed / non-object schema would 400 before sampling, so\n    // we degrade to `json_object` rather than crash.\n    if (this.isStrictCompatible(responseSchema)) {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: {\n            name: \"response\",\n            schema: responseSchema,\n            strict: true,\n          },\n        },\n      };\n    }\n\n    return { response_format: { type: \"json_object\" } };\n  }\n\n  /**\n   * OpenAI strict `json_schema` mode requires the root to be a JSON\n   * Schema object type (`{ type: \"object\", properties: ... }`). Anything\n   * else (top-level arrays, primitives, unknown shapes) is rejected with\n   * a 400 before a token is sampled. We check structurally here so the\n   * first call doesn't crash on a malformed extraction — loose\n   * `json_object` mode is a safe degradation.\n   */\n  private isStrictCompatible(schema: Record<string, unknown>): boolean {\n    return (\n      schema.type === \"object\" &&\n      typeof schema.properties === \"object\" &&\n      schema.properties !== null &&\n      this.isStrictSafeNode(schema)\n    );\n  }\n\n  /**\n   * Recursively check the one strict-mode rule schemas most often trip on:\n   * every object must list ALL of its `properties` in `required` (OpenAI\n   * strict has no notion of optional — optional fields must be expressed\n   * as nullable, e.g. `type: [\"string\", \"null\"]`, and still appear in\n   * `required`). A schema that violates this anywhere in the tree is NOT\n   * sent in strict `json_schema` mode — it degrades to loose\n   * `json_object` so a hand-built or optional-bearing schema can't 400\n   * the call (\"'required' ... must include every key in properties\").\n   * Client-side `validate()` still enforces the full shape.\n   */\n  private isStrictSafeNode(node: unknown): boolean {\n    if (!node || typeof node !== \"object\") {\n      return true;\n    }\n\n    const record = node as Record<string, unknown>;\n\n    if (record.type === \"object\" && record.properties && typeof record.properties === \"object\") {\n      const properties = record.properties as Record<string, unknown>;\n      const keys = Object.keys(properties);\n      const required = Array.isArray(record.required) ? (record.required as unknown[]) : [];\n\n      if (keys.some((key) => !required.includes(key))) {\n        return false;\n      }\n\n      for (const key of keys) {\n        if (!this.isStrictSafeNode(properties[key])) {\n          return false;\n        }\n      }\n    }\n\n    if (record.items !== undefined && !this.isStrictSafeNode(record.items)) {\n      return false;\n    }\n\n    for (const branch of [\"anyOf\", \"allOf\", \"oneOf\"] as const) {\n      const value = record[branch];\n      if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) {\n        return false;\n      }\n    }\n\n    return true;\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    const cachedTokens = raw.prompt_tokens_details?.cached_tokens;\n    const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;\n\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n      ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\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   * No-ops in two cases so the adapter never forwards an unsupported\n   * param: (1) the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`), or (2) the caller\n   * supplied no `effort`. The neutral `ReasoningEffort`\n   * (`\"low\" | \"medium\" | \"high\"`) is a strict subset of OpenAI's accepted\n   * values, so it forwards verbatim.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  private buildReasoningParams(reasoning: ModelCallOptions[\"reasoning\"]): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning || !reasoning?.effort) {\n      return {};\n    }\n\n    return { reasoning_effort: 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","import {\n  InvalidRequestError,\n  type SpeechGenerationResponse,\n  type SpeechModelContract,\n  type SpeechModelPricing,\n  type SpeechOptions,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAISpeechConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Model-id prefixes OpenAI exposes through the **Speech** (TTS) API. */\nconst SPEECH_MODEL_PREFIXES = [\"tts-1\", \"gpt-4o-mini-tts\", \"gpt-audio\"] as const;\n\n/** True when `name` is a recognized OpenAI text-to-speech model. */\nexport function isOpenAISpeechModel(name: string): boolean {\n  return SPEECH_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n\n/** Map a neutral output container hint to its IANA audio media type. */\nfunction audioMediaType(format: string | undefined): string {\n  switch (format) {\n    case \"opus\":\n      return \"audio/opus\";\n    case \"aac\":\n      return \"audio/aac\";\n    case \"flac\":\n      return \"audio/flac\";\n    case \"wav\":\n      return \"audio/wav\";\n    case \"pcm\":\n      return \"audio/pcm\";\n    default:\n      return \"audio/mpeg\";\n  }\n}\n\n/**\n * OpenAI-backed implementation of `SpeechModelContract` (text-to-speech)\n * via `audio.speech.create`. Standalone primitive — no relation to chat\n * completions or the agent loop. Consumed by the `ai.speech()` verb.\n *\n * **Capability guard.** The constructor rejects a non-TTS model id up\n * front (`tts-1` / `gpt-4o-mini-tts` only) so the mistake surfaces at\n * wiring time, mirroring the embedder / image guards.\n *\n * @example\n * const tts = new OpenAISpeechModel(client, { name: \"tts-1\", voice: \"alloy\" }, \"openai\");\n * const { audio } = await tts.generate(\"Welcome aboard.\");\n */\nexport class OpenAISpeechModel implements SpeechModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: SpeechModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly defaultVoice?: string;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAISpeechConfig, provider: string = \"openai\") {\n    if (!isOpenAISpeechModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI text-to-speech model. ` +\n          \"Use a `tts-1` / `tts-1-hd` / `gpt-4o-mini-tts` model with openai.speech({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.defaultVoice = config.voice;\n  }\n\n  public async generate(text: string, options?: SpeechOptions): Promise<SpeechGenerationResponse> {\n    const format = options?.format ?? \"mp3\";\n\n    this.logger.debug(LOG_MODULE, \"speech.request\", \"audio.speech.create\", {\n      model: this.name,\n      characters: text.length,\n    });\n\n    let response: Response;\n\n    try {\n      response = await this.client.audio.speech.create(\n        {\n          model: this.name,\n          input: text,\n          voice: options?.voice ?? this.defaultVoice ?? \"alloy\",\n          response_format: format as OpenAI.Audio.SpeechCreateParams[\"response_format\"],\n          ...(options?.speed !== undefined ? { speed: options.speed } : {}),\n          ...(options?.instructions !== undefined ? { instructions: options.instructions } : {}),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n      this.logger.error(LOG_MODULE, \"speech.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n      throw wrapped;\n    }\n\n    const base64 = Buffer.from(await response.arrayBuffer()).toString(\"base64\");\n\n    return {\n      audio: { type: \"base64\", base64, mediaType: audioMediaType(format) },\n      // The Speech API reports no token usage; spend is priced per\n      // character (or per token for gpt-4o-mini-tts) by `ai.speech()`.\n      usage: { input: 0, output: 0, total: 0 },\n      characters: text.length,\n    };\n  }\n}\n","import {\n  InvalidRequestError,\n  type AudioInput,\n  type TranscribeOptions,\n  type TranscriptionModelContract,\n  type TranscriptionModelPricing,\n  type TranscriptionResponse,\n  type TranscriptionSegment,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport OpenAI, { toFile } from \"openai\";\nimport type { OpenAITranscriptionConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Model-id prefixes OpenAI exposes through the **Transcription** (STT) API. */\nconst TRANSCRIPTION_MODEL_PREFIXES = [\"whisper\", \"gpt-4o-transcribe\", \"gpt-4o-mini-transcribe\"] as const;\n\n/** True when `name` is a recognized OpenAI speech-to-text model. */\nexport function isOpenAITranscriptionModel(name: string): boolean {\n  return TRANSCRIPTION_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n\n/** Defensive view over the response, whose shape varies by `response_format`. */\ntype RawTranscription = {\n  text: string;\n  duration?: number;\n  language?: string;\n  segments?: Array<{ text: string; start?: number; end?: number }>;\n  usage?: {\n    type?: string;\n    seconds?: number;\n    input_tokens?: number;\n    output_tokens?: number;\n    total_tokens?: number;\n  };\n};\n\n/**\n * OpenAI-backed implementation of `TranscriptionModelContract`\n * (speech-to-text) via `audio.transcriptions.create`. Consumed by the\n * `ai.transcribe()` verb.\n *\n * **Response format.** Defaults to `verbose_json` for `whisper-1` (so\n * the run gets a `duration` + timestamped `segments`) and `json` for\n * the token-metered `gpt-4o-transcribe` family. Base64 audio is wrapped\n * in an uploadable via the SDK's `toFile`.\n *\n * @example\n * const stt = new OpenAITranscriptionModel(client, { name: \"whisper-1\" }, \"openai\");\n * const { text } = await stt.transcribe({ base64, mediaType: \"audio/mpeg\" });\n */\nexport class OpenAITranscriptionModel implements TranscriptionModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: TranscriptionModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: OpenAI,\n    config: OpenAITranscriptionConfig,\n    provider: string = \"openai\",\n  ) {\n    if (!isOpenAITranscriptionModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI transcription model. ` +\n          \"Use a `whisper-1` / `gpt-4o-transcribe` / `gpt-4o-mini-transcribe` model with openai.transcribe({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n  }\n\n  public async transcribe(\n    audio: AudioInput,\n    options?: TranscribeOptions,\n  ): Promise<TranscriptionResponse> {\n    const isWhisper = this.name.startsWith(\"whisper\");\n    const format = options?.format ?? (isWhisper ? \"verbose_json\" : \"json\");\n\n    const file = await toFile(Buffer.from(audio.base64, \"base64\"), audio.filename ?? \"audio\", {\n      type: audio.mediaType,\n    });\n\n    this.logger.debug(LOG_MODULE, \"transcription.request\", \"audio.transcriptions.create\", {\n      model: this.name,\n      format,\n    });\n\n    let raw: unknown;\n\n    try {\n      raw = await this.client.audio.transcriptions.create(\n        {\n          model: this.name,\n          file,\n          response_format: format as OpenAI.Audio.TranscriptionCreateParams[\"response_format\"],\n          ...(options?.language ? { language: options.language } : {}),\n          ...(options?.prompt ? { prompt: options.prompt } : {}),\n        } as OpenAI.Audio.TranscriptionCreateParamsNonStreaming,\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n      this.logger.error(LOG_MODULE, \"transcription.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n      throw wrapped;\n    }\n\n    const response = raw as RawTranscription;\n\n    const segments: TranscriptionSegment[] | undefined = response.segments?.map((segment) => ({\n      text: segment.text,\n      ...(segment.start !== undefined ? { start: segment.start } : {}),\n      ...(segment.end !== undefined ? { end: segment.end } : {}),\n    }));\n\n    const durationSeconds =\n      response.duration ?? (response.usage?.type === \"duration\" ? response.usage.seconds : undefined);\n\n    const usage =\n      response.usage?.type === \"tokens\"\n        ? {\n            input: response.usage.input_tokens ?? 0,\n            output: response.usage.output_tokens ?? 0,\n            total: response.usage.total_tokens ?? 0,\n          }\n        : { input: 0, output: 0, total: 0 };\n\n    return {\n      text: response.text,\n      ...(segments && segments.length > 0 ? { segments } : {}),\n      ...(durationSeconds !== undefined ? { durationSeconds } : {}),\n      usage,\n    };\n  }\n}\n","import OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n  SpeechModelContract,\n  TranscriptionModelContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  OpenAIEmbedderConfig,\n  OpenAIImageConfig,\n  OpenAIModelConfig,\n  OpenAISDKConfig,\n  OpenAISpeechConfig,\n  OpenAITranscriptionConfig,\n} from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIImageModel } from \"./image\";\nimport { OpenAIModel } from \"./model\";\nimport { OpenAISpeechModel } from \"./speech\";\nimport { OpenAITranscriptionModel } from \"./transcription\";\n\n/**\n * OpenAI-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any OpenAI-compatible provider\n * (OpenAI, Azure OpenAI, OpenRouter, local gateways speaking the Chat\n * Completions protocol). A single `OpenAISDK` instance holds one live\n * `OpenAI` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per provider/account and reuse it\n * across all agents, workflows, and supervisors that target that\n * provider.\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client (authentication, base URL) and\n *   its lifetime scope. Factory for `OpenAIModel` instances — each\n *   model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `OpenAIModel` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `OpenAI` client is heavy to construct and\n * designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the PascalCase naming convention readers\n * expect from a constructor.\n *\n * @example\n * const openai = new OpenAISDK({ apiKey: process.env.OPENAI_API_KEY! });\n * const model = openai.model({ name: \"gpt-4o\", temperature: 0.7 });\n * const tokens = await openai.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.openai` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, persona, instruction, openai: new OpenAISDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.openai.model({ name: \"gpt-4o-mini\" }) });\n */\nexport class OpenAISDK implements SDKAdapterContract {\n  private readonly client: OpenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: OpenAISDKConfig) {\n    // Peel off the framework-only keys and forward every other upstream\n    // `ClientOptions` (timeout, maxRetries, defaultHeaders, fetch,\n    // organization, project, …) verbatim — they type-check, so dropping them\n    // is a silent footgun. Mirrors the Bedrock/Google/Ollama adapters.\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.client = new OpenAI(clientOptions);\n    this.provider = provider ?? \"openai\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build an `OpenAIModel` bound to this SDK's client. Each call returns\n   * a fresh model instance, but all instances share the underlying\n   * `OpenAI` client — connection pools, rate limits, and authentication\n   * state stay unified across every model produced here. The SDK's\n   * `provider` label is forwarded so every model self-identifies as\n   * coming from the same upstream.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: OpenAIModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses a\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Accepts an optional model id for future per-model tokenizer\n   * dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build an `OpenAIEmbedder` bound to this SDK's client. Each call\n   * returns a fresh embedder instance sharing the same underlying\n   * `OpenAI` client — connection pools and authentication stay unified\n   * across every embedder produced here.\n   *\n   * @example\n   * const embedder = openai.embedder({ name: \"text-embedding-3-small\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: OpenAIEmbedderConfig): EmbedderContract {\n    return new OpenAIEmbedder(this.client, config);\n  }\n\n  /**\n   * Build an `OpenAIImageModel` bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. Accepts the `gpt-image-*` (token-metered)\n   * and `dall-e-*` (per-image-metered) families; a non-image model id\n   * is rejected at construction.\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined` (no cost computed). A token-priced\n   * `gpt-image-1` entry can live in the same SDK registry as the chat\n   * models.\n   *\n   * @example\n   * const model = openai.image({ name: \"gpt-image-1\" });\n   * const { data } = await ai.image({ model, prompt: \"a red bicycle\" });\n   */\n  public image(config: OpenAIImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIImageModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Build an `OpenAISpeechModel` (text-to-speech) bound to this SDK's\n   * client, for use with `ai.speech({ model, text })`. Accepts the\n   * `tts-1` / `gpt-4o-mini-tts` families; a non-TTS model id is rejected\n   * at construction.\n   *\n   * @example\n   * const tts = openai.speech({ name: \"tts-1\", voice: \"alloy\" });\n   * const { data } = await ai.speech({ model: tts, text: \"Hello\" });\n   */\n  public speech(config: OpenAISpeechConfig): SpeechModelContract {\n    return new OpenAISpeechModel(this.client, config, this.provider);\n  }\n\n  /**\n   * Build an `OpenAITranscriptionModel` (speech-to-text) bound to this\n   * SDK's client, for use with `ai.transcribe({ model, audio })`.\n   * Accepts the `whisper-1` / `gpt-4o-transcribe` families; a non-STT\n   * model id is rejected at construction.\n   *\n   * @example\n   * const stt = openai.transcribe({ name: \"whisper-1\" });\n   * const { data } = await ai.transcribe({ model: stt, audio });\n   */\n  public transcribe(config: OpenAITranscriptionConfig): TranscriptionModelContract {\n    return new OpenAITranscriptionModel(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAEA,MAAM,kBAAgD;CACpD,MAAM;CACN,YAAY;CACZ,QAAQ;AACV;;;;;;;;;;AAWA,SAAgB,gBAAgB,KAA8C;CAC5E,OAAO,gBAAgB,OAAO,OAAO;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACQA,SAAgB,iBACd,UACsD;CACtD,OAAO,SAAS,KAAK,MAAM;EACzB,IAAI,EAAE,SAAS,QACb,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,cAAc,EAAE,cAAc;EAChC;EAEF,IAAI,EAAE,SAAS,eAAe,EAAE,aAAa,EAAE,UAAU,SAAS,GAChE,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,YAAY,EAAE,UAAU,KAAK,QAAQ;IACnC,IAAI,GAAG;IACP,MAAM;IACN,UAAU;KAAE,MAAM,GAAG;KAAM,WAAW,KAAK,UAAU,GAAG,SAAS,CAAC,CAAC;IAAE;GACvE,EAAE;EACJ;EAGF,IAAI,EAAE,SAAS,UAAU,MAAM,QAAQ,EAAE,OAAO,GAC9C,OAAO;GACL,MAAM;GACN,SAAS,EAAE,QAAQ,IAAI,mBAAmB;EAC5C;EAGF,OAAO;GAAE,MAAM,EAAE;GAAM,SAAS,iBAAiB,EAAE,OAAO;EAAE;CAI9D,CAAC;AACH;;;;;;;AAQA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,KAAK,SAAS,SAMhB,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;CAGjD,IAAI,KAAK,SAAS,OAAO;EACvB,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,8FACF;EAGF,OAAO;GACL,MAAM;GACN,MAAM;IACJ,UAAU;IACV,WAAW,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO;GACjE;EACF;CACF;CAGA,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,2FACF;CAGF,OAAO;EACL,MAAM;EACN,aAAa;GACX,MAAM,KAAK,OAAO;GAClB,QAAQ,oBAAoB,KAAK,OAAO,SAAS;EACnD;CACF;AACF;;;;;;AAOA,SAAS,oBAAoB,WAAkC;CAC7D,IAAI,cAAc,eAAe,cAAc,iBAAiB,cAAc,cAC5E,OAAO;CAGT,IAAI,cAAc,eAAe,cAAc,gBAAgB,cAAc,cAC3E,OAAO;CAGT,MAAM,IAAIA,mCACR,0DAA0D,UAAU,GACtE;AACF;;;;;;;;;;;;;ACvJA,SAAgB,cACd,OAC0D;CAC1D,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,MAAM,KAAK,UAAU;EAC1B,MAAM;EACN,UAAU;GACR,MAAM,KAAK;GACX,aAAa,KAAK;GAClB,YAAY,aAAa,KAAK,KAAK;EACrC;CACF,EAAE;AACJ;;;;;;;AAQA,SAAS,aAAa,OAAuE;CAC3F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO;EAAE,MAAM;EAAU,YAAY,CAAC;CAAE;AAC1C;;;;;;;;;;;;;;;;;;;;;;;;;ACOA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkBC,wBACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,QAAQ,KAAK;CAC1C,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAE1F,IAAI,UAAU,QAAQ,KAAK,GACzB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,mBACzC,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,sBACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,uBACzC,OAAO,IAAIC,sCAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BACjB,OAAO,IAAIC,0CAA2B,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAG3E,IAAI,MAAM,SAAS,kBACjB,OAAO,IAAIC,kCAAmB,SAAS;EACrC,OAAO;EACP;EACA,QAAQ;CACV,CAAC;CAGH,IAAI,OAAO,MAAM,WAAW,YAAY,MAAM,UAAU,OAAO,MAAM,SAAS,KAC5E,OAAO,IAAIC,mCAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;;AAQA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkBC,eAAO,UAC3B,OAAO;EACL,QAAQ,OAAO;EACf,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,MAAM,OAAO;CACf;CAGF,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SACE,OAAO,IAAI,YAAY,YAAY,IAAI,YAAY,OAC9C,IAAI,UACL;GACN,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAClD;CACF;CAEA,OAAO,CAAC;AACV;;;;;;;AAQA,SAAS,UAAU,QAAiB,OAAkC;CACpE,IAAI,kBAAkBA,eAAO,2BAC3B,OAAO;CAGT,IAAI,MAAM,SAAS,6BACjB,OAAO;CAGT,IAAI,MAAM,SAAS,eAAe,MAAM,SAAS,gBAC/C,OAAO;CAGT,OAAO;AACT;;;;;;;AAQA,SAAS,aACP,QACA,OACyB;CACzB,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,MAAM,YAAY,cAAc,MAAM;CAEtC,IAAI,WACF,QAAQ,YAAY;CAGtB,OAAO;AACT;;;;;AAMA,SAAS,cAAc,QAAqC;CAC1D,IAAI,OAAO,WAAW,YAAY,WAAW,MAC3C;CAGF,MAAM,MAAM;CAEZ,IAAI,OAAO,IAAI,eAAe,UAC5B,OAAO,IAAI;CAGb,IAAI,OAAO,IAAI,cAAc,UAC3B,OAAO,IAAI;AAIf;;;;;;AAOA,SAAS,gBAAgB,SAAiE;CACxF,IAAI,CAAC,SACH;CAGF,MAAM,MAAM,QAAQ,kBAAkB,QAAQ;CAE9C,IAAI,CAAC,KACH;CAGF,MAAM,UAAU,OAAO,GAAG;CAE1B,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GACzC;CAGF,OAAO,KAAK,MAAM,UAAU,GAAI;AAClC;;;;AChOA,MAAMC,eAAa;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BnB,IAAa,iBAAb,MAAwD;CAetD,AAAO,YAAY,QAAgB,QAA8B;kBAbtC;gBAWOC;EAGhC,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,uBAAuB,OAAO;EACnC,KAAK,aAAa,OAAO,cAAc;CACzC;CAEA,MAAa,MAAM,OAAyC;EAC1D,MAAM,EAAE,UAAU,UAAU,MAAM,KAAK,QAAQ,KAAK;EAEpD,OAAO;GACL,QAAQ,SAAS,KAAK,EAAE,CAAC;GACzB,YAAY,KAAK;GACjB;EACF;CACF;CAEA,MAAa,UAAU,QAAiD;EACtE,MAAM,EAAE,UAAU,UAAU,MAAM,KAAK,QAAQ,MAAM;EAErD,OAAO;GACL,SAAS,SAAS,KAAK,KAAK,MAAM,EAAE,SAAS;GAC7C,YAAY,KAAK;GACjB;EACF;CACF;;;;;;;CAQA,MAAc,QAAQ,OAGnB;EACD,KAAK,OAAO,MAAMD,cAAY,oBAAoB,qBAAqB;GACrE,OAAO,KAAK;GACZ,OAAO,MAAM,QAAQ,KAAK;GAC1B,OAAO,MAAM,QAAQ,KAAK,IAAI,MAAM,SAAS;EAC/C,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,WAAW,OAAO;IAC7C,OAAO,KAAK;IACZ;IACA,GAAI,KAAK,yBAAyB,SAC9B,EAAE,YAAY,KAAK,qBAAqB,IACxC,CAAC;GACP,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,kBAAkB,QAAQ,SAAS;IAC/D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,KAAK,OAAO,MAAMA,cAAY,qBAAqB,8BAA8B;GAC/E,YAAY,SAAS,KAAK,EAAE,EAAE,UAAU;GACxC,OAAO;IACL,cAAc,SAAS,MAAM;IAC7B,aAAa,SAAS,MAAM;GAC9B;EACF,CAAC;EAID,IAAI,KAAK,eAAe,GACtB,KAAK,aAAa,SAAS,KAAK,EAAE,CAAC,UAAU;EAG/C,MAAM,QAAwB;GAC5B,cAAc,SAAS,MAAM;GAC7B,aAAa,SAAS,MAAM;EAC9B;EAEA,OAAO;GAAE;GAAU;EAAM;CAC3B;AACF;;;;;;;;;;;;;;;;;ACjIA,MAAa,8BAA8B,CAAC,aAAa,QAAQ;;;;;;;;;;;;AAajE,SAAgB,mBAAmB,MAAuB;CACxD,OAAO,4BAA4B,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AAC7E;;;;ACbA,MAAME,eAAa;;AAGnB,SAAS,aAAa,QAAoC;CACxD,QAAQ,QAAR;EACE,KAAK;EACL,KAAK,OACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,SACE,OAAO;CACX;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,IAAa,mBAAb,MAA4D;CAQ1D,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzDC;EAGhC,IAAI,CAAC,mBAAmB,OAAO,IAAI,GACjC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,yHAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;CACxB;CAEA,MAAa,SACX,QACA,SACkC;EAClC,MAAM,aAAa,KAAK,KAAK,WAAW,WAAW;EAInD,MAAM,iBACH,SAAS,mBACT,aAAa,SAAY;EAE5B,MAAM,OAAsD;GAC1D,OAAO,KAAK;GACZ;EACF;EAEA,IAAI,SAAS,UAAU,QAAW,KAAK,IAAI,QAAQ;EACnD,IAAI,SAAS,SAAS,QAAW,KAAK,OAAO,QAAQ;EACrD,IAAI,SAAS,YAAY,QACvB,KAAK,UAAU,QAAQ;EAEzB,IAAI,CAAC,cAAc,gBAAgB,KAAK,kBAAkB;EAC1D,IAAI,cAAc,SAAS,WAAW,QACpC,KAAK,gBAAgB,QAAQ;EAE/B,IAAI,SAAS,eAAe,QAC1B,KAAK,aAAa,QAAQ;EAG5B,KAAK,OAAO,MAAMF,cAAY,iBAAiB,mBAAmB;GAChE,OAAO,KAAK;GACZ,OAAO,SAAS,SAAS;EAC3B,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,OAAO,SAClC,MACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,eAAe,QAAQ,SAAS;IAC5D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,UAAU,SAAS,QAAQ,CAAC,EAAC,CAAE,KAAK,UACxC,KAAK,iBAAiB,OAAO,SAAS,MAAM,CAC9C;EAEA,MAAM,QAAQ,SAAS,QACnB;GACE,OAAO,SAAS,MAAM;GACtB,QAAQ,SAAS,MAAM;GACvB,OAAO,SAAS,MAAM;EACxB,IACA;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAEpC,KAAK,OAAO,MAAMA,cAAY,kBAAkB,6BAA6B;GAC3E,QAAQ,OAAO;GACf;EACF,CAAC;EAED,OAAO;GAAE;GAAQ;EAAM;CACzB;;;;;;;;CASA,AAAQ,iBAAiB,OAA4B,QAA4C;EAC/F,IAAI,MAAM,UACR,OAAO;GACL,MAAM;GACN,QAAQ,MAAM;GACd,WAAW,aAAa,MAAM;GAC9B,GAAI,MAAM,iBAAiB,EAAE,eAAe,MAAM,eAAe,IAAI,CAAC;EACxE;EAGF,IAAI,MAAM,KACR,OAAO;GACL,MAAM;GACN,KAAK,MAAM;GACX,GAAI,MAAM,iBAAiB,EAAE,eAAe,MAAM,eAAe,IAAI,CAAC;EACxE;EAGF,MAAM,IAAIG,6BAAc,iEAAiE;CAC3F;AACF;;;;;;;;;;;;;;;;;;ACpKA,MAAM,6BAA6B;CAAC;CAAM;CAAM;CAAM;AAAO;;;;;;;;;;;;;;AAe7D,SAAgB,yBAAyB,WAA4B;CACnE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,2BAA2B,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAClF;;;;;;;;;;;;;;;;;ACpBA,MAAM,0BAA0B;CAC9B;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;AAcA,SAAgB,sBAAsB,WAA4B;CAChE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,wBAAwB,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAC/E;;;;ACnBA,MAAMC,eAAa;;;;;;;;;;AAWnB,SAAS,sBAAsB,gBAA2D;CACxF,IAAI,mBAAmB,iBAAiB,mBAAmB,QACzD,OAAO;CAGT,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzDC;EAGhC,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;GASnE,eAAe;GAMf,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,KAAK,OAAO,MAAMD,cAAY,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,SAAS,SAAS,QAAQ;EAChC,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAMA,cAAY,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,MAAMA,cAAY,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,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,yCAA8C,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAMA,cAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4BA,AAAQ,oBAAoB,gBAE1B;EACA,IAAI,CAAC,gBACH,OAAO,CAAC;EAGV,MAAM,WAAW,KAAK,OAAO;EAE7B,IAAI,aAAa,QACf,OAAO,CAAC;EAGV,IAAI,aAAa,eACf,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAOpD,IAAI,KAAK,mBAAmB,cAAc,GACxC,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IACX,MAAM;IACN,QAAQ;IACR,QAAQ;GACV;EACF,EACF;EAGF,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;CACpD;;;;;;;;;CAUA,AAAQ,mBAAmB,QAA0C;EACnE,OACE,OAAO,SAAS,YAChB,OAAO,OAAO,eAAe,YAC7B,OAAO,eAAe,QACtB,KAAK,iBAAiB,MAAM;CAEhC;;;;;;;;;;;;CAaA,AAAQ,iBAAiB,MAAwB;EAC/C,IAAI,CAAC,QAAQ,OAAO,SAAS,UAC3B,OAAO;EAGT,MAAM,SAAS;EAEf,IAAI,OAAO,SAAS,YAAY,OAAO,cAAc,OAAO,OAAO,eAAe,UAAU;GAC1F,MAAM,aAAa,OAAO;GAC1B,MAAM,OAAO,OAAO,KAAK,UAAU;GACnC,MAAM,WAAW,MAAM,QAAQ,OAAO,QAAQ,IAAK,OAAO,WAAyB,CAAC;GAEpF,IAAI,KAAK,MAAM,QAAQ,CAAC,SAAS,SAAS,GAAG,CAAC,GAC5C,OAAO;GAGT,KAAK,MAAM,OAAO,MAChB,IAAI,CAAC,KAAK,iBAAiB,WAAW,IAAI,GACxC,OAAO;EAGb;EAEA,IAAI,OAAO,UAAU,UAAa,CAAC,KAAK,iBAAiB,OAAO,KAAK,GACnE,OAAO;EAGT,KAAK,MAAM,UAAU;GAAC;GAAS;GAAS;EAAO,GAAY;GACzD,MAAM,QAAQ,OAAO;GACrB,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,MAAM,QAAQ,CAAC,KAAK,iBAAiB,GAAG,CAAC,GACzE,OAAO;EAEX;EAEA,OAAO;CACT;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAChD,MAAM,kBAAkB,IAAI,2BAA2B;EAEvD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;GACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EACpF;CACF;;;;;;;;;;;;;;;;;;CAmBA,AAAQ,qBAAqB,WAE3B;EACA,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WAAW,QAC9C,OAAO,CAAC;EAGV,OAAO,EAAE,kBAAkB,UAAU,OAAO;CAC9C;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,yCAA+C,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF;;;;ACnfA,MAAME,eAAa;;AAGnB,MAAM,wBAAwB;CAAC;CAAS;CAAmB;AAAW;;AAGtE,SAAgB,oBAAoB,MAAuB;CACzD,OAAO,sBAAsB,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AACvE;;AAGA,SAAS,eAAe,QAAoC;CAC1D,QAAQ,QAAR;EACE,KAAK,QACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,SACE,OAAO;CACX;AACF;;;;;;;;;;;;;;AAeA,IAAa,oBAAb,MAA8D;CAS5D,AAAO,YAAY,QAAgB,QAA4B,WAAmB,UAAU;gBAF1DC;EAGhC,IAAI,CAAC,oBAAoB,OAAO,IAAI,GAClC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,uIAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe,OAAO;CAC7B;CAEA,MAAa,SAAS,MAAc,SAA4D;EAC9F,MAAM,SAAS,SAAS,UAAU;EAElC,KAAK,OAAO,MAAMF,cAAY,kBAAkB,uBAAuB;GACrE,OAAO,KAAK;GACZ,YAAY,KAAK;EACnB,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,MAAM,OAAO,OACxC;IACE,OAAO,KAAK;IACZ,OAAO;IACP,OAAO,SAAS,SAAS,KAAK,gBAAgB;IAC9C,iBAAiB;IACjB,GAAI,SAAS,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;IAC/D,GAAI,SAAS,iBAAiB,SAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;GACtF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GACtC,KAAK,OAAO,MAAMA,cAAY,gBAAgB,QAAQ,SAAS;IAC7D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GACD,MAAM;EACR;EAIA,OAAO;GACL,OAAO;IAAE,MAAM;IAAU,QAHZ,OAAO,KAAK,MAAM,SAAS,YAAY,CAAC,CAAC,CAAC,SAAS,QAGlC;IAAG,WAAW,eAAe,MAAM;GAAE;GAGnE,OAAO;IAAE,OAAO;IAAG,QAAQ;IAAG,OAAO;GAAE;GACvC,YAAY,KAAK;EACnB;CACF;AACF;;;;ACxGA,MAAM,aAAa;;AAGnB,MAAM,+BAA+B;CAAC;CAAW;CAAqB;AAAwB;;AAG9F,SAAgB,2BAA2B,MAAuB;CAChE,OAAO,6BAA6B,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AAC9E;;;;;;;;;;;;;;;AA+BA,IAAa,2BAAb,MAA4E;CAQ1E,AAAO,YACL,QACA,QACA,WAAmB,UACnB;gBANgCG;EAOhC,IAAI,CAAC,2BAA2B,OAAO,IAAI,GACzC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,8JAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;CACxB;CAEA,MAAa,WACX,OACA,SACgC;EAChC,MAAM,YAAY,KAAK,KAAK,WAAW,SAAS;EAChD,MAAM,SAAS,SAAS,WAAW,YAAY,iBAAiB;EAEhE,MAAM,OAAO,yBAAa,OAAO,KAAK,MAAM,QAAQ,QAAQ,GAAG,MAAM,YAAY,SAAS,EACxF,MAAM,MAAM,UACd,CAAC;EAED,KAAK,OAAO,MAAM,YAAY,yBAAyB,+BAA+B;GACpF,OAAO,KAAK;GACZ;EACF,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,MAAM,MAAM,KAAK,OAAO,MAAM,eAAe,OAC3C;IACE,OAAO,KAAK;IACZ;IACA,iBAAiB;IACjB,GAAI,SAAS,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;IAC1D,GAAI,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GACtD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GACtC,KAAK,OAAO,MAAM,YAAY,uBAAuB,QAAQ,SAAS;IACpE,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GACD,MAAM;EACR;EAEA,MAAM,WAAW;EAEjB,MAAM,WAA+C,SAAS,UAAU,KAAK,aAAa;GACxF,MAAM,QAAQ;GACd,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;GAC9D,GAAI,QAAQ,QAAQ,SAAY,EAAE,KAAK,QAAQ,IAAI,IAAI,CAAC;EAC1D,EAAE;EAEF,MAAM,kBACJ,SAAS,aAAa,SAAS,OAAO,SAAS,aAAa,SAAS,MAAM,UAAU;EAEvF,MAAM,QACJ,SAAS,OAAO,SAAS,WACrB;GACE,OAAO,SAAS,MAAM,gBAAgB;GACtC,QAAQ,SAAS,MAAM,iBAAiB;GACxC,OAAO,SAAS,MAAM,gBAAgB;EACxC,IACA;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAEtC,OAAO;GACL,MAAM,SAAS;GACf,GAAI,YAAY,SAAS,SAAS,IAAI,EAAE,SAAS,IAAI,CAAC;GACtD,GAAI,oBAAoB,SAAY,EAAE,gBAAgB,IAAI,CAAC;GAC3D;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACrFA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAK1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,SAAS,IAAIC,eAAO,aAAa;EACtC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;;;CAcA,AAAO,MAAM,QAA0C;EACrD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,YAAY,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACnE;;;;;;;;CASA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;;;;;;;;;;;CAYA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,QAAQ,MAAM;CAC/C;;;;;;;;;;;;;;;;;CAkBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,iBAAiB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACxE;;;;;;;;;;;CAYA,AAAO,OAAO,QAAiD;EAC7D,OAAO,IAAI,kBAAkB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACjE;;;;;;;;;;;CAYA,AAAO,WAAW,QAA+D;EAC/E,OAAO,IAAI,yBAAyB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACxE;AACF"}
+{"version":3,"file":"index.cjs","names":["InvalidRequestError","AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","ContentFilterError","InvalidRequestError","ProviderError","OpenAI","LOG_MODULE","log","LOG_MODULE","log","InvalidRequestError","ProviderError","LOG_MODULE","log","LOG_MODULE","log","InvalidRequestError","log","InvalidRequestError","OpenAI"],"sources":["../../../../../../@warlock.js/ai-openai/src/utils/map-finish-reason.ts","../../../../../../@warlock.js/ai-openai/src/utils/to-openai-messages.ts","../../../../../../@warlock.js/ai-openai/src/utils/to-openai-tools.ts","../../../../../../@warlock.js/ai-openai/src/utils/wrap-openai-error.ts","../../../../../../@warlock.js/ai-openai/src/embedder.ts","../../../../../../@warlock.js/ai-openai/src/known-image-models.ts","../../../../../../@warlock.js/ai-openai/src/image.ts","../../../../../../@warlock.js/ai-openai/src/known-reasoning-models.ts","../../../../../../@warlock.js/ai-openai/src/known-vision-models.ts","../../../../../../@warlock.js/ai-openai/src/model.ts","../../../../../../@warlock.js/ai-openai/src/speech.ts","../../../../../../@warlock.js/ai-openai/src/transcription.ts","../../../../../../@warlock.js/ai-openai/src/sdk.ts"],"sourcesContent":["import type { FinishReason } from \"@warlock.js/ai\";\n\nconst finishReasonMap: Record<string, FinishReason> = {\n  stop: \"stop\",\n  tool_calls: \"tool_calls\",\n  length: \"length\",\n};\n\n/**\n * Map the raw OpenAI `finish_reason` string to the normalized FinishReason union.\n * Unknown/unexpected values fall through to \"error\".\n *\n * @example\n * mapFinishReason(\"stop\");        // \"stop\"\n * mapFinishReason(\"tool_calls\");  // \"tool_calls\"\n * mapFinishReason(null);          // \"error\"\n */\nexport function mapFinishReason(raw: string | null | undefined): FinishReason {\n  return finishReasonMap[raw ?? \"\"] ?? \"error\";\n}\n","import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral Message[] to OpenAI's chat message shape.\n * Handles the `tool` role (requires `tool_call_id`) and assistant messages\n * that carry `toolCalls` from a prior model response.\n *\n * Multipart `content` (a `ContentPart[]`) is mapped into OpenAI's user-message\n * content-parts shape: text becomes `{ type: \"text\", text }`, images become\n * `{ type: \"image_url\", image_url: { url } }` — with base64 sources rendered\n * as `data:` URLs inline.\n *\n * @example\n * const openaiMessages = toOpenAIMessages([\n *   { role: \"user\", content: \"Hi\" },\n *   { role: \"tool\", toolCallId: \"call_1\", content: '{\"ok\":true}' },\n * ]);\n *\n * @example\n * toOpenAIMessages([\n *   { role: \"user\", content: [\n *     { type: \"text\", text: \"What is this?\" },\n *     { type: \"image\", source: { url: \"https://example.com/cat.jpg\" } },\n *   ]},\n * ]);\n */\nexport function toOpenAIMessages(\n  messages: Message[],\n): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {\n  return messages.map((m) => {\n    if (m.role === \"tool\") {\n      return {\n        role: \"tool\",\n        content: stringifyContent(m.content),\n        tool_call_id: m.toolCallId ?? \"\",\n      };\n    }\n    if (m.role === \"assistant\" && m.toolCalls && m.toolCalls.length > 0) {\n      return {\n        role: \"assistant\",\n        content: stringifyContent(m.content),\n        tool_calls: m.toolCalls.map((tc) => ({\n          id: tc.id,\n          type: \"function\" as const,\n          function: { name: tc.name, arguments: JSON.stringify(tc.input ?? {}) },\n        })),\n      };\n    }\n\n    if (m.role === \"user\" && Array.isArray(m.content)) {\n      return {\n        role: \"user\",\n        content: m.content.map(toOpenAIContentPart),\n      };\n    }\n\n    return { role: m.role, content: stringifyContent(m.content) } as\n      | OpenAI.Chat.Completions.ChatCompletionUserMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionSystemMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionAssistantMessageParam;\n  });\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role (system / assistant text / tool), collapse a `ContentPart[]` to\n * its concatenated text so OpenAI's wire format stays valid. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to an OpenAI chat content part — one\n * branch per modality, each to its real wire shape:\n *\n * - `text` → `{ type: \"text\" }`.\n * - `image` → `{ type: \"image_url\" }` (remote URL, or a `data:` URL for\n *   inlined base64 bytes).\n * - `pdf` → `{ type: \"file\", file: { file_data } }` (OpenAI document\n *   input; base64 only — there is no remote-URL file source).\n * - `audio` → `{ type: \"input_audio\", input_audio: { data, format } }`\n *   (base64 only; `wav` / `mp3` are the only formats OpenAI accepts).\n *\n * PDF and audio reach this point ONLY when the model declared the\n * matching capability (`openai.model({ name, pdf: true })` /\n * `{ audio: true }`) — the agent's modality gate throws upfront\n * otherwise, so capability and behavior stay in lockstep. A remote-URL\n * pdf/audio source raises a typed `InvalidRequestError` here rather\n * than a downstream provider fault.\n */\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (part.type === \"image\") {\n    const url =\n      \"url\" in part.source\n        ? part.source.url\n        : `data:${part.source.mediaType};base64,${part.source.base64}`;\n\n    return { type: \"image_url\", image_url: { url } };\n  }\n\n  if (part.type === \"pdf\") {\n    if (\"url\" in part.source) {\n      throw new InvalidRequestError(\n        \"OpenAI chat completions cannot fetch a remote-URL PDF; supply base64 document bytes instead.\",\n      );\n    }\n\n    return {\n      type: \"file\",\n      file: {\n        filename: \"document.pdf\",\n        file_data: `data:${part.source.mediaType};base64,${part.source.base64}`,\n      },\n    };\n  }\n\n  // Audio — the remaining `ContentPart` variant.\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"OpenAI chat completions cannot fetch remote-URL audio; supply base64 audio bytes instead.\",\n    );\n  }\n\n  return {\n    type: \"input_audio\",\n    input_audio: {\n      data: part.source.base64,\n      format: toOpenAIAudioFormat(part.source.mediaType),\n    },\n  };\n}\n\n/**\n * Narrow a neutral audio media type to the two formats OpenAI's\n * `input_audio` accepts (`wav` / `mp3`). An unsupported type raises a\n * typed `InvalidRequestError` up front rather than a provider 400.\n */\nfunction toOpenAIAudioFormat(mediaType: string): \"wav\" | \"mp3\" {\n  if (mediaType === \"audio/wav\" || mediaType === \"audio/x-wav\" || mediaType === \"audio/wave\") {\n    return \"wav\";\n  }\n\n  if (mediaType === \"audio/mp3\" || mediaType === \"audio/mpeg\" || mediaType === \"audio/mpga\") {\n    return \"mp3\";\n  }\n\n  throw new InvalidRequestError(\n    `OpenAI input_audio supports only \"wav\" and \"mp3\"; got \"${mediaType}\".`,\n  );\n}\n","import { extractJsonSchema, type ToolConfig } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral ToolConfig[] to OpenAI's tools array.\n * Uses the shared `extractJsonSchema` helper; falls back to an empty-object\n * schema when extraction fails so the tool still registers with the provider.\n *\n * @example\n * const tools = toOpenAITools([weatherTool, calculatorTool]);\n * await client.chat.completions.create({ model, messages, tools });\n */\nexport function toOpenAITools(\n  tools: ToolConfig<unknown, unknown>[] | undefined,\n): OpenAI.Chat.Completions.ChatCompletionTool[] | undefined {\n  if (!tools || tools.length === 0) {\n    return undefined;\n  }\n\n  return tools.map((tool) => ({\n    type: \"function\",\n    function: {\n      name: tool.name,\n      description: tool.description,\n      parameters: toParameters(tool.input),\n    },\n  }));\n}\n\n/**\n * Resolve a tool's input schema to a JSON-Schema object. OpenAI's\n * function `parameters` expects an object root; anything else (or a\n * failed extraction) degrades to an empty-object schema so the tool\n * still registers and the model simply sees no parameters.\n */\nfunction toParameters(input: ToolConfig<unknown, unknown>[\"input\"]): Record<string, unknown> {\n  const schema = extractJsonSchema(input);\n\n  if (schema && schema.type === \"object\") {\n    return schema;\n  }\n\n  return { type: \"object\", properties: {} };\n}\n","import {\n  AIError,\n  ContentFilterError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n  QuotaExceededError,\n} from \"@warlock.js/ai\";\nimport OpenAI from \"openai\";\n\n/**\n * Raw-error fields the wrapper reads off an OpenAI SDK error.\n *\n * `APIError` exposes `status`, `code`, `message`, `type`, `headers` —\n * we duck-type because wrapped retries, proxied errors, and custom\n * error subclasses sometimes lose the `instanceof` relationship.\n */\ntype OpenAIErrorShape = {\n  status?: number;\n  code?: string | null;\n  message?: string;\n  type?: string | null;\n  headers?: Record<string, string> | undefined;\n  name?: string;\n};\n\n/**\n * Wrap any thrown value caught inside the OpenAI adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Prefers `APIError.code` when present (stable\n * machine identifier across SDK versions), falls back to `status` when\n * `code` is missing (common with proxied deployments that strip the\n * field). Name-based detection (`APIConnectionTimeoutError`) catches\n * transport-layer errors that never produced an HTTP response.\n *\n * `AIError` instances are returned unchanged — callers can pass the\n * error through `try/catch/throw wrap(e)` pipelines without accidental\n * double-wrapping.\n *\n * @example\n * try {\n *   return await this.client.chat.completions.create(...);\n * } catch (thrown) {\n *   throw wrapOpenAIError(thrown);\n * }\n */\nexport function wrapOpenAIError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(thrown, shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n\n  if (isTimeout(thrown, shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 401 || shape.code === \"invalid_api_key\") {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.code === \"insufficient_quota\") {\n    return new QuotaExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 429 || shape.code === \"rate_limit_exceeded\") {\n    return new ProviderRateLimitError(message, {\n      cause: thrown,\n      context,\n      retryAfter: parseRetryAfter(shape.headers),\n    });\n  }\n\n  if (shape.code === \"context_length_exceeded\") {\n    return new ContextLengthExceededError(message, { cause: thrown, context });\n  }\n\n  if (shape.code === \"content_filter\") {\n    return new ContentFilterError(message, {\n      cause: thrown,\n      context,\n      reason: message,\n    });\n  }\n\n  if (typeof shape.status === \"number\" && shape.status >= 400 && shape.status < 500) {\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  return new ProviderError(message, { cause: thrown, context });\n}\n\n/**\n * Read the raw error shape without depending on `instanceof APIError`\n * — some consumers wrap the SDK, and proxies sometimes strip the\n * prototype chain. Duck-typing on the visible fields is resilient to\n * both.\n */\nfunction toShape(thrown: unknown): OpenAIErrorShape {\n  if (thrown instanceof OpenAI.APIError) {\n    return {\n      status: thrown.status,\n      code: thrown.code,\n      message: thrown.message,\n      type: thrown.type,\n      headers: thrown.headers as Record<string, string> | undefined,\n      name: thrown.name,\n    };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      code: typeof raw.code === \"string\" ? raw.code : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      type: typeof raw.type === \"string\" ? raw.type : undefined,\n      headers:\n        typeof raw.headers === \"object\" && raw.headers !== null\n          ? (raw.headers as Record<string, string>)\n          : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the thrown value represents a timeout. OpenAI's SDK\n * throws `APIConnectionTimeoutError` for transport-level timeouts, and\n * Node surfaces `ETIMEDOUT` / `ECONNABORTED` on the lower socket\n * layer. Either signal counts.\n */\nfunction isTimeout(thrown: unknown, shape: OpenAIErrorShape): boolean {\n  if (thrown instanceof OpenAI.APIConnectionTimeoutError) {\n    return true;\n  }\n\n  if (shape.name === \"APIConnectionTimeoutError\") {\n    return true;\n  }\n\n  if (shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\") {\n    return true;\n  }\n\n  return false;\n}\n\n/**\n * Attach the raw diagnostic fields to `error.context` so consumers\n * have everything the provider surfaced without each subclass having\n * to redeclare them. Never includes `cause` — that lives on\n * `error.cause`.\n */\nfunction buildContext(\n  thrown: unknown,\n  shape: OpenAIErrorShape,\n): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.code) {\n    context.code = shape.code;\n  }\n\n  if (shape.type) {\n    context.type = shape.type;\n  }\n\n  const requestId = readRequestId(thrown);\n\n  if (requestId) {\n    context.requestId = requestId;\n  }\n\n  return context;\n}\n\n/**\n * OpenAI puts the request id on `APIError.request_id`. Extract\n * defensively — both camel and snake keys exist across SDK versions.\n */\nfunction readRequestId(thrown: unknown): string | undefined {\n  if (typeof thrown !== \"object\" || thrown === null) {\n    return undefined;\n  }\n\n  const raw = thrown as Record<string, unknown>;\n\n  if (typeof raw.request_id === \"string\") {\n    return raw.request_id;\n  }\n\n  if (typeof raw.requestId === \"string\") {\n    return raw.requestId;\n  }\n\n  return undefined;\n}\n\n/**\n * Parse the `Retry-After` response header (seconds per HTTP spec)\n * into milliseconds so consumers can feed it straight to `setTimeout`.\n * Returns `undefined` when missing or unparseable.\n */\nfunction parseRetryAfter(headers: Record<string, string> | undefined): number | undefined {\n  if (!headers) {\n    return undefined;\n  }\n\n  const raw = headers[\"retry-after\"] ?? headers[\"Retry-After\"];\n\n  if (!raw) {\n    return undefined;\n  }\n\n  const seconds = Number(raw);\n\n  if (!Number.isFinite(seconds) || seconds < 0) {\n    return undefined;\n  }\n\n  return Math.round(seconds * 1000);\n}\n","import {\n  type EmbeddingBatchResult,\n  type EmbeddingResult,\n  type EmbeddingUsage,\n  type EmbedderContract,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIEmbedderConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\ntype EmbeddingsResponse = Awaited<\n  ReturnType<OpenAI[\"embeddings\"][\"create\"]>\n>;\n\n/**\n * OpenAI-backed implementation of `EmbedderContract`.\n *\n * **Role.** Converts text (or a batch of texts) into floating-point\n * vectors via OpenAI's Embeddings API. Standalone primitive — no\n * relationship to chat completions, tools, or the agent loop.\n *\n * **Dimensions.** When no `dimensions` override is supplied in config,\n * `this.dimensions` starts at `0` and is populated from the first\n * response's vector length, then cached for all subsequent calls —\n * even if a later response were to return a different length, the\n * first value wins so batches stay dimensionally consistent. Passing\n * `dimensions` in config both forwards the truncation hint to the API\n * (for models like `text-embedding-3-*`) and sets the initial value.\n *\n * **Error handling.** Raw OpenAI SDK errors are wrapped into the\n * typed `@warlock.js/ai` `AIError` hierarchy via `wrapOpenAIError` —\n * callers catch `AIError` subclasses (`ProviderRateLimitError`,\n * `ProviderAuthError`, etc.) instead of OpenAI's own classes.\n *\n * @example\n * const embedder = new OpenAIEmbedder(client, { name: \"text-embedding-3-small\" });\n * const { vector, dimensions, usage } = await embedder.embed(\"Hello world\");\n * const { vectors } = await embedder.embedMany([\"doc 1\", \"doc 2\"]);\n */\nexport class OpenAIEmbedder implements EmbedderContract {\n  public readonly name: string;\n  public readonly provider = \"openai\";\n  public dimensions: number;\n\n  private readonly client: OpenAI;\n\n  /**\n   * User-specified truncation hint, or `undefined` if omitted.\n   * Forwarded to the API on every call so OpenAI can truncate the\n   * embedding server-side for models that support it.\n   */\n  private readonly configuredDimensions: number | undefined;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIEmbedderConfig) {\n    this.client = client;\n    this.name = config.name;\n    this.configuredDimensions = config.dimensions;\n    this.dimensions = config.dimensions ?? 0;\n  }\n\n  public async embed(input: string): Promise<EmbeddingResult> {\n    const { response, usage } = await this.request(input);\n\n    return {\n      vector: response.data[0].embedding,\n      dimensions: this.dimensions,\n      usage,\n    };\n  }\n\n  public async embedMany(inputs: string[]): Promise<EmbeddingBatchResult> {\n    const { response, usage } = await this.request(inputs);\n\n    return {\n      vectors: response.data.map((d) => d.embedding),\n      dimensions: this.dimensions,\n      usage,\n    };\n  }\n\n  /**\n   * Shared transport for both `embed()` and `embedMany()` — issues the\n   * `embeddings.create` call, wraps provider errors, caches dimensions\n   * on the first successful response, and returns the raw response\n   * plus a camelCase usage object for the caller to shape.\n   */\n  private async request(input: string | string[]): Promise<{\n    response: EmbeddingsResponse;\n    usage: EmbeddingUsage;\n  }> {\n    this.logger.debug(LOG_MODULE, \"embedder.request\", \"embeddings.create\", {\n      model: this.name,\n      batch: Array.isArray(input),\n      count: Array.isArray(input) ? input.length : 1,\n    });\n\n    let response: EmbeddingsResponse;\n\n    try {\n      response = await this.client.embeddings.create({\n        model: this.name,\n        input,\n        ...(this.configuredDimensions !== undefined\n          ? { dimensions: this.configuredDimensions }\n          : {}),\n      });\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"embedder.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    this.logger.debug(LOG_MODULE, \"embedder.response\", \"embeddings.create returned\", {\n      dimensions: response.data[0]?.embedding.length,\n      usage: {\n        promptTokens: response.usage.prompt_tokens,\n        totalTokens: response.usage.total_tokens,\n      },\n    });\n\n    // Cache dimensions on the first response. Once set, stays set —\n    // we trust the first call to define the shape for this embedder.\n    if (this.dimensions === 0) {\n      this.dimensions = response.data[0].embedding.length;\n    }\n\n    const usage: EmbeddingUsage = {\n      promptTokens: response.usage.prompt_tokens,\n      totalTokens: response.usage.total_tokens,\n    };\n\n    return { response, usage };\n  }\n}\n","/**\n * Model-id prefixes OpenAI exposes through the **Images** API\n * (`client.images.generate`). The two live families:\n *\n * - `gpt-image-*` — token-metered, always returns base64 bytes (no\n *   `response_format` knob), supports `output_format` + `background`.\n * - `dall-e-*` — per-image-metered, returns a URL or base64 via\n *   `response_format`.\n *\n * Used by {@link isOpenAIImageModel} for the construction-time guard so\n * `openai.image({ name: \"gpt-4o\" })` fails fast with a curated error\n * instead of a downstream 400 — mirroring the embedder/vision guards.\n */\nexport const OPENAI_IMAGE_MODEL_PREFIXES = [\"gpt-image\", \"dall-e\"] as const;\n\n/**\n * True when `name` is a recognized OpenAI image-generation model. A\n * prefix match (not an exact list) so dated snapshots\n * (`gpt-image-1-mini`, `dall-e-3`) are covered without a maintenance\n * burden every time OpenAI ships a point release.\n *\n * @example\n * isOpenAIImageModel(\"gpt-image-1\"); // true\n * isOpenAIImageModel(\"dall-e-3\");    // true\n * isOpenAIImageModel(\"gpt-4o\");      // false\n */\nexport function isOpenAIImageModel(name: string): boolean {\n  return OPENAI_IMAGE_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n","import {\n  InvalidRequestError,\n  ProviderError,\n  type GeneratedImage,\n  type ImageGenerationOptions,\n  type ImageGenerationResponse,\n  type ImageModelContract,\n  type ImageModelPricing,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIImageConfig } from \"./config.type\";\nimport { isOpenAIImageModel } from \"./known-image-models\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Map a neutral output container to its IANA media type. */\nfunction mediaTypeFor(format: string | undefined): string {\n  switch (format) {\n    case \"jpeg\":\n    case \"jpg\":\n      return \"image/jpeg\";\n    case \"webp\":\n      return \"image/webp\";\n    default:\n      return \"image/png\";\n  }\n}\n\n/**\n * OpenAI-backed implementation of `ImageModelContract`.\n *\n * **Role.** Bridges the vendor-neutral `ai.image()` verb to OpenAI's\n * **Images** API for the two image families OpenAI ships: the\n * token-metered `gpt-image-*` models (always return base64 bytes) and\n * the per-image-metered `dall-e-*` models (URL or base64). Like\n * `OpenAIEmbedder`, it's a standalone primitive — no relationship to\n * chat completions, tools, or the agent loop.\n *\n * **Capability guard.** The constructor rejects a non-image model id\n * up front (`gpt-4o` → typed `InvalidRequestError`) so the mistake\n * surfaces at wiring time, not as a downstream provider 400 — the\n * \"fail fast at construction\" rule shared with the embedder/vision\n * guards.\n *\n * **Error handling.** Raw OpenAI SDK errors are wrapped into the typed\n * `@warlock.js/ai` `AIError` hierarchy via `wrapOpenAIError`, so a\n * caller catches `ProviderRateLimitError` / `ContentFilterError` /\n * `ProviderAuthError` rather than OpenAI's own classes. `ai.image()`\n * turns those throws into `result.error`.\n *\n * @example\n * const model = new OpenAIImageModel(client, { name: \"gpt-image-1\" }, \"openai\");\n * const { images, usage } = await model.generate(\"a teal ceramic mug, studio light\");\n */\nexport class OpenAIImageModel implements ImageModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: ImageModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIImageConfig, provider: string = \"openai\") {\n    if (!isOpenAIImageModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI image-generation model. ` +\n          \"Use a `gpt-image-*` or `dall-e-*` model with openai.image({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n  }\n\n  public async generate(\n    prompt: string,\n    options?: ImageGenerationOptions,\n  ): Promise<ImageGenerationResponse> {\n    const isGptImage = this.name.startsWith(\"gpt-image\");\n    // gpt-image always returns base64 bytes (no `response_format` knob);\n    // DALL·E defaults to self-contained base64 here (URLs expire in ~60\n    // min), but the caller can ask for a URL via `options.responseFormat`.\n    const responseFormat =\n      (options?.responseFormat as \"url\" | \"b64_json\" | undefined) ??\n      (isGptImage ? undefined : \"b64_json\");\n\n    const body: OpenAI.Images.ImageGenerateParamsNonStreaming = {\n      model: this.name,\n      prompt,\n    };\n\n    if (options?.count !== undefined) body.n = options.count;\n    if (options?.size !== undefined) body.size = options.size;\n    if (options?.quality !== undefined) {\n      body.quality = options.quality as OpenAI.Images.ImageGenerateParamsBase[\"quality\"];\n    }\n    if (!isGptImage && responseFormat) body.response_format = responseFormat;\n    if (isGptImage && options?.format !== undefined) {\n      body.output_format = options.format as OpenAI.Images.ImageGenerateParamsBase[\"output_format\"];\n    }\n    if (options?.background !== undefined) {\n      body.background = options.background as OpenAI.Images.ImageGenerateParamsBase[\"background\"];\n    }\n\n    this.logger.debug(LOG_MODULE, \"image.request\", \"images.generate\", {\n      model: this.name,\n      count: options?.count ?? 1,\n    });\n\n    let response: OpenAI.Images.ImagesResponse;\n\n    try {\n      response = await this.client.images.generate(\n        body,\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"image.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const images = (response.data ?? []).map((image) =>\n      this.toGeneratedImage(image, options?.format),\n    );\n\n    const usage = response.usage\n      ? {\n          input: response.usage.input_tokens,\n          output: response.usage.output_tokens,\n          total: response.usage.total_tokens,\n        }\n      : { input: 0, output: 0, total: 0 };\n\n    this.logger.debug(LOG_MODULE, \"image.response\", \"images.generate succeeded\", {\n      images: images.length,\n      usage,\n    });\n\n    return { images, usage };\n  }\n\n  /**\n   * Normalize one OpenAI `Image` into the neutral discriminated shape.\n   * Base64 wins when present (gpt-image, and DALL·E in b64 mode);\n   * otherwise a hosted URL. A response carrying neither is a provider\n   * contract violation — surface it as a typed `ProviderError` rather\n   * than emitting a half-formed part.\n   */\n  private toGeneratedImage(image: OpenAI.Images.Image, format: string | undefined): GeneratedImage {\n    if (image.b64_json) {\n      return {\n        type: \"base64\",\n        base64: image.b64_json,\n        mediaType: mediaTypeFor(format),\n        ...(image.revised_prompt ? { revisedPrompt: image.revised_prompt } : {}),\n      };\n    }\n\n    if (image.url) {\n      return {\n        type: \"url\",\n        url: image.url,\n        ...(image.revised_prompt ? { revisedPrompt: image.revised_prompt } : {}),\n      };\n    }\n\n    throw new ProviderError(\"OpenAI image response contained neither base64 bytes nor a URL.\");\n  }\n}\n","/**\n * Model-name prefixes for OpenAI families that expose internal\n * reasoning / thinking tokens and accept the `reasoning_effort`\n * request parameter on the Chat Completions API.\n *\n * Matched as a prefix so dated variants (`o3-2025-04-16`) and\n * `-mini` / `-pro` suffixes (`o4-mini`, `gpt-5-pro`) are covered\n * without listing every release tag explicitly.\n *\n * Maintenance: append a new prefix when OpenAI ships a reasoning\n * model family that doesn't already match. Devs can always override\n * per-model via `openai.model({ name, reasoning: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst REASONING_CAPABLE_PREFIXES = [\"o1\", \"o3\", \"o4\", \"gpt-5\"];\n\n/**\n * Infer whether a given OpenAI model name is a reasoning model (o-series\n * and the gpt-5 family) based on the known-prefix list. Unknown models\n * default to `false` so the adapter never forwards an unsupported\n * `reasoning_effort` param to a non-reasoning model (which would 400).\n *\n * @example\n * inferReasoningCapability(\"o3-mini\");          // → true\n * inferReasoningCapability(\"o4-mini\");          // → true\n * inferReasoningCapability(\"gpt-5-pro\");        // → true\n * inferReasoningCapability(\"gpt-4o\");           // → false\n * inferReasoningCapability(\"custom-llm\");       // → false\n */\nexport function inferReasoningCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return REASONING_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","/**\n * Model-name prefixes for OpenAI families that support vision input\n * (image attachments) on the Chat Completions API.\n *\n * Matched as a prefix so dated variants (`gpt-4o-2024-08-06`) and\n * `-mini` / `-preview` suffixes (`gpt-4o-mini`, `gpt-4-turbo-preview`)\n * are covered without listing every release tag explicitly.\n *\n * Maintenance: append a new prefix when OpenAI ships a vision-capable\n * model family that doesn't already match. Devs can always override\n * per-model via `openai.model({ name, vision: true | false })` —\n * explicit config wins over inference in either direction.\n */\nconst VISION_CAPABLE_PREFIXES = [\n  \"gpt-4o\",\n  \"gpt-4-turbo\",\n  \"gpt-4.1\",\n  \"o1\",\n  \"o3\",\n  \"chatgpt-4o\",\n];\n\n/**\n * Infer whether a given OpenAI model name supports vision based on the\n * known-prefix list. Unknown models default to `false` so that passing\n * an image attachment to an unsupported model surfaces a clear,\n * agent-side capability error instead of an opaque OpenAI 400.\n *\n * @example\n * inferVisionCapability(\"gpt-4o-mini\");          // → true\n * inferVisionCapability(\"gpt-4o-2024-08-06\");    // → true\n * inferVisionCapability(\"gpt-3.5-turbo\");        // → false\n * inferVisionCapability(\"custom-llm\");           // → false\n */\nexport function inferVisionCapability(modelName: string): boolean {\n  const normalized = modelName.toLowerCase();\n\n  return VISION_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));\n}\n","import {\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, OpenAIResponseFormat } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toOpenAIMessages, toOpenAITools, wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * Map an explicit `responseFormat` override to the default\n * `structuredOutput` capability. Loose wire modes (`\"json_object\"`,\n * `\"text\"`) don't enforce shape, so the agent needs to see the soft\n * schema hint in the system prompt — that only happens when the\n * capability is `false`. Default (no override) stays `true` to\n * preserve the prior assumption that OpenAI models support strict\n * structured output.\n */\nfunction inferStructuredOutput(responseFormat: OpenAIResponseFormat | undefined): boolean {\n  if (responseFormat === \"json_object\" || responseFormat === \"text\") {\n    return false;\n  }\n\n  return true;\n}\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    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 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      // OpenAI prompt caching is automatic on the Chat Completions API\n      // (no caller-supplied breakpoints — the platform caches long\n      // prompt prefixes server-side and reports the hit count via\n      // `prompt_tokens_details.cached_tokens`). We therefore advertise\n      // the read-side accounting capability as always available while\n      // treating `ModelCallOptions.cacheControl` write breakpoints as a\n      // no-op (see `buildReasoningParams` siblings — there is no cache\n      // param to emit).\n      promptCaching: true,\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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[0];\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter.\n   *\n   * When `config.responseFormat` is set, it wins: `\"text\"` emits no\n   * `response_format` at all, `\"json_object\"` always picks the loose\n   * mode, and `\"json_schema\"` picks strict mode (with the same\n   * `isStrictCompatible` safety check — a malformed schema still\n   * degrades to `json_object` rather than 400). The override exists\n   * because some targets (older OpenAI models, OpenRouter routes,\n   * Ollama OpenAI-compat) reject strict `json_schema` outright.\n   *\n   * When the override is omitted, uses strict `json_schema` mode\n   * (token-level enforcement) only when the schema is a proper\n   * root-object JSON Schema (`{ type: \"object\", properties: ... }`).\n   * For anything else — malformed extractor output, non-object\n   * schemas, or future shapes we haven't tested — falls back to loose\n   * `json_object` mode, which guarantees *some* valid JSON without\n   * enforcing shape. The agent's soft instruction already embeds the\n   * schema text in the system prompt when the model declares no\n   * native structured-output capability, so shape validation still\n   * runs client-side via the Standard Schema `validate()` call.\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    if (!responseSchema) {\n      return {};\n    }\n\n    const override = this.config.responseFormat;\n\n    if (override === \"text\") {\n      return {};\n    }\n\n    if (override === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    // Either auto-select (no override) or explicit `\"json_schema\"`.\n    // The strict-compat check still applies in the explicit case —\n    // a malformed / non-object schema would 400 before sampling, so\n    // we degrade to `json_object` rather than crash.\n    if (this.isStrictCompatible(responseSchema)) {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: {\n            name: \"response\",\n            schema: responseSchema,\n            strict: true,\n          },\n        },\n      };\n    }\n\n    return { response_format: { type: \"json_object\" } };\n  }\n\n  /**\n   * OpenAI strict `json_schema` mode requires the root to be a JSON\n   * Schema object type (`{ type: \"object\", properties: ... }`). Anything\n   * else (top-level arrays, primitives, unknown shapes) is rejected with\n   * a 400 before a token is sampled. We check structurally here so the\n   * first call doesn't crash on a malformed extraction — loose\n   * `json_object` mode is a safe degradation.\n   */\n  private isStrictCompatible(schema: Record<string, unknown>): boolean {\n    return (\n      schema.type === \"object\" &&\n      typeof schema.properties === \"object\" &&\n      schema.properties !== null &&\n      this.isStrictSafeNode(schema)\n    );\n  }\n\n  /**\n   * Recursively check the one strict-mode rule schemas most often trip on:\n   * every object must list ALL of its `properties` in `required` (OpenAI\n   * strict has no notion of optional — optional fields must be expressed\n   * as nullable, e.g. `type: [\"string\", \"null\"]`, and still appear in\n   * `required`). A schema that violates this anywhere in the tree is NOT\n   * sent in strict `json_schema` mode — it degrades to loose\n   * `json_object` so a hand-built or optional-bearing schema can't 400\n   * the call (\"'required' ... must include every key in properties\").\n   * Client-side `validate()` still enforces the full shape.\n   */\n  private isStrictSafeNode(node: unknown): boolean {\n    if (!node || typeof node !== \"object\") {\n      return true;\n    }\n\n    const record = node as Record<string, unknown>;\n\n    if (record.type === \"object\" && record.properties && typeof record.properties === \"object\") {\n      const properties = record.properties as Record<string, unknown>;\n      const keys = Object.keys(properties);\n      const required = Array.isArray(record.required) ? (record.required as unknown[]) : [];\n\n      if (keys.some((key) => !required.includes(key))) {\n        return false;\n      }\n\n      for (const key of keys) {\n        if (!this.isStrictSafeNode(properties[key])) {\n          return false;\n        }\n      }\n    }\n\n    if (record.items !== undefined && !this.isStrictSafeNode(record.items)) {\n      return false;\n    }\n\n    for (const branch of [\"anyOf\", \"allOf\", \"oneOf\"] as const) {\n      const value = record[branch];\n      if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) {\n        return false;\n      }\n    }\n\n    return true;\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    const cachedTokens = raw.prompt_tokens_details?.cached_tokens;\n    const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;\n\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n      ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\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 in two cases so the adapter never forwards an unsupported\n   * param: (1) the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`), or (2) the caller supplied no `effort`.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  private buildReasoningParams(reasoning: ModelCallOptions[\"reasoning\"]): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning || !reasoning?.effort) {\n      return {};\n    }\n\n    return { reasoning_effort: 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","import {\n  InvalidRequestError,\n  type SpeechGenerationResponse,\n  type SpeechModelContract,\n  type SpeechModelPricing,\n  type SpeechOptions,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAISpeechConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Model-id prefixes OpenAI exposes through the **Speech** (TTS) API. */\nconst SPEECH_MODEL_PREFIXES = [\"tts-1\", \"gpt-4o-mini-tts\", \"gpt-audio\"] as const;\n\n/** True when `name` is a recognized OpenAI text-to-speech model. */\nexport function isOpenAISpeechModel(name: string): boolean {\n  return SPEECH_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n\n/** Map a neutral output container hint to its IANA audio media type. */\nfunction audioMediaType(format: string | undefined): string {\n  switch (format) {\n    case \"opus\":\n      return \"audio/opus\";\n    case \"aac\":\n      return \"audio/aac\";\n    case \"flac\":\n      return \"audio/flac\";\n    case \"wav\":\n      return \"audio/wav\";\n    case \"pcm\":\n      return \"audio/pcm\";\n    default:\n      return \"audio/mpeg\";\n  }\n}\n\n/**\n * OpenAI-backed implementation of `SpeechModelContract` (text-to-speech)\n * via `audio.speech.create`. Standalone primitive — no relation to chat\n * completions or the agent loop. Consumed by the `ai.speech()` verb.\n *\n * **Capability guard.** The constructor rejects a non-TTS model id up\n * front (`tts-1` / `gpt-4o-mini-tts` only) so the mistake surfaces at\n * wiring time, mirroring the embedder / image guards.\n *\n * @example\n * const tts = new OpenAISpeechModel(client, { name: \"tts-1\", voice: \"alloy\" }, \"openai\");\n * const { audio } = await tts.generate(\"Welcome aboard.\");\n */\nexport class OpenAISpeechModel implements SpeechModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: SpeechModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly defaultVoice?: string;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAISpeechConfig, provider: string = \"openai\") {\n    if (!isOpenAISpeechModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI text-to-speech model. ` +\n          \"Use a `tts-1` / `tts-1-hd` / `gpt-4o-mini-tts` model with openai.speech({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.defaultVoice = config.voice;\n  }\n\n  public async generate(text: string, options?: SpeechOptions): Promise<SpeechGenerationResponse> {\n    const format = options?.format ?? \"mp3\";\n\n    this.logger.debug(LOG_MODULE, \"speech.request\", \"audio.speech.create\", {\n      model: this.name,\n      characters: text.length,\n    });\n\n    let response: Response;\n\n    try {\n      response = await this.client.audio.speech.create(\n        {\n          model: this.name,\n          input: text,\n          voice: options?.voice ?? this.defaultVoice ?? \"alloy\",\n          response_format: format as OpenAI.Audio.SpeechCreateParams[\"response_format\"],\n          ...(options?.speed !== undefined ? { speed: options.speed } : {}),\n          ...(options?.instructions !== undefined ? { instructions: options.instructions } : {}),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n      this.logger.error(LOG_MODULE, \"speech.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n      throw wrapped;\n    }\n\n    const base64 = Buffer.from(await response.arrayBuffer()).toString(\"base64\");\n\n    return {\n      audio: { type: \"base64\", base64, mediaType: audioMediaType(format) },\n      // The Speech API reports no token usage; spend is priced per\n      // character (or per token for gpt-4o-mini-tts) by `ai.speech()`.\n      usage: { input: 0, output: 0, total: 0 },\n      characters: text.length,\n    };\n  }\n}\n","import {\n  InvalidRequestError,\n  type AudioInput,\n  type TranscribeOptions,\n  type TranscriptionModelContract,\n  type TranscriptionModelPricing,\n  type TranscriptionResponse,\n  type TranscriptionSegment,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport OpenAI, { toFile } from \"openai\";\nimport type { OpenAITranscriptionConfig } from \"./config.type\";\nimport { wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/** Model-id prefixes OpenAI exposes through the **Transcription** (STT) API. */\nconst TRANSCRIPTION_MODEL_PREFIXES = [\"whisper\", \"gpt-4o-transcribe\", \"gpt-4o-mini-transcribe\"] as const;\n\n/** True when `name` is a recognized OpenAI speech-to-text model. */\nexport function isOpenAITranscriptionModel(name: string): boolean {\n  return TRANSCRIPTION_MODEL_PREFIXES.some((prefix) => name.startsWith(prefix));\n}\n\n/** Defensive view over the response, whose shape varies by `response_format`. */\ntype RawTranscription = {\n  text: string;\n  duration?: number;\n  language?: string;\n  segments?: Array<{ text: string; start?: number; end?: number }>;\n  usage?: {\n    type?: string;\n    seconds?: number;\n    input_tokens?: number;\n    output_tokens?: number;\n    total_tokens?: number;\n  };\n};\n\n/**\n * OpenAI-backed implementation of `TranscriptionModelContract`\n * (speech-to-text) via `audio.transcriptions.create`. Consumed by the\n * `ai.transcribe()` verb.\n *\n * **Response format.** Defaults to `verbose_json` for `whisper-1` (so\n * the run gets a `duration` + timestamped `segments`) and `json` for\n * the token-metered `gpt-4o-transcribe` family. Base64 audio is wrapped\n * in an uploadable via the SDK's `toFile`.\n *\n * @example\n * const stt = new OpenAITranscriptionModel(client, { name: \"whisper-1\" }, \"openai\");\n * const { text } = await stt.transcribe({ base64, mediaType: \"audio/mpeg\" });\n */\nexport class OpenAITranscriptionModel implements TranscriptionModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: TranscriptionModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly logger: Logger = log;\n\n  public constructor(\n    client: OpenAI,\n    config: OpenAITranscriptionConfig,\n    provider: string = \"openai\",\n  ) {\n    if (!isOpenAITranscriptionModel(config.name)) {\n      throw new InvalidRequestError(\n        `\"${config.name}\" is not a known OpenAI transcription model. ` +\n          \"Use a `whisper-1` / `gpt-4o-transcribe` / `gpt-4o-mini-transcribe` model with openai.transcribe({ name }).\",\n      );\n    }\n\n    this.client = client;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n  }\n\n  public async transcribe(\n    audio: AudioInput,\n    options?: TranscribeOptions,\n  ): Promise<TranscriptionResponse> {\n    const isWhisper = this.name.startsWith(\"whisper\");\n    const format = options?.format ?? (isWhisper ? \"verbose_json\" : \"json\");\n\n    const file = await toFile(Buffer.from(audio.base64, \"base64\"), audio.filename ?? \"audio\", {\n      type: audio.mediaType,\n    });\n\n    this.logger.debug(LOG_MODULE, \"transcription.request\", \"audio.transcriptions.create\", {\n      model: this.name,\n      format,\n    });\n\n    let raw: unknown;\n\n    try {\n      raw = await this.client.audio.transcriptions.create(\n        {\n          model: this.name,\n          file,\n          response_format: format as OpenAI.Audio.TranscriptionCreateParams[\"response_format\"],\n          ...(options?.language ? { language: options.language } : {}),\n          ...(options?.prompt ? { prompt: options.prompt } : {}),\n        } as OpenAI.Audio.TranscriptionCreateParamsNonStreaming,\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n      this.logger.error(LOG_MODULE, \"transcription.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n      throw wrapped;\n    }\n\n    const response = raw as RawTranscription;\n\n    const segments: TranscriptionSegment[] | undefined = response.segments?.map((segment) => ({\n      text: segment.text,\n      ...(segment.start !== undefined ? { start: segment.start } : {}),\n      ...(segment.end !== undefined ? { end: segment.end } : {}),\n    }));\n\n    const durationSeconds =\n      response.duration ?? (response.usage?.type === \"duration\" ? response.usage.seconds : undefined);\n\n    const usage =\n      response.usage?.type === \"tokens\"\n        ? {\n            input: response.usage.input_tokens ?? 0,\n            output: response.usage.output_tokens ?? 0,\n            total: response.usage.total_tokens ?? 0,\n          }\n        : { input: 0, output: 0, total: 0 };\n\n    return {\n      text: response.text,\n      ...(segments && segments.length > 0 ? { segments } : {}),\n      ...(durationSeconds !== undefined ? { durationSeconds } : {}),\n      usage,\n    };\n  }\n}\n","import OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n  SpeechModelContract,\n  TranscriptionModelContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  OpenAIEmbedderConfig,\n  OpenAIImageConfig,\n  OpenAIModelConfig,\n  OpenAISDKConfig,\n  OpenAISpeechConfig,\n  OpenAITranscriptionConfig,\n} from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIImageModel } from \"./image\";\nimport { OpenAIModel } from \"./model\";\nimport { OpenAISpeechModel } from \"./speech\";\nimport { OpenAITranscriptionModel } from \"./transcription\";\n\n/**\n * OpenAI-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for any OpenAI-compatible provider\n * (OpenAI, Azure OpenAI, OpenRouter, local gateways speaking the Chat\n * Completions protocol). A single `OpenAISDK` instance holds one live\n * `OpenAI` client, shared by every `ModelContract` it produces via\n * `model()`. Users construct one SDK per provider/account and reuse it\n * across all agents, workflows, and supervisors that target that\n * provider.\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client (authentication, base URL) and\n *   its lifetime scope. Factory for `OpenAIModel` instances — each\n *   model call gets a reference to the same client.\n * - Does NOT own: anything per-call (tool execution, message history,\n *   streaming loop) — those live in `OpenAIModel` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"): the `OpenAI` client is heavy to construct and\n * designed to be reused; keeping it on `this` makes that reuse\n * explicit and aligns with the PascalCase naming convention readers\n * expect from a constructor.\n *\n * @example\n * const openai = new OpenAISDK({ apiKey: process.env.OPENAI_API_KEY! });\n * const model = openai.model({ name: \"gpt-4o\", temperature: 0.7 });\n * const tokens = await openai.count(\"Hello world\");\n *\n * @example\n * // Compose into an `ai.openai` namespace for ergonomic agent wiring\n * const ai = { agent, tool, systemPrompt, persona, instruction, openai: new OpenAISDK({ apiKey }) };\n * const myAgent = ai.agent({ model: ai.openai.model({ name: \"gpt-4o-mini\" }) });\n */\nexport class OpenAISDK implements SDKAdapterContract {\n  private readonly client: OpenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: OpenAISDKConfig) {\n    // Peel off the framework-only keys and forward every other upstream\n    // `ClientOptions` (timeout, maxRetries, defaultHeaders, fetch,\n    // organization, project, …) verbatim — they type-check, so dropping them\n    // is a silent footgun. Mirrors the Bedrock/Google/Ollama adapters.\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.client = new OpenAI(clientOptions);\n    this.provider = provider ?? \"openai\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build an `OpenAIModel` bound to this SDK's client. Each call returns\n   * a fresh model instance, but all instances share the underlying\n   * `OpenAI` client — connection pools, rate limits, and authentication\n   * state stay unified across every model produced here. The SDK's\n   * `provider` label is forwarded so every model self-identifies as\n   * coming from the same upstream.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: OpenAIModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate for a given text. Uses a\n   * character-heuristic (`approximateTokenCount`) from the core package\n   * — good enough for budgeting and quota guards, not for billing.\n   * Accepts an optional model id for future per-model tokenizer\n   * dispatch; currently ignored.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build an `OpenAIEmbedder` bound to this SDK's client. Each call\n   * returns a fresh embedder instance sharing the same underlying\n   * `OpenAI` client — connection pools and authentication stay unified\n   * across every embedder produced here.\n   *\n   * @example\n   * const embedder = openai.embedder({ name: \"text-embedding-3-small\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: OpenAIEmbedderConfig): EmbedderContract {\n    return new OpenAIEmbedder(this.client, config);\n  }\n\n  /**\n   * Build an `OpenAIImageModel` bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. Accepts the `gpt-image-*` (token-metered)\n   * and `dall-e-*` (per-image-metered) families; a non-image model id\n   * is rejected at construction.\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined` (no cost computed). A token-priced\n   * `gpt-image-1` entry can live in the same SDK registry as the chat\n   * models.\n   *\n   * @example\n   * const model = openai.image({ name: \"gpt-image-1\" });\n   * const { data } = await ai.image({ model, prompt: \"a red bicycle\" });\n   */\n  public image(config: OpenAIImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: OpenAIImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new OpenAIImageModel(this.client, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Build an `OpenAISpeechModel` (text-to-speech) bound to this SDK's\n   * client, for use with `ai.speech({ model, text })`. Accepts the\n   * `tts-1` / `gpt-4o-mini-tts` families; a non-TTS model id is rejected\n   * at construction.\n   *\n   * @example\n   * const tts = openai.speech({ name: \"tts-1\", voice: \"alloy\" });\n   * const { data } = await ai.speech({ model: tts, text: \"Hello\" });\n   */\n  public speech(config: OpenAISpeechConfig): SpeechModelContract {\n    return new OpenAISpeechModel(this.client, config, this.provider);\n  }\n\n  /**\n   * Build an `OpenAITranscriptionModel` (speech-to-text) bound to this\n   * SDK's client, for use with `ai.transcribe({ model, audio })`.\n   * Accepts the `whisper-1` / `gpt-4o-transcribe` families; a non-STT\n   * model id is rejected at construction.\n   *\n   * @example\n   * const stt = openai.transcribe({ name: \"whisper-1\" });\n   * const { data } = await ai.transcribe({ model: stt, audio });\n   */\n  public transcribe(config: OpenAITranscriptionConfig): TranscriptionModelContract {\n    return new OpenAITranscriptionModel(this.client, config, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAEA,MAAM,kBAAgD;CACpD,MAAM;CACN,YAAY;CACZ,QAAQ;AACV;;;;;;;;;;AAWA,SAAgB,gBAAgB,KAA8C;CAC5E,OAAO,gBAAgB,OAAO,OAAO;AACvC;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACQA,SAAgB,iBACd,UACsD;CACtD,OAAO,SAAS,KAAK,MAAM;EACzB,IAAI,EAAE,SAAS,QACb,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,cAAc,EAAE,cAAc;EAChC;EAEF,IAAI,EAAE,SAAS,eAAe,EAAE,aAAa,EAAE,UAAU,SAAS,GAChE,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,YAAY,EAAE,UAAU,KAAK,QAAQ;IACnC,IAAI,GAAG;IACP,MAAM;IACN,UAAU;KAAE,MAAM,GAAG;KAAM,WAAW,KAAK,UAAU,GAAG,SAAS,CAAC,CAAC;IAAE;GACvE,EAAE;EACJ;EAGF,IAAI,EAAE,SAAS,UAAU,MAAM,QAAQ,EAAE,OAAO,GAC9C,OAAO;GACL,MAAM;GACN,SAAS,EAAE,QAAQ,IAAI,mBAAmB;EAC5C;EAGF,OAAO;GAAE,MAAM,EAAE;GAAM,SAAS,iBAAiB,EAAE,OAAO;EAAE;CAI9D,CAAC;AACH;;;;;;;AAQA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,KAAK,SAAS,SAMhB,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;CAGjD,IAAI,KAAK,SAAS,OAAO;EACvB,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,8FACF;EAGF,OAAO;GACL,MAAM;GACN,MAAM;IACJ,UAAU;IACV,WAAW,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO;GACjE;EACF;CACF;CAGA,IAAI,SAAS,KAAK,QAChB,MAAM,IAAIA,mCACR,2FACF;CAGF,OAAO;EACL,MAAM;EACN,aAAa;GACX,MAAM,KAAK,OAAO;GAClB,QAAQ,oBAAoB,KAAK,OAAO,SAAS;EACnD;CACF;AACF;;;;;;AAOA,SAAS,oBAAoB,WAAkC;CAC7D,IAAI,cAAc,eAAe,cAAc,iBAAiB,cAAc,cAC5E,OAAO;CAGT,IAAI,cAAc,eAAe,cAAc,gBAAgB,cAAc,cAC3E,OAAO;CAGT,MAAM,IAAIA,mCACR,0DAA0D,UAAU,GACtE;AACF;;;;;;;;;;;;;ACvJA,SAAgB,cACd,OAC0D;CAC1D,IAAI,CAAC,SAAS,MAAM,WAAW,GAC7B;CAGF,OAAO,MAAM,KAAK,UAAU;EAC1B,MAAM;EACN,UAAU;GACR,MAAM,KAAK;GACX,aAAa,KAAK;GAClB,YAAY,aAAa,KAAK,KAAK;EACrC;CACF,EAAE;AACJ;;;;;;;AAQA,SAAS,aAAa,OAAuE;CAC3F,MAAM,+CAA2B,KAAK;CAEtC,IAAI,UAAU,OAAO,SAAS,UAC5B,OAAO;CAGT,OAAO;EAAE,MAAM;EAAU,YAAY,CAAC;CAAE;AAC1C;;;;;;;;;;;;;;;;;;;;;;;;;ACOA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkBC,wBACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,QAAQ,KAAK;CAC1C,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAE1F,IAAI,UAAU,QAAQ,KAAK,GACzB,OAAO,IAAIC,oCAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,mBACzC,OAAO,IAAIC,iCAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,sBACjB,OAAO,IAAIC,kCAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,uBACzC,OAAO,IAAIC,sCAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BACjB,OAAO,IAAIC,0CAA2B,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAG3E,IAAI,MAAM,SAAS,kBACjB,OAAO,IAAIC,kCAAmB,SAAS;EACrC,OAAO;EACP;EACA,QAAQ;CACV,CAAC;CAGH,IAAI,OAAO,MAAM,WAAW,YAAY,MAAM,UAAU,OAAO,MAAM,SAAS,KAC5E,OAAO,IAAIC,mCAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAIC,6BAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;;AAQA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkBC,eAAO,UAC3B,OAAO;EACL,QAAQ,OAAO;EACf,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,MAAM,OAAO;CACf;CAGF,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,SACE,OAAO,IAAI,YAAY,YAAY,IAAI,YAAY,OAC9C,IAAI,UACL;GACN,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAClD;CACF;CAEA,OAAO,CAAC;AACV;;;;;;;AAQA,SAAS,UAAU,QAAiB,OAAkC;CACpE,IAAI,kBAAkBA,eAAO,2BAC3B,OAAO;CAGT,IAAI,MAAM,SAAS,6BACjB,OAAO;CAGT,IAAI,MAAM,SAAS,eAAe,MAAM,SAAS,gBAC/C,OAAO;CAGT,OAAO;AACT;;;;;;;AAQA,SAAS,aACP,QACA,OACyB;CACzB,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,MAAM,YAAY,cAAc,MAAM;CAEtC,IAAI,WACF,QAAQ,YAAY;CAGtB,OAAO;AACT;;;;;AAMA,SAAS,cAAc,QAAqC;CAC1D,IAAI,OAAO,WAAW,YAAY,WAAW,MAC3C;CAGF,MAAM,MAAM;CAEZ,IAAI,OAAO,IAAI,eAAe,UAC5B,OAAO,IAAI;CAGb,IAAI,OAAO,IAAI,cAAc,UAC3B,OAAO,IAAI;AAIf;;;;;;AAOA,SAAS,gBAAgB,SAAiE;CACxF,IAAI,CAAC,SACH;CAGF,MAAM,MAAM,QAAQ,kBAAkB,QAAQ;CAE9C,IAAI,CAAC,KACH;CAGF,MAAM,UAAU,OAAO,GAAG;CAE1B,IAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GACzC;CAGF,OAAO,KAAK,MAAM,UAAU,GAAI;AAClC;;;;AChOA,MAAMC,eAAa;;;;;;;;;;;;;;;;;;;;;;;;;;AA+BnB,IAAa,iBAAb,MAAwD;CAetD,AAAO,YAAY,QAAgB,QAA8B;kBAbtC;gBAWOC;EAGhC,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,uBAAuB,OAAO;EACnC,KAAK,aAAa,OAAO,cAAc;CACzC;CAEA,MAAa,MAAM,OAAyC;EAC1D,MAAM,EAAE,UAAU,UAAU,MAAM,KAAK,QAAQ,KAAK;EAEpD,OAAO;GACL,QAAQ,SAAS,KAAK,EAAE,CAAC;GACzB,YAAY,KAAK;GACjB;EACF;CACF;CAEA,MAAa,UAAU,QAAiD;EACtE,MAAM,EAAE,UAAU,UAAU,MAAM,KAAK,QAAQ,MAAM;EAErD,OAAO;GACL,SAAS,SAAS,KAAK,KAAK,MAAM,EAAE,SAAS;GAC7C,YAAY,KAAK;GACjB;EACF;CACF;;;;;;;CAQA,MAAc,QAAQ,OAGnB;EACD,KAAK,OAAO,MAAMD,cAAY,oBAAoB,qBAAqB;GACrE,OAAO,KAAK;GACZ,OAAO,MAAM,QAAQ,KAAK;GAC1B,OAAO,MAAM,QAAQ,KAAK,IAAI,MAAM,SAAS;EAC/C,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,WAAW,OAAO;IAC7C,OAAO,KAAK;IACZ;IACA,GAAI,KAAK,yBAAyB,SAC9B,EAAE,YAAY,KAAK,qBAAqB,IACxC,CAAC;GACP,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,kBAAkB,QAAQ,SAAS;IAC/D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,KAAK,OAAO,MAAMA,cAAY,qBAAqB,8BAA8B;GAC/E,YAAY,SAAS,KAAK,EAAE,EAAE,UAAU;GACxC,OAAO;IACL,cAAc,SAAS,MAAM;IAC7B,aAAa,SAAS,MAAM;GAC9B;EACF,CAAC;EAID,IAAI,KAAK,eAAe,GACtB,KAAK,aAAa,SAAS,KAAK,EAAE,CAAC,UAAU;EAG/C,MAAM,QAAwB;GAC5B,cAAc,SAAS,MAAM;GAC7B,aAAa,SAAS,MAAM;EAC9B;EAEA,OAAO;GAAE;GAAU;EAAM;CAC3B;AACF;;;;;;;;;;;;;;;;;ACjIA,MAAa,8BAA8B,CAAC,aAAa,QAAQ;;;;;;;;;;;;AAajE,SAAgB,mBAAmB,MAAuB;CACxD,OAAO,4BAA4B,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AAC7E;;;;ACbA,MAAME,eAAa;;AAGnB,SAAS,aAAa,QAAoC;CACxD,QAAQ,QAAR;EACE,KAAK;EACL,KAAK,OACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,SACE,OAAO;CACX;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;AA4BA,IAAa,mBAAb,MAA4D;CAQ1D,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzDC;EAGhC,IAAI,CAAC,mBAAmB,OAAO,IAAI,GACjC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,yHAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;CACxB;CAEA,MAAa,SACX,QACA,SACkC;EAClC,MAAM,aAAa,KAAK,KAAK,WAAW,WAAW;EAInD,MAAM,iBACH,SAAS,mBACT,aAAa,SAAY;EAE5B,MAAM,OAAsD;GAC1D,OAAO,KAAK;GACZ;EACF;EAEA,IAAI,SAAS,UAAU,QAAW,KAAK,IAAI,QAAQ;EACnD,IAAI,SAAS,SAAS,QAAW,KAAK,OAAO,QAAQ;EACrD,IAAI,SAAS,YAAY,QACvB,KAAK,UAAU,QAAQ;EAEzB,IAAI,CAAC,cAAc,gBAAgB,KAAK,kBAAkB;EAC1D,IAAI,cAAc,SAAS,WAAW,QACpC,KAAK,gBAAgB,QAAQ;EAE/B,IAAI,SAAS,eAAe,QAC1B,KAAK,aAAa,QAAQ;EAG5B,KAAK,OAAO,MAAMF,cAAY,iBAAiB,mBAAmB;GAChE,OAAO,KAAK;GACZ,OAAO,SAAS,SAAS;EAC3B,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,OAAO,SAClC,MACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,eAAe,QAAQ,SAAS;IAC5D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,UAAU,SAAS,QAAQ,CAAC,EAAC,CAAE,KAAK,UACxC,KAAK,iBAAiB,OAAO,SAAS,MAAM,CAC9C;EAEA,MAAM,QAAQ,SAAS,QACnB;GACE,OAAO,SAAS,MAAM;GACtB,QAAQ,SAAS,MAAM;GACvB,OAAO,SAAS,MAAM;EACxB,IACA;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAEpC,KAAK,OAAO,MAAMA,cAAY,kBAAkB,6BAA6B;GAC3E,QAAQ,OAAO;GACf;EACF,CAAC;EAED,OAAO;GAAE;GAAQ;EAAM;CACzB;;;;;;;;CASA,AAAQ,iBAAiB,OAA4B,QAA4C;EAC/F,IAAI,MAAM,UACR,OAAO;GACL,MAAM;GACN,QAAQ,MAAM;GACd,WAAW,aAAa,MAAM;GAC9B,GAAI,MAAM,iBAAiB,EAAE,eAAe,MAAM,eAAe,IAAI,CAAC;EACxE;EAGF,IAAI,MAAM,KACR,OAAO;GACL,MAAM;GACN,KAAK,MAAM;GACX,GAAI,MAAM,iBAAiB,EAAE,eAAe,MAAM,eAAe,IAAI,CAAC;EACxE;EAGF,MAAM,IAAIG,6BAAc,iEAAiE;CAC3F;AACF;;;;;;;;;;;;;;;;;;ACpKA,MAAM,6BAA6B;CAAC;CAAM;CAAM;CAAM;AAAO;;;;;;;;;;;;;;AAe7D,SAAgB,yBAAyB,WAA4B;CACnE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,2BAA2B,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAClF;;;;;;;;;;;;;;;;;ACpBA,MAAM,0BAA0B;CAC9B;CACA;CACA;CACA;CACA;CACA;AACF;;;;;;;;;;;;;AAcA,SAAgB,sBAAsB,WAA4B;CAChE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,wBAAwB,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAC/E;;;;ACnBA,MAAMC,eAAa;;;;;;;;;;AAWnB,SAAS,sBAAsB,gBAA2D;CACxF,IAAI,mBAAmB,iBAAiB,mBAAmB,QACzD,OAAO;CAGT,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzDC;EAGhC,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;GASnE,eAAe;GAMf,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,KAAK,OAAO,MAAMD,cAAY,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,SAAS,SAAS,QAAQ;EAChC,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAMA,cAAY,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,MAAMA,cAAY,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,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,yCAA8C,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAMA,cAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAMA,cAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4BA,AAAQ,oBAAoB,gBAE1B;EACA,IAAI,CAAC,gBACH,OAAO,CAAC;EAGV,MAAM,WAAW,KAAK,OAAO;EAE7B,IAAI,aAAa,QACf,OAAO,CAAC;EAGV,IAAI,aAAa,eACf,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAOpD,IAAI,KAAK,mBAAmB,cAAc,GACxC,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IACX,MAAM;IACN,QAAQ;IACR,QAAQ;GACV;EACF,EACF;EAGF,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;CACpD;;;;;;;;;CAUA,AAAQ,mBAAmB,QAA0C;EACnE,OACE,OAAO,SAAS,YAChB,OAAO,OAAO,eAAe,YAC7B,OAAO,eAAe,QACtB,KAAK,iBAAiB,MAAM;CAEhC;;;;;;;;;;;;CAaA,AAAQ,iBAAiB,MAAwB;EAC/C,IAAI,CAAC,QAAQ,OAAO,SAAS,UAC3B,OAAO;EAGT,MAAM,SAAS;EAEf,IAAI,OAAO,SAAS,YAAY,OAAO,cAAc,OAAO,OAAO,eAAe,UAAU;GAC1F,MAAM,aAAa,OAAO;GAC1B,MAAM,OAAO,OAAO,KAAK,UAAU;GACnC,MAAM,WAAW,MAAM,QAAQ,OAAO,QAAQ,IAAK,OAAO,WAAyB,CAAC;GAEpF,IAAI,KAAK,MAAM,QAAQ,CAAC,SAAS,SAAS,GAAG,CAAC,GAC5C,OAAO;GAGT,KAAK,MAAM,OAAO,MAChB,IAAI,CAAC,KAAK,iBAAiB,WAAW,IAAI,GACxC,OAAO;EAGb;EAEA,IAAI,OAAO,UAAU,UAAa,CAAC,KAAK,iBAAiB,OAAO,KAAK,GACnE,OAAO;EAGT,KAAK,MAAM,UAAU;GAAC;GAAS;GAAS;EAAO,GAAY;GACzD,MAAM,QAAQ,OAAO;GACrB,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,MAAM,QAAQ,CAAC,KAAK,iBAAiB,GAAG,CAAC,GACzE,OAAO;EAEX;EAEA,OAAO;CACT;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAChD,MAAM,kBAAkB,IAAI,2BAA2B;EAEvD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;GACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EACpF;CACF;;;;;;;;;;;;;;;;;;;;;;;;;;CA2BA,AAAQ,qBAAqB,WAE3B;EACA,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WAAW,QAC9C,OAAO,CAAC;EAGV,OAAO,EAAE,kBAAkB,UAAU,OAAO;CAC9C;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,yCAA+C,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF;;;;AC3fA,MAAME,eAAa;;AAGnB,MAAM,wBAAwB;CAAC;CAAS;CAAmB;AAAW;;AAGtE,SAAgB,oBAAoB,MAAuB;CACzD,OAAO,sBAAsB,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AACvE;;AAGA,SAAS,eAAe,QAAoC;CAC1D,QAAQ,QAAR;EACE,KAAK,QACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,KAAK,OACH,OAAO;EACT,SACE,OAAO;CACX;AACF;;;;;;;;;;;;;;AAeA,IAAa,oBAAb,MAA8D;CAS5D,AAAO,YAAY,QAAgB,QAA4B,WAAmB,UAAU;gBAF1DC;EAGhC,IAAI,CAAC,oBAAoB,OAAO,IAAI,GAClC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,uIAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe,OAAO;CAC7B;CAEA,MAAa,SAAS,MAAc,SAA4D;EAC9F,MAAM,SAAS,SAAS,UAAU;EAElC,KAAK,OAAO,MAAMF,cAAY,kBAAkB,uBAAuB;GACrE,OAAO,KAAK;GACZ,YAAY,KAAK;EACnB,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,MAAM,OAAO,OACxC;IACE,OAAO,KAAK;IACZ,OAAO;IACP,OAAO,SAAS,SAAS,KAAK,gBAAgB;IAC9C,iBAAiB;IACjB,GAAI,SAAS,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;IAC/D,GAAI,SAAS,iBAAiB,SAAY,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;GACtF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GACtC,KAAK,OAAO,MAAMA,cAAY,gBAAgB,QAAQ,SAAS;IAC7D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GACD,MAAM;EACR;EAIA,OAAO;GACL,OAAO;IAAE,MAAM;IAAU,QAHZ,OAAO,KAAK,MAAM,SAAS,YAAY,CAAC,CAAC,CAAC,SAAS,QAGlC;IAAG,WAAW,eAAe,MAAM;GAAE;GAGnE,OAAO;IAAE,OAAO;IAAG,QAAQ;IAAG,OAAO;GAAE;GACvC,YAAY,KAAK;EACnB;CACF;AACF;;;;ACxGA,MAAM,aAAa;;AAGnB,MAAM,+BAA+B;CAAC;CAAW;CAAqB;AAAwB;;AAG9F,SAAgB,2BAA2B,MAAuB;CAChE,OAAO,6BAA6B,MAAM,WAAW,KAAK,WAAW,MAAM,CAAC;AAC9E;;;;;;;;;;;;;;;AA+BA,IAAa,2BAAb,MAA4E;CAQ1E,AAAO,YACL,QACA,QACA,WAAmB,UACnB;gBANgCG;EAOhC,IAAI,CAAC,2BAA2B,OAAO,IAAI,GACzC,MAAM,IAAIC,mCACR,IAAI,OAAO,KAAK,8JAElB;EAGF,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;CACxB;CAEA,MAAa,WACX,OACA,SACgC;EAChC,MAAM,YAAY,KAAK,KAAK,WAAW,SAAS;EAChD,MAAM,SAAS,SAAS,WAAW,YAAY,iBAAiB;EAEhE,MAAM,OAAO,yBAAa,OAAO,KAAK,MAAM,QAAQ,QAAQ,GAAG,MAAM,YAAY,SAAS,EACxF,MAAM,MAAM,UACd,CAAC;EAED,KAAK,OAAO,MAAM,YAAY,yBAAyB,+BAA+B;GACpF,OAAO,KAAK;GACZ;EACF,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,MAAM,MAAM,KAAK,OAAO,MAAM,eAAe,OAC3C;IACE,OAAO,KAAK;IACZ;IACA,iBAAiB;IACjB,GAAI,SAAS,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;IAC1D,GAAI,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;GACtD,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GACtC,KAAK,OAAO,MAAM,YAAY,uBAAuB,QAAQ,SAAS;IACpE,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GACD,MAAM;EACR;EAEA,MAAM,WAAW;EAEjB,MAAM,WAA+C,SAAS,UAAU,KAAK,aAAa;GACxF,MAAM,QAAQ;GACd,GAAI,QAAQ,UAAU,SAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;GAC9D,GAAI,QAAQ,QAAQ,SAAY,EAAE,KAAK,QAAQ,IAAI,IAAI,CAAC;EAC1D,EAAE;EAEF,MAAM,kBACJ,SAAS,aAAa,SAAS,OAAO,SAAS,aAAa,SAAS,MAAM,UAAU;EAEvF,MAAM,QACJ,SAAS,OAAO,SAAS,WACrB;GACE,OAAO,SAAS,MAAM,gBAAgB;GACtC,QAAQ,SAAS,MAAM,iBAAiB;GACxC,OAAO,SAAS,MAAM,gBAAgB;EACxC,IACA;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAEtC,OAAO;GACL,MAAM,SAAS;GACf,GAAI,YAAY,SAAS,SAAS,IAAI,EAAE,SAAS,IAAI,CAAC;GACtD,GAAI,oBAAoB,SAAY,EAAE,gBAAgB,IAAI,CAAC;GAC3D;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACrFA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAK1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,SAAS,IAAIC,eAAO,aAAa;EACtC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;;;CAcA,AAAO,MAAM,QAA0C;EACrD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,YAAY,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACnE;;;;;;;;CASA,MAAa,MAAM,MAAc,QAAkC;EACjE,iDAA6B,IAAI;CACnC;;;;;;;;;;;CAYA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,QAAQ,MAAM;CAC/C;;;;;;;;;;;;;;;;;CAkBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,iBAAiB,KAAK,QAAQ,gBAAgB,KAAK,QAAQ;CACxE;;;;;;;;;;;CAYA,AAAO,OAAO,QAAiD;EAC7D,OAAO,IAAI,kBAAkB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACjE;;;;;;;;;;;CAYA,AAAO,WAAW,QAA+D;EAC/E,OAAO,IAAI,yBAAyB,KAAK,QAAQ,QAAQ,KAAK,QAAQ;CACxE;AACF"}

package/esm/model.mjs

@@ -344,12 +344,20 @@ var OpenAIModel = class {
 	* so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no
 	* wire equivalent here and is silently ignored.
 	*
+	* The neutral `ReasoningEffort` (`"low" | "medium" | "high" | "none"`)
+	* is a subset of OpenAI's accepted values, so it forwards verbatim —
+	* `"none"` included. `"none"` is load-bearing: gpt-5 / o-series models
+	* **reject function tools** on Chat Completions while reasoning is
+	* active, and the endpoint accepts tools only when `reasoning_effort`
+	* is `"none"` (the alternative is the Responses API). This is why the
+	* param is EMITTED for `"none"` rather than omitted — omitting it
+	* leaves the model reasoning server-side by default, so tools would
+	* still be rejected.
+	*
 	* No-ops in two cases so the adapter never forwards an unsupported
 	* param: (1) the model is not reasoning-capable
-	* (`capabilities.reasoning` is false — e.g. `gpt-4o`), or (2) the caller
-	* supplied no `effort`. The neutral `ReasoningEffort`
-	* (`"low" | "medium" | "high"`) is a strict subset of OpenAI's accepted
-	* values, so it forwards verbatim.
+	* (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any
+	* `reasoning_effort`), or (2) the caller supplied no `effort`.
 	*
 	* Returns an empty spread when nothing applies, so the caller can
 	* unconditionally `...buildReasoningParams(...)` into the request.

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-openai/src/model.ts"],"sourcesContent":["import {\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, OpenAIResponseFormat } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toOpenAIMessages, toOpenAITools, wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * Map an explicit `responseFormat` override to the default\n * `structuredOutput` capability. Loose wire modes (`\"json_object\"`,\n * `\"text\"`) don't enforce shape, so the agent needs to see the soft\n * schema hint in the system prompt — that only happens when the\n * capability is `false`. Default (no override) stays `true` to\n * preserve the prior assumption that OpenAI models support strict\n * structured output.\n */\nfunction inferStructuredOutput(responseFormat: OpenAIResponseFormat | undefined): boolean {\n  if (responseFormat === \"json_object\" || responseFormat === \"text\") {\n    return false;\n  }\n\n  return true;\n}\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    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 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      // OpenAI prompt caching is automatic on the Chat Completions API\n      // (no caller-supplied breakpoints — the platform caches long\n      // prompt prefixes server-side and reports the hit count via\n      // `prompt_tokens_details.cached_tokens`). We therefore advertise\n      // the read-side accounting capability as always available while\n      // treating `ModelCallOptions.cacheControl` write breakpoints as a\n      // no-op (see `buildReasoningParams` siblings — there is no cache\n      // param to emit).\n      promptCaching: true,\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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[0];\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter.\n   *\n   * When `config.responseFormat` is set, it wins: `\"text\"` emits no\n   * `response_format` at all, `\"json_object\"` always picks the loose\n   * mode, and `\"json_schema\"` picks strict mode (with the same\n   * `isStrictCompatible` safety check — a malformed schema still\n   * degrades to `json_object` rather than 400). The override exists\n   * because some targets (older OpenAI models, OpenRouter routes,\n   * Ollama OpenAI-compat) reject strict `json_schema` outright.\n   *\n   * When the override is omitted, uses strict `json_schema` mode\n   * (token-level enforcement) only when the schema is a proper\n   * root-object JSON Schema (`{ type: \"object\", properties: ... }`).\n   * For anything else — malformed extractor output, non-object\n   * schemas, or future shapes we haven't tested — falls back to loose\n   * `json_object` mode, which guarantees *some* valid JSON without\n   * enforcing shape. The agent's soft instruction already embeds the\n   * schema text in the system prompt when the model declares no\n   * native structured-output capability, so shape validation still\n   * runs client-side via the Standard Schema `validate()` call.\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    if (!responseSchema) {\n      return {};\n    }\n\n    const override = this.config.responseFormat;\n\n    if (override === \"text\") {\n      return {};\n    }\n\n    if (override === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    // Either auto-select (no override) or explicit `\"json_schema\"`.\n    // The strict-compat check still applies in the explicit case —\n    // a malformed / non-object schema would 400 before sampling, so\n    // we degrade to `json_object` rather than crash.\n    if (this.isStrictCompatible(responseSchema)) {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: {\n            name: \"response\",\n            schema: responseSchema,\n            strict: true,\n          },\n        },\n      };\n    }\n\n    return { response_format: { type: \"json_object\" } };\n  }\n\n  /**\n   * OpenAI strict `json_schema` mode requires the root to be a JSON\n   * Schema object type (`{ type: \"object\", properties: ... }`). Anything\n   * else (top-level arrays, primitives, unknown shapes) is rejected with\n   * a 400 before a token is sampled. We check structurally here so the\n   * first call doesn't crash on a malformed extraction — loose\n   * `json_object` mode is a safe degradation.\n   */\n  private isStrictCompatible(schema: Record<string, unknown>): boolean {\n    return (\n      schema.type === \"object\" &&\n      typeof schema.properties === \"object\" &&\n      schema.properties !== null &&\n      this.isStrictSafeNode(schema)\n    );\n  }\n\n  /**\n   * Recursively check the one strict-mode rule schemas most often trip on:\n   * every object must list ALL of its `properties` in `required` (OpenAI\n   * strict has no notion of optional — optional fields must be expressed\n   * as nullable, e.g. `type: [\"string\", \"null\"]`, and still appear in\n   * `required`). A schema that violates this anywhere in the tree is NOT\n   * sent in strict `json_schema` mode — it degrades to loose\n   * `json_object` so a hand-built or optional-bearing schema can't 400\n   * the call (\"'required' ... must include every key in properties\").\n   * Client-side `validate()` still enforces the full shape.\n   */\n  private isStrictSafeNode(node: unknown): boolean {\n    if (!node || typeof node !== \"object\") {\n      return true;\n    }\n\n    const record = node as Record<string, unknown>;\n\n    if (record.type === \"object\" && record.properties && typeof record.properties === \"object\") {\n      const properties = record.properties as Record<string, unknown>;\n      const keys = Object.keys(properties);\n      const required = Array.isArray(record.required) ? (record.required as unknown[]) : [];\n\n      if (keys.some((key) => !required.includes(key))) {\n        return false;\n      }\n\n      for (const key of keys) {\n        if (!this.isStrictSafeNode(properties[key])) {\n          return false;\n        }\n      }\n    }\n\n    if (record.items !== undefined && !this.isStrictSafeNode(record.items)) {\n      return false;\n    }\n\n    for (const branch of [\"anyOf\", \"allOf\", \"oneOf\"] as const) {\n      const value = record[branch];\n      if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) {\n        return false;\n      }\n    }\n\n    return true;\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    const cachedTokens = raw.prompt_tokens_details?.cached_tokens;\n    const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;\n\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n      ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\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   * No-ops in two cases so the adapter never forwards an unsupported\n   * param: (1) the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`), or (2) the caller\n   * supplied no `effort`. The neutral `ReasoningEffort`\n   * (`\"low\" | \"medium\" | \"high\"`) is a strict subset of OpenAI's accepted\n   * values, so it forwards verbatim.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  private buildReasoningParams(reasoning: ModelCallOptions[\"reasoning\"]): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning || !reasoning?.effort) {\n      return {};\n    }\n\n    return { reasoning_effort: 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":";;;;;;;;;;;AAmBA,MAAM,aAAa;;;;;;;;;;AAWnB,SAAS,sBAAsB,gBAA2D;CACxF,IAAI,mBAAmB,iBAAiB,mBAAmB,QACzD,OAAO;CAGT,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAGhC,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;GASnE,eAAe;GAMf,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,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,SAAS,SAAS,QAAQ;EAChC,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,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;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4BA,AAAQ,oBAAoB,gBAE1B;EACA,IAAI,CAAC,gBACH,OAAO,CAAC;EAGV,MAAM,WAAW,KAAK,OAAO;EAE7B,IAAI,aAAa,QACf,OAAO,CAAC;EAGV,IAAI,aAAa,eACf,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAOpD,IAAI,KAAK,mBAAmB,cAAc,GACxC,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IACX,MAAM;IACN,QAAQ;IACR,QAAQ;GACV;EACF,EACF;EAGF,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;CACpD;;;;;;;;;CAUA,AAAQ,mBAAmB,QAA0C;EACnE,OACE,OAAO,SAAS,YAChB,OAAO,OAAO,eAAe,YAC7B,OAAO,eAAe,QACtB,KAAK,iBAAiB,MAAM;CAEhC;;;;;;;;;;;;CAaA,AAAQ,iBAAiB,MAAwB;EAC/C,IAAI,CAAC,QAAQ,OAAO,SAAS,UAC3B,OAAO;EAGT,MAAM,SAAS;EAEf,IAAI,OAAO,SAAS,YAAY,OAAO,cAAc,OAAO,OAAO,eAAe,UAAU;GAC1F,MAAM,aAAa,OAAO;GAC1B,MAAM,OAAO,OAAO,KAAK,UAAU;GACnC,MAAM,WAAW,MAAM,QAAQ,OAAO,QAAQ,IAAK,OAAO,WAAyB,CAAC;GAEpF,IAAI,KAAK,MAAM,QAAQ,CAAC,SAAS,SAAS,GAAG,CAAC,GAC5C,OAAO;GAGT,KAAK,MAAM,OAAO,MAChB,IAAI,CAAC,KAAK,iBAAiB,WAAW,IAAI,GACxC,OAAO;EAGb;EAEA,IAAI,OAAO,UAAU,UAAa,CAAC,KAAK,iBAAiB,OAAO,KAAK,GACnE,OAAO;EAGT,KAAK,MAAM,UAAU;GAAC;GAAS;GAAS;EAAO,GAAY;GACzD,MAAM,QAAQ,OAAO;GACrB,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,MAAM,QAAQ,CAAC,KAAK,iBAAiB,GAAG,CAAC,GACzE,OAAO;EAEX;EAEA,OAAO;CACT;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAChD,MAAM,kBAAkB,IAAI,2BAA2B;EAEvD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;GACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EACpF;CACF;;;;;;;;;;;;;;;;;;CAmBA,AAAQ,qBAAqB,WAE3B;EACA,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WAAW,QAC9C,OAAO,CAAC;EAGV,OAAO,EAAE,kBAAkB,UAAU,OAAO;CAC9C;;;;;;;;;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":["../../../../../../@warlock.js/ai-openai/src/model.ts"],"sourcesContent":["import {\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, OpenAIResponseFormat } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toOpenAIMessages, toOpenAITools, wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * Map an explicit `responseFormat` override to the default\n * `structuredOutput` capability. Loose wire modes (`\"json_object\"`,\n * `\"text\"`) don't enforce shape, so the agent needs to see the soft\n * schema hint in the system prompt — that only happens when the\n * capability is `false`. Default (no override) stays `true` to\n * preserve the prior assumption that OpenAI models support strict\n * structured output.\n */\nfunction inferStructuredOutput(responseFormat: OpenAIResponseFormat | undefined): boolean {\n  if (responseFormat === \"json_object\" || responseFormat === \"text\") {\n    return false;\n  }\n\n  return true;\n}\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    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 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      // OpenAI prompt caching is automatic on the Chat Completions API\n      // (no caller-supplied breakpoints — the platform caches long\n      // prompt prefixes server-side and reports the hit count via\n      // `prompt_tokens_details.cached_tokens`). We therefore advertise\n      // the read-side accounting capability as always available while\n      // treating `ModelCallOptions.cacheControl` write breakpoints as a\n      // no-op (see `buildReasoningParams` siblings — there is no cache\n      // param to emit).\n      promptCaching: true,\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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[0];\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),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning),\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  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter.\n   *\n   * When `config.responseFormat` is set, it wins: `\"text\"` emits no\n   * `response_format` at all, `\"json_object\"` always picks the loose\n   * mode, and `\"json_schema\"` picks strict mode (with the same\n   * `isStrictCompatible` safety check — a malformed schema still\n   * degrades to `json_object` rather than 400). The override exists\n   * because some targets (older OpenAI models, OpenRouter routes,\n   * Ollama OpenAI-compat) reject strict `json_schema` outright.\n   *\n   * When the override is omitted, uses strict `json_schema` mode\n   * (token-level enforcement) only when the schema is a proper\n   * root-object JSON Schema (`{ type: \"object\", properties: ... }`).\n   * For anything else — malformed extractor output, non-object\n   * schemas, or future shapes we haven't tested — falls back to loose\n   * `json_object` mode, which guarantees *some* valid JSON without\n   * enforcing shape. The agent's soft instruction already embeds the\n   * schema text in the system prompt when the model declares no\n   * native structured-output capability, so shape validation still\n   * runs client-side via the Standard Schema `validate()` call.\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    if (!responseSchema) {\n      return {};\n    }\n\n    const override = this.config.responseFormat;\n\n    if (override === \"text\") {\n      return {};\n    }\n\n    if (override === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    // Either auto-select (no override) or explicit `\"json_schema\"`.\n    // The strict-compat check still applies in the explicit case —\n    // a malformed / non-object schema would 400 before sampling, so\n    // we degrade to `json_object` rather than crash.\n    if (this.isStrictCompatible(responseSchema)) {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: {\n            name: \"response\",\n            schema: responseSchema,\n            strict: true,\n          },\n        },\n      };\n    }\n\n    return { response_format: { type: \"json_object\" } };\n  }\n\n  /**\n   * OpenAI strict `json_schema` mode requires the root to be a JSON\n   * Schema object type (`{ type: \"object\", properties: ... }`). Anything\n   * else (top-level arrays, primitives, unknown shapes) is rejected with\n   * a 400 before a token is sampled. We check structurally here so the\n   * first call doesn't crash on a malformed extraction — loose\n   * `json_object` mode is a safe degradation.\n   */\n  private isStrictCompatible(schema: Record<string, unknown>): boolean {\n    return (\n      schema.type === \"object\" &&\n      typeof schema.properties === \"object\" &&\n      schema.properties !== null &&\n      this.isStrictSafeNode(schema)\n    );\n  }\n\n  /**\n   * Recursively check the one strict-mode rule schemas most often trip on:\n   * every object must list ALL of its `properties` in `required` (OpenAI\n   * strict has no notion of optional — optional fields must be expressed\n   * as nullable, e.g. `type: [\"string\", \"null\"]`, and still appear in\n   * `required`). A schema that violates this anywhere in the tree is NOT\n   * sent in strict `json_schema` mode — it degrades to loose\n   * `json_object` so a hand-built or optional-bearing schema can't 400\n   * the call (\"'required' ... must include every key in properties\").\n   * Client-side `validate()` still enforces the full shape.\n   */\n  private isStrictSafeNode(node: unknown): boolean {\n    if (!node || typeof node !== \"object\") {\n      return true;\n    }\n\n    const record = node as Record<string, unknown>;\n\n    if (record.type === \"object\" && record.properties && typeof record.properties === \"object\") {\n      const properties = record.properties as Record<string, unknown>;\n      const keys = Object.keys(properties);\n      const required = Array.isArray(record.required) ? (record.required as unknown[]) : [];\n\n      if (keys.some((key) => !required.includes(key))) {\n        return false;\n      }\n\n      for (const key of keys) {\n        if (!this.isStrictSafeNode(properties[key])) {\n          return false;\n        }\n      }\n    }\n\n    if (record.items !== undefined && !this.isStrictSafeNode(record.items)) {\n      return false;\n    }\n\n    for (const branch of [\"anyOf\", \"allOf\", \"oneOf\"] as const) {\n      const value = record[branch];\n      if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) {\n        return false;\n      }\n    }\n\n    return true;\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    const cachedTokens = raw.prompt_tokens_details?.cached_tokens;\n    const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;\n\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n      ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\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 in two cases so the adapter never forwards an unsupported\n   * param: (1) the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`), or (2) the caller supplied no `effort`.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  private buildReasoningParams(reasoning: ModelCallOptions[\"reasoning\"]): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning || !reasoning?.effort) {\n      return {};\n    }\n\n    return { reasoning_effort: 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":";;;;;;;;;;;AAmBA,MAAM,aAAa;;;;;;;;;;AAWnB,SAAS,sBAAsB,gBAA2D;CACxF,IAAI,mBAAmB,iBAAiB,mBAAmB,QACzD,OAAO;CAGT,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAGhC,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;GASnE,eAAe;GAMf,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,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,SAAS,SAAS,QAAQ;EAChC,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,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,SAAS;GACjD,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;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4BA,AAAQ,oBAAoB,gBAE1B;EACA,IAAI,CAAC,gBACH,OAAO,CAAC;EAGV,MAAM,WAAW,KAAK,OAAO;EAE7B,IAAI,aAAa,QACf,OAAO,CAAC;EAGV,IAAI,aAAa,eACf,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAOpD,IAAI,KAAK,mBAAmB,cAAc,GACxC,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IACX,MAAM;IACN,QAAQ;IACR,QAAQ;GACV;EACF,EACF;EAGF,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;CACpD;;;;;;;;;CAUA,AAAQ,mBAAmB,QAA0C;EACnE,OACE,OAAO,SAAS,YAChB,OAAO,OAAO,eAAe,YAC7B,OAAO,eAAe,QACtB,KAAK,iBAAiB,MAAM;CAEhC;;;;;;;;;;;;CAaA,AAAQ,iBAAiB,MAAwB;EAC/C,IAAI,CAAC,QAAQ,OAAO,SAAS,UAC3B,OAAO;EAGT,MAAM,SAAS;EAEf,IAAI,OAAO,SAAS,YAAY,OAAO,cAAc,OAAO,OAAO,eAAe,UAAU;GAC1F,MAAM,aAAa,OAAO;GAC1B,MAAM,OAAO,OAAO,KAAK,UAAU;GACnC,MAAM,WAAW,MAAM,QAAQ,OAAO,QAAQ,IAAK,OAAO,WAAyB,CAAC;GAEpF,IAAI,KAAK,MAAM,QAAQ,CAAC,SAAS,SAAS,GAAG,CAAC,GAC5C,OAAO;GAGT,KAAK,MAAM,OAAO,MAChB,IAAI,CAAC,KAAK,iBAAiB,WAAW,IAAI,GACxC,OAAO;EAGb;EAEA,IAAI,OAAO,UAAU,UAAa,CAAC,KAAK,iBAAiB,OAAO,KAAK,GACnE,OAAO;EAGT,KAAK,MAAM,UAAU;GAAC;GAAS;GAAS;EAAO,GAAY;GACzD,MAAM,QAAQ,OAAO;GACrB,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,MAAM,QAAQ,CAAC,KAAK,iBAAiB,GAAG,CAAC,GACzE,OAAO;EAEX;EAEA,OAAO;CACT;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAChD,MAAM,kBAAkB,IAAI,2BAA2B;EAEvD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;GACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EACpF;CACF;;;;;;;;;;;;;;;;;;;;;;;;;;CA2BA,AAAQ,qBAAqB,WAE3B;EACA,IAAI,CAAC,KAAK,aAAa,aAAa,CAAC,WAAW,QAC9C,OAAO,CAAC;EAGV,OAAO,EAAE,kBAAkB,UAAU,OAAO;CAC9C;;;;;;;;;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"}

package/llms-full.txt

@@ -129,10 +129,29 @@ const model = openai.model({ name: "o3-mini" });   // reasoning auto-true
 await model.complete(messages, { reasoning: { effort: "high" } });  // → reasoning_effort: "high"
 ```
 
-- `reasoning.effort` (`"low" | "medium" | "high"`) maps verbatim to OpenAI's `reasoning_effort` request param.
+- `reasoning.effort` (`"low" | "medium" | "high" | "none"`) maps verbatim to OpenAI's `reasoning_effort` request param.
 - `reasoning.maxTokens` has **no Chat Completions equivalent** (it's the Anthropic extended-thinking budget) and is silently ignored here.
 - When `capabilities.reasoning` is `false` (e.g. `gpt-4o`), the option is dropped — the adapter never forwards `reasoning_effort` to a model that would 400 on it. Pin `reasoning: true` to force it for a custom/fine-tuned reasoning model.
 
+### `effort: "none"` — reasoning off, tools on
+
+gpt-5 / o-series models **reject function tools** on the Chat Completions API while reasoning is active:
+
+```
+400 — Function tools with reasoning_effort are not supported for <model>
+in /v1/chat/completions. To use function tools, use /v1/responses or set
+reasoning_effort to 'none'.
+```
+
+Pass `reasoning: { effort: "none" }` to run such a model **without reasoning** so tool-using agents work:
+
+```ts
+const model = openai.model({ name: "gpt-5-mini" });               // reasoning auto-true
+await model.complete(messages, { reasoning: { effort: "none" }, tools }); // → reasoning_effort: "none"
+```
+
+The adapter **emits** `reasoning_effort: "none"` explicitly — it does **not** omit the param. Omitting it leaves the model reasoning server-side by default, so tools would still be rejected; `"none"` is the switch that turns reasoning off on the wire. Trade-off: you lose reasoning. For tool-heavy agent work (function calls + good replies rather than deep analysis) this is usually the right call — it's the difference between empty replies and working ones. When you need reasoning **and** tools together, use the Responses API (planned).
+
 ## Token usage — what's reported
 
 `usage` on the result carries the neutral channel breakdown, populated from OpenAI's `usage` block:

package/package.json

@@ -14,12 +14,12 @@
   },
   "dependencies": {
     "openai": "^6.34.0",
-    "@warlock.js/logger": "4.7.0"
+    "@warlock.js/logger": "4.8.0"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "4.7.0"
+    "@warlock.js/ai": "4.8.0"
   },
-  "version": "4.7.0",
+  "version": "4.8.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/setup-openai/SKILL.md

@@ -121,10 +121,29 @@ const model = openai.model({ name: "o3-mini" });   // reasoning auto-true
 await model.complete(messages, { reasoning: { effort: "high" } });  // → reasoning_effort: "high"
 ```
 
-- `reasoning.effort` (`"low" | "medium" | "high"`) maps verbatim to OpenAI's `reasoning_effort` request param.
+- `reasoning.effort` (`"low" | "medium" | "high" | "none"`) maps verbatim to OpenAI's `reasoning_effort` request param.
 - `reasoning.maxTokens` has **no Chat Completions equivalent** (it's the Anthropic extended-thinking budget) and is silently ignored here.
 - When `capabilities.reasoning` is `false` (e.g. `gpt-4o`), the option is dropped — the adapter never forwards `reasoning_effort` to a model that would 400 on it. Pin `reasoning: true` to force it for a custom/fine-tuned reasoning model.
 
+### `effort: "none"` — reasoning off, tools on
+
+gpt-5 / o-series models **reject function tools** on the Chat Completions API while reasoning is active:
+
+```
+400 — Function tools with reasoning_effort are not supported for <model>
+in /v1/chat/completions. To use function tools, use /v1/responses or set
+reasoning_effort to 'none'.
+```
+
+Pass `reasoning: { effort: "none" }` to run such a model **without reasoning** so tool-using agents work:
+
+```ts
+const model = openai.model({ name: "gpt-5-mini" });               // reasoning auto-true
+await model.complete(messages, { reasoning: { effort: "none" }, tools }); // → reasoning_effort: "none"
+```
+
+The adapter **emits** `reasoning_effort: "none"` explicitly — it does **not** omit the param. Omitting it leaves the model reasoning server-side by default, so tools would still be rejected; `"none"` is the switch that turns reasoning off on the wire. Trade-off: you lose reasoning. For tool-heavy agent work (function calls + good replies rather than deep analysis) this is usually the right call — it's the difference between empty replies and working ones. When you need reasoning **and** tools together, use the Responses API (planned).
+
 ## Token usage — what's reported
 
 `usage` on the result carries the neutral channel breakdown, populated from OpenAI's `usage` block: