@warlock.js/ai-openai 4.2.10 → 4.3.0

package/CHANGELOG.md

@@ -6,6 +6,16 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## 4.3.0 - 2026-06-21
+
+### Added
+
+- Cost-truth wiring (additive, non-breaking):
+  - `Usage.reasoningTokens` now populated from `completion_tokens_details.reasoning_tokens` on both `complete()` and the streaming `done` event (o-series / gpt-5 hidden reasoning channel). Emitted only when > 0.
+  - `ModelCallOptions.reasoning.effort` mapped to the provider-native `reasoning_effort` request param. Forwarded only for reasoning-capable models; `reasoning.maxTokens` has no Chat Completions equivalent and is ignored.
+  - `ModelCapabilities.reasoning` inferred from the model name (`o1*` / `o3*` / `o4*` / `gpt-5*`), overridable via `.model({ name, reasoning })`. New `known-reasoning-models.ts` prefix helper.
+  - `ModelCapabilities.promptCaching` advertised as always `true` (OpenAI caches automatically and reports hits via `Usage.cachedTokens`). `ModelCallOptions.cacheControl` write breakpoints are a graceful no-op.
+
 ## 4.1.15
 
 - Baseline — per-package changelog tracking starts at this version.

package/cjs/index.cjs

@@ -404,6 +404,46 @@ var OpenAIEmbedder = class {
 	}
 };
 
+//#endregion
+//#region ../@warlock.js/ai-openai/src/known-reasoning-models.ts
+/**
+* Model-name prefixes for OpenAI families that expose internal
+* reasoning / thinking tokens and accept the `reasoning_effort`
+* request parameter on the Chat Completions API.
+*
+* Matched as a prefix so dated variants (`o3-2025-04-16`) and
+* `-mini` / `-pro` suffixes (`o4-mini`, `gpt-5-pro`) are covered
+* without listing every release tag explicitly.
+*
+* Maintenance: append a new prefix when OpenAI ships a reasoning
+* model family that doesn't already match. Devs can always override
+* per-model via `openai.model({ name, reasoning: true | false })` —
+* explicit config wins over inference in either direction.
+*/
+const REASONING_CAPABLE_PREFIXES = [
+	"o1",
+	"o3",
+	"o4",
+	"gpt-5"
+];
+/**
+* Infer whether a given OpenAI model name is a reasoning model (o-series
+* and the gpt-5 family) based on the known-prefix list. Unknown models
+* default to `false` so the adapter never forwards an unsupported
+* `reasoning_effort` param to a non-reasoning model (which would 400).
+*
+* @example
+* inferReasoningCapability("o3-mini");          // → true
+* inferReasoningCapability("o4-mini");          // → true
+* inferReasoningCapability("gpt-5-pro");        // → true
+* inferReasoningCapability("gpt-4o");           // → false
+* inferReasoningCapability("custom-llm");       // → false
+*/
+function inferReasoningCapability(modelName) {
+	const normalized = modelName.toLowerCase();
+	return REASONING_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));
+}
+
 //#endregion
 //#region ../@warlock.js/ai-openai/src/known-vision-models.ts
 /**
@@ -507,7 +547,9 @@ var OpenAIModel = class {
 		this.pricing = config.pricing;
 		this.capabilities = {
 			structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),
-			vision: config.vision ?? inferVisionCapability(config.name)
+			vision: config.vision ?? inferVisionCapability(config.name),
+			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
+			promptCaching: true
 		};
 	}
 	/**
@@ -531,7 +573,8 @@ var OpenAIModel = class {
 				temperature: options?.temperature ?? this.config.temperature,
 				max_tokens: options?.maxTokens ?? this.config.maxTokens,
 				tools: toOpenAITools(options?.tools),
-				...this.buildResponseFormat(options?.responseSchema)
+				...this.buildResponseFormat(options?.responseSchema),
+				...this.buildReasoningParams(options?.reasoning)
 			}, options?.signal ? { signal: options.signal } : void 0);
 		} catch (thrown) {
 			const wrapped = wrapOpenAIError(thrown);
@@ -578,7 +621,8 @@ var OpenAIModel = class {
 				tools: toOpenAITools(options?.tools),
 				stream: true,
 				stream_options: { include_usage: true },
-				...this.buildResponseFormat(options?.responseSchema)
+				...this.buildResponseFormat(options?.responseSchema),
+				...this.buildReasoningParams(options?.reasoning)
 			}, options?.signal ? { signal: options.signal } : void 0);
 		} catch (thrown) {
 			const wrapped = wrapOpenAIError(thrown);
@@ -622,6 +666,8 @@ var OpenAIModel = class {
 					usage.total = chunk.usage.total_tokens ?? 0;
 					const cached = chunk.usage.prompt_tokens_details?.cached_tokens;
 					if (cached !== void 0 && cached > 0) usage.cachedTokens = cached;
+					const reasoning = chunk.usage.completion_tokens_details?.reasoning_tokens;
+					if (reasoning !== void 0 && reasoning > 0) usage.reasoningTokens = reasoning;
 				}
 			}
 			for (const acc of toolCallAccum.values()) {
@@ -709,6 +755,14 @@ var OpenAIModel = class {
 	* or partials) into the neutral `Usage` shape. Missing usage collapses to
 	* zeros rather than propagating `undefined`, so downstream aggregation
 	* math stays safe.
+	*
+	* `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the
+	* subset of the prompt served from OpenAI's automatic prompt cache);
+	* `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`
+	* (the hidden reasoning channel on o-series / gpt-5 models, already
+	* counted within `output`). Both are emitted only when the provider
+	* reports a positive value, so non-reasoning / uncached calls keep the
+	* lean `{ input, output, total }` shape.
 	*/
 	extractUsage(raw) {
 		if (!raw) return {
@@ -717,14 +771,37 @@ var OpenAIModel = class {
 			total: 0
 		};
 		const cachedTokens = raw.prompt_tokens_details?.cached_tokens;
+		const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;
 		return {
 			input: raw.prompt_tokens,
 			output: raw.completion_tokens,
 			total: raw.total_tokens,
-			...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {}
+			...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {},
+			...reasoningTokens !== void 0 && reasoningTokens > 0 ? { reasoningTokens } : {}
 		};
 	}
 	/**
+	* Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's
+	* `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat
+	* Completions API exposes a discrete effort knob, not a token budget,
+	* so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no
+	* wire equivalent here and is silently ignored.
+	*
+	* 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.
+	*
+	* Returns an empty spread when nothing applies, so the caller can
+	* unconditionally `...buildReasoningParams(...)` into the request.
+	*/
+	buildReasoningParams(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning?.effort) return {};
+		return { reasoning_effort: reasoning.effort };
+	}
+	/**
 	* Reshape OpenAI's `tool_calls` array into the neutral
 	* `ModelToolCallRequest[]`. The raw `arguments` field is a JSON string
 	* per OpenAI's protocol — we parse it defensively via `safeJsonParse` so

package/cjs/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","ContentFilterError","InvalidRequestError","ProviderError","OpenAI","LOG_MODULE","log","log","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-vision-models.ts","../../../../../../@warlock.js/ai-openai/src/model.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 type { ContentPart, 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\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  // TODO: Allow other types for urls not just images\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","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-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 { 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    };\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        },\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        },\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        }\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    );\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  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\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n    };\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 OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type { OpenAIEmbedderConfig, OpenAIModelConfig, OpenAISDKConfig } from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIModel } from \"./model\";\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    this.client = new OpenAI({\n      apiKey: config.apiKey,\n      baseURL: config.baseURL,\n    });\n    this.provider = config.provider ?? \"openai\";\n    this.pricing = config.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"],"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;AAEA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CASzC,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;AACjD;;;;;;;;;;;;;ACjFA,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,kBAAkBA,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,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;;;;ACpBA,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;gBAFzDE;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;EAC5D;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;GACrD,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;GACrD,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;IAEzB;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,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;CAE1B;;;;;;;CAQA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAEhD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;EAC3E;CACF;;;;;;;;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AChWA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAC1C,KAAK,SAAS,IAAIC,eAAO;GACvB,QAAQ,OAAO;GACf,SAAS,OAAO;EAClB,CAAC;EACD,KAAK,WAAW,OAAO,YAAY;EACnC,KAAK,UAAU,OAAO;CACxB;;;;;;;;;;;;;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;AACF"}
+{"version":3,"file":"index.cjs","names":["AIError","ProviderTimeoutError","ProviderAuthError","QuotaExceededError","ProviderRateLimitError","ContextLengthExceededError","ContentFilterError","InvalidRequestError","ProviderError","OpenAI","LOG_MODULE","log","log","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-reasoning-models.ts","../../../../../../@warlock.js/ai-openai/src/known-vision-models.ts","../../../../../../@warlock.js/ai-openai/src/model.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 type { ContentPart, 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\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  // TODO: Allow other types for urls not just images\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","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-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    };\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    );\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 OpenAI from \"openai\";\nimport type {\n  EmbedderContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type { OpenAIEmbedderConfig, OpenAIModelConfig, OpenAISDKConfig } from \"./config.type\";\nimport { OpenAIEmbedder } from \"./embedder\";\nimport { OpenAIModel } from \"./model\";\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    this.client = new OpenAI({\n      apiKey: config.apiKey,\n      baseURL: config.baseURL,\n    });\n    this.provider = config.provider ?? \"openai\";\n    this.pricing = config.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"],"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;AAEA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CASzC,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;AACjD;;;;;;;;;;;;;ACjFA,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,kBAAkBA,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;;;;;;;;;;;;;;;;;;AChIA,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,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;gBAFzDE;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;EACjB;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,yCAA8C,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;CAE1B;;;;;;;;;;;;;;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACzZA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAC1C,KAAK,SAAS,IAAIC,eAAO;GACvB,QAAQ,OAAO;GACf,SAAS,OAAO;EAClB,CAAC;EACD,KAAK,WAAW,OAAO,YAAY;EACnC,KAAK,UAAU,OAAO;CACxB;;;;;;;;;;;;;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;AACF"}

package/esm/config.type.d.mts

@@ -96,6 +96,17 @@ type OpenAIModelConfig = ModelConfig & {
    * `responseFormat`.
    */
   structuredOutput?: boolean;
+  /**
+   * Override the auto-inferred `reasoning` capability. When omitted,
+   * the adapter checks the model name against a known-prefix list (see
+   * `known-reasoning-models.ts`) — `true` for the o-series (`o1*`,
+   * `o3*`, `o4*`) and the `gpt-5*` family, `false` otherwise. Setting
+   * this explicitly always wins over inference — useful for fine-tuned
+   * reasoning models or gateways exposing reasoning under a custom
+   * name. When `false`, `ModelCallOptions.reasoning` is ignored rather
+   * than forwarded as an unsupported `reasoning_effort` param.
+   */
+  reasoning?: boolean;
 };
 /**
  * Per-embedder configuration for `OpenAISDK.embedder()`. Mirrors the

package/esm/config.type.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-openai/src/config.type.ts"],"mappings":";;;;;;AAiCA;;;;;;;;;;;;;;AAOuC;AAiBvC;;;;AAAgC;AAUhC;;;;;;;;KAlCY,eAAA,GAAkB,aAAA;EAC5B,QAAA;EAgEgB;AAalB;;;;EAvEE,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;;;;;;;KAiBf,oBAAA;;;;;;;;;KAUA,iBAAA,GAAoB,WAAA;;;;;;;;EAQ9B,MAAA;;;;;;;;;;;;;;EAcA,cAAA,GAAiB,oBAAoB;;;;;;;;;EASrC,gBAAA;AAAA;;;;;;;;;;;KAaU,oBAAA,GAAuB,cAAc"}
+{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../@warlock.js/ai-openai/src/config.type.ts"],"mappings":";;;;;;AAiCA;;;;;;;;;;;;;;AAOuC;AAiBvC;;;;AAAgC;AAUhC;;;;;;;;KAlCY,eAAA,GAAkB,aAAA;EAC5B,QAAA;EA2ES;AAAA;AAaX;;;EAlFE,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;;;;;;;KAiBf,oBAAA;;;;;;;;;KAUA,iBAAA,GAAoB,WAAA;;;;;;;;EAQ9B,MAAA;;;;;;;;;;;;;;EAcA,cAAA,GAAiB,oBAAoB;;;;;;;;;EASrC,gBAAA;;;;;;;;;;;EAWA,SAAA;AAAA;;;;;;;;;;;KAaU,oBAAA,GAAuB,cAAc"}

package/esm/known-reasoning-models.mjs

@@ -0,0 +1,42 @@
+//#region ../@warlock.js/ai-openai/src/known-reasoning-models.ts
+/**
+* Model-name prefixes for OpenAI families that expose internal
+* reasoning / thinking tokens and accept the `reasoning_effort`
+* request parameter on the Chat Completions API.
+*
+* Matched as a prefix so dated variants (`o3-2025-04-16`) and
+* `-mini` / `-pro` suffixes (`o4-mini`, `gpt-5-pro`) are covered
+* without listing every release tag explicitly.
+*
+* Maintenance: append a new prefix when OpenAI ships a reasoning
+* model family that doesn't already match. Devs can always override
+* per-model via `openai.model({ name, reasoning: true | false })` —
+* explicit config wins over inference in either direction.
+*/
+const REASONING_CAPABLE_PREFIXES = [
+	"o1",
+	"o3",
+	"o4",
+	"gpt-5"
+];
+/**
+* Infer whether a given OpenAI model name is a reasoning model (o-series
+* and the gpt-5 family) based on the known-prefix list. Unknown models
+* default to `false` so the adapter never forwards an unsupported
+* `reasoning_effort` param to a non-reasoning model (which would 400).
+*
+* @example
+* inferReasoningCapability("o3-mini");          // → true
+* inferReasoningCapability("o4-mini");          // → true
+* inferReasoningCapability("gpt-5-pro");        // → true
+* inferReasoningCapability("gpt-4o");           // → false
+* inferReasoningCapability("custom-llm");       // → false
+*/
+function inferReasoningCapability(modelName) {
+	const normalized = modelName.toLowerCase();
+	return REASONING_CAPABLE_PREFIXES.some((prefix) => normalized.startsWith(prefix));
+}
+
+//#endregion
+export { inferReasoningCapability };
+//# sourceMappingURL=known-reasoning-models.mjs.map

package/esm/known-reasoning-models.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"known-reasoning-models.mjs","names":[],"sources":["../../../../../../@warlock.js/ai-openai/src/known-reasoning-models.ts"],"sourcesContent":["/**\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"],"mappings":";;;;;;;;;;;;;;;AAcA,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"}

package/esm/model.mjs

@@ -3,6 +3,7 @@ import { toOpenAIMessages } from "./utils/to-openai-messages.mjs";
 import { toOpenAITools } from "./utils/to-openai-tools.mjs";
 import { wrapOpenAIError } from "./utils/wrap-openai-error.mjs";
 import "./utils/index.mjs";
+import { inferReasoningCapability } from "./known-reasoning-models.mjs";
 import { inferVisionCapability } from "./known-vision-models.mjs";
 import { safeJsonParse } from "@warlock.js/ai";
 import { log } from "@warlock.js/logger";
@@ -69,7 +70,9 @@ var OpenAIModel = class {
 		this.pricing = config.pricing;
 		this.capabilities = {
 			structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),
-			vision: config.vision ?? inferVisionCapability(config.name)
+			vision: config.vision ?? inferVisionCapability(config.name),
+			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
+			promptCaching: true
 		};
 	}
 	/**
@@ -93,7 +96,8 @@ var OpenAIModel = class {
 				temperature: options?.temperature ?? this.config.temperature,
 				max_tokens: options?.maxTokens ?? this.config.maxTokens,
 				tools: toOpenAITools(options?.tools),
-				...this.buildResponseFormat(options?.responseSchema)
+				...this.buildResponseFormat(options?.responseSchema),
+				...this.buildReasoningParams(options?.reasoning)
 			}, options?.signal ? { signal: options.signal } : void 0);
 		} catch (thrown) {
 			const wrapped = wrapOpenAIError(thrown);
@@ -140,7 +144,8 @@ var OpenAIModel = class {
 				tools: toOpenAITools(options?.tools),
 				stream: true,
 				stream_options: { include_usage: true },
-				...this.buildResponseFormat(options?.responseSchema)
+				...this.buildResponseFormat(options?.responseSchema),
+				...this.buildReasoningParams(options?.reasoning)
 			}, options?.signal ? { signal: options.signal } : void 0);
 		} catch (thrown) {
 			const wrapped = wrapOpenAIError(thrown);
@@ -184,6 +189,8 @@ var OpenAIModel = class {
 					usage.total = chunk.usage.total_tokens ?? 0;
 					const cached = chunk.usage.prompt_tokens_details?.cached_tokens;
 					if (cached !== void 0 && cached > 0) usage.cachedTokens = cached;
+					const reasoning = chunk.usage.completion_tokens_details?.reasoning_tokens;
+					if (reasoning !== void 0 && reasoning > 0) usage.reasoningTokens = reasoning;
 				}
 			}
 			for (const acc of toolCallAccum.values()) {
@@ -271,6 +278,14 @@ var OpenAIModel = class {
 	* or partials) into the neutral `Usage` shape. Missing usage collapses to
 	* zeros rather than propagating `undefined`, so downstream aggregation
 	* math stays safe.
+	*
+	* `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the
+	* subset of the prompt served from OpenAI's automatic prompt cache);
+	* `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`
+	* (the hidden reasoning channel on o-series / gpt-5 models, already
+	* counted within `output`). Both are emitted only when the provider
+	* reports a positive value, so non-reasoning / uncached calls keep the
+	* lean `{ input, output, total }` shape.
 	*/
 	extractUsage(raw) {
 		if (!raw) return {
@@ -279,14 +294,37 @@ var OpenAIModel = class {
 			total: 0
 		};
 		const cachedTokens = raw.prompt_tokens_details?.cached_tokens;
+		const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;
 		return {
 			input: raw.prompt_tokens,
 			output: raw.completion_tokens,
 			total: raw.total_tokens,
-			...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {}
+			...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {},
+			...reasoningTokens !== void 0 && reasoningTokens > 0 ? { reasoningTokens } : {}
 		};
 	}
 	/**
+	* Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's
+	* `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat
+	* Completions API exposes a discrete effort knob, not a token budget,
+	* so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no
+	* wire equivalent here and is silently ignored.
+	*
+	* 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.
+	*
+	* Returns an empty spread when nothing applies, so the caller can
+	* unconditionally `...buildReasoningParams(...)` into the request.
+	*/
+	buildReasoningParams(reasoning) {
+		if (!this.capabilities.reasoning || !reasoning?.effort) return {};
+		return { reasoning_effort: reasoning.effort };
+	}
+	/**
 	* Reshape OpenAI's `tool_calls` array into the neutral
 	* `ModelToolCallRequest[]`. The raw `arguments` field is a JSON string
 	* per OpenAI's protocol — we parse it defensively via `safeJsonParse` so

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 { 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    };\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        },\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        },\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        }\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    );\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  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\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n    };\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":";;;;;;;;;;AAkBA,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;EAC5D;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;GACrD,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;GACrD,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;IAEzB;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;CAE1B;;;;;;;CAQA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAEhD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;EAC3E;CACF;;;;;;;;;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    };\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    );\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;EACjB;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;CAE1B;;;;;;;;;;;;;;;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"}

package/llms-full.txt

@@ -8,7 +8,7 @@
 
 ---
 name: setup-openai
-description: 'Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`; "wire openai into a warlock agent", "configure gpt-4o", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.'
+description: 'Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, reasoning?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`, `reasoning_effort`, `reasoningTokens`, `cachedTokens`, o-series / gpt-5 reasoning, prompt caching; "wire openai into a warlock agent", "configure gpt-4o", "use o3 / gpt-5 reasoning effort", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.'
 ---
 
 # `@warlock.js/ai-openai`
@@ -52,8 +52,10 @@ Returns a `ModelContract` you pass straight into `ai.agent({ model })`.
 | --- | --- |
 | `structuredOutput` | `true`, unless `responseFormat` is set to `"json_object"` or `"text"` (loose modes) — then `false`. |
 | `vision` | Inferred from model name. `true` for `gpt-4o*`, `gpt-4-turbo*`, `gpt-4.1*`, `o1*`, `o3*`, `chatgpt-4o*`; `false` otherwise. |
+| `reasoning` | Inferred from model name. `true` for the o-series (`o1*`, `o3*`, `o4*`) and the `gpt-5*` family; `false` otherwise. Drives whether `reasoning_effort` is forwarded. |
+| `promptCaching` | Always `true`. OpenAI caches long prompt prefixes automatically and reports hits via `usage.cachedTokens` — there are no caller-supplied write breakpoints. |
 
-**Override either flag explicitly** via `.model({ name, vision?, structuredOutput? })` — an explicit value always wins over inference.
+**Override `vision`, `structuredOutput`, or `reasoning` explicitly** via `.model({ name, vision?, structuredOutput?, reasoning? })` — an explicit value always wins over inference.
 
 ## Structured output
 
@@ -99,6 +101,33 @@ const { vectors } = await embedder.embedMany(["doc 1", "doc 2", "doc 3"]);
 openai.embedder({ name: "text-embedding-3-large", dimensions: 256 });
 ```
 
+## Reasoning (o-series / gpt-5)
+
+Reasoning models accept a discrete effort knob. The agent passes it through `ModelCallOptions.reasoning`:
+
+```ts
+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.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.
+
+## Token usage — what's reported
+
+`usage` on the result carries the neutral channel breakdown, populated from OpenAI's `usage` block:
+
+| `Usage` field | OpenAI source | Notes |
+| --- | --- | --- |
+| `input` / `output` / `total` | `prompt_tokens` / `completion_tokens` / `total_tokens` | always present (zeroed if OpenAI omits the block). |
+| `cachedTokens` | `prompt_tokens_details.cached_tokens` | subset of `input` served from OpenAI's automatic prompt cache. Emitted only when > 0. |
+| `reasoningTokens` | `completion_tokens_details.reasoning_tokens` | hidden reasoning channel on o-series / gpt-5 (already counted within `output`). Emitted only when > 0. |
+
+Both `cachedTokens` and `reasoningTokens` are omitted (not set to `0`) when the provider reports zero, so non-reasoning / uncached calls keep the lean `{ input, output, total }` shape. Reported identically on `complete()` and the streaming `done` event.
+
+`cacheControl` write breakpoints (`ModelCallOptions.cacheControl`) are a **no-op** — OpenAI has no caller-driven cache marker on the Chat Completions wire. The read-side `cachedTokens` accounting above works without any caller action.
+
 ## Token counting
 
 ```ts

package/llms.txt

@@ -6,4 +6,4 @@
 
 ## Skills
 
-- [setup-openai](@warlock.js/ai-openai/setup-openai/SKILL.md): Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`; "wire openai into a warlock agent", "configure gpt-4o", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.
+- [setup-openai](@warlock.js/ai-openai/setup-openai/SKILL.md): Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, reasoning?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`, `reasoning_effort`, `reasoningTokens`, `cachedTokens`, o-series / gpt-5 reasoning, prompt caching; "wire openai into a warlock agent", "configure gpt-4o", "use o3 / gpt-5 reasoning effort", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.

package/package.json

@@ -14,12 +14,12 @@
   },
   "dependencies": {
     "openai": "^6.34.0",
-    "@warlock.js/logger": "4.2.10"
+    "@warlock.js/logger": "4.3.0"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "4.2.10"
+    "@warlock.js/ai": "4.3.0"
   },
-  "version": "4.2.10",
+  "version": "4.3.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/README.md

@@ -6,4 +6,4 @@ Per-task skills. All cross-references use the form `@warlock.js/<pkg>/<skill>/SK
 
 ### [`setup-openai/`](./setup-openai/SKILL.md)
 
-Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, structuredOutput?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Load when wiring an OpenAI-backed model into a @warlock.js agent or routing via an OpenAI-compatible gateway.
+Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, reasoning?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Covers o-series / gpt-5 reasoning (reasoning_effort), automatic prompt caching with cost-truth usage (cachedTokens/reasoningTokens), and the per-model pricing registry. Load when wiring an OpenAI-backed model into a @warlock.js agent or routing via an OpenAI-compatible gateway.

package/skills/setup-openai/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: setup-openai
-description: 'Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`; "wire openai into a warlock agent", "configure gpt-4o", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.'
+description: 'Wire @warlock.js/ai-openai — new OpenAISDK({apiKey, baseURL?, provider?, pricing?}) for OpenAI / Azure / OpenRouter, .model({name, vision?, reasoning?, structuredOutput?, responseFormat?}) for ModelContract, .embedder({name, dimensions?}) for embeddings. Triggers: `OpenAISDK`, `.model`, `.embedder`, `.embed`, `.embedMany`, `baseURL`, `pricing`, `responseSchema`, `responseFormat`, `reasoning_effort`, `reasoningTokens`, `cachedTokens`, o-series / gpt-5 reasoning, prompt caching; "wire openai into a warlock agent", "configure gpt-4o", "use o3 / gpt-5 reasoning effort", "route through openrouter or azure openai", "openai embeddings with warlock"; typical import `import { OpenAISDK } from "@warlock.js/ai-openai"`. Skip: agent wiring — `@warlock.js/ai/run-ai-agent/SKILL.md`; adapter comparison — `@warlock.js/ai/pick-ai-provider/SKILL.md`; competing adapters `@warlock.js/ai-anthropic`, `@warlock.js/ai-bedrock`, `@warlock.js/ai-google`, `@warlock.js/ai-ollama`; raw `openai` SDK, Vercel `@ai-sdk/openai`.'
 ---
 
 # `@warlock.js/ai-openai`
@@ -44,8 +44,10 @@ Returns a `ModelContract` you pass straight into `ai.agent({ model })`.
 | --- | --- |
 | `structuredOutput` | `true`, unless `responseFormat` is set to `"json_object"` or `"text"` (loose modes) — then `false`. |
 | `vision` | Inferred from model name. `true` for `gpt-4o*`, `gpt-4-turbo*`, `gpt-4.1*`, `o1*`, `o3*`, `chatgpt-4o*`; `false` otherwise. |
+| `reasoning` | Inferred from model name. `true` for the o-series (`o1*`, `o3*`, `o4*`) and the `gpt-5*` family; `false` otherwise. Drives whether `reasoning_effort` is forwarded. |
+| `promptCaching` | Always `true`. OpenAI caches long prompt prefixes automatically and reports hits via `usage.cachedTokens` — there are no caller-supplied write breakpoints. |
 
-**Override either flag explicitly** via `.model({ name, vision?, structuredOutput? })` — an explicit value always wins over inference.
+**Override `vision`, `structuredOutput`, or `reasoning` explicitly** via `.model({ name, vision?, structuredOutput?, reasoning? })` — an explicit value always wins over inference.
 
 ## Structured output
 
@@ -91,6 +93,33 @@ const { vectors } = await embedder.embedMany(["doc 1", "doc 2", "doc 3"]);
 openai.embedder({ name: "text-embedding-3-large", dimensions: 256 });
 ```
 
+## Reasoning (o-series / gpt-5)
+
+Reasoning models accept a discrete effort knob. The agent passes it through `ModelCallOptions.reasoning`:
+
+```ts
+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.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.
+
+## Token usage — what's reported
+
+`usage` on the result carries the neutral channel breakdown, populated from OpenAI's `usage` block:
+
+| `Usage` field | OpenAI source | Notes |
+| --- | --- | --- |
+| `input` / `output` / `total` | `prompt_tokens` / `completion_tokens` / `total_tokens` | always present (zeroed if OpenAI omits the block). |
+| `cachedTokens` | `prompt_tokens_details.cached_tokens` | subset of `input` served from OpenAI's automatic prompt cache. Emitted only when > 0. |
+| `reasoningTokens` | `completion_tokens_details.reasoning_tokens` | hidden reasoning channel on o-series / gpt-5 (already counted within `output`). Emitted only when > 0. |
+
+Both `cachedTokens` and `reasoningTokens` are omitted (not set to `0`) when the provider reports zero, so non-reasoning / uncached calls keep the lean `{ input, output, total }` shape. Reported identically on `complete()` and the streaming `done` event.
+
+`cacheControl` write breakpoints (`ModelCallOptions.cacheControl`) are a **no-op** — OpenAI has no caller-driven cache marker on the Chat Completions wire. The read-side `cachedTokens` accounting above works without any caller action.
+
 ## Token counting
 
 ```ts