@warlock.js/ai-openai 5.25.0 → 5.27.0

package/esm/config.type.d.mts

@@ -1,4 +1,4 @@
-import { ClientOptions } from "openai";
+import OpenAI, { ClientOptions } from "openai";
 import { EmbedderConfig, ImageModelConfig, ModelConfig, ModelPricing, SpeechModelConfig, TranscriptionModelConfig } from "@warlock.js/ai";
 
 //#region ../ai-openai/src/config.type.d.ts
@@ -64,6 +64,18 @@ type OpenAISDKConfig = Omit<ClientOptions, "provider"> & {
  * openai.model({ name: "some-legacy-model", responseFormat: "json_object" });
  */
 type OpenAIResponseFormat = "json_schema" | "json_object" | "text";
+/**
+ * Which OpenAI endpoint a model speaks.
+ *
+ * - `"chat"` (default) - Chat Completions. Reasoning models reject function
+ *   tools here unless reasoning is off, so the adapter sends
+ *   `reasoning_effort: "none"` when tools are attached.
+ * - `"responses"` - the Responses API, stateless (`store: false`). Reasoning
+ *   stays at the provider default even with tools attached. Direct OpenAI
+ *   only: `OpenAISDK.model()` throws for any other provider label, so the
+ *   wrapper packages (deepseek, groq, xai, mistral) can never reach it.
+ */
+type OpenAIApi = "chat" | "responses";
 /**
  * Per-model configuration for `OpenAISDK.model()`. Extends the neutral
  * `ModelConfig` with OpenAI-specific capability overrides.
@@ -71,8 +83,18 @@ type OpenAIResponseFormat = "json_schema" | "json_object" | "text";
  * @example
  * openai.model({ name: "gpt-4o-mini" });               // vision auto-true
  * openai.model({ name: "fine-tuned-x", vision: true }); // dev override
+ *
+ * @example
+ * // Reasoning model + tools on the Responses API (direct OpenAI only):
+ * openai.model({ name: "gpt-5.6", api: "responses" });
  */
 type OpenAIModelConfig = ModelConfig & {
+  /**
+   * Wire API to use. Default `"chat"` (Chat Completions). `"responses"`
+   * selects the stateless Responses adapter and is accepted only on the
+   * direct `"openai"` provider label; see {@link OpenAIApi}.
+   */
+  api?: OpenAIApi;
   /**
    * Override the auto-inferred vision capability. When omitted, the
    * adapter checks the model name against a known-prefix list (see
@@ -108,13 +130,25 @@ type OpenAIModelConfig = ModelConfig & {
    * 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
+   * `o3*`, `o4*`) and the `gpt-5*` / `gpt-6*` families, `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;
+  /**
+   * Stable request bucket used by OpenAI to improve prompt-cache matching.
+   * Sent only by models whose SDK provider label is `"openai"`; compatible
+   * wrapper endpoints deliberately never receive this OpenAI-only field.
+   */
+  promptCacheKey?: string;
+  /**
+   * OpenAI's maximum prompt-cache retention policy. `"24h"` enables
+   * extended retention; `"in_memory"` keeps the cache within memory.
+   * Sent only by models whose SDK provider label is `"openai"`.
+   */
+  promptCacheRetention?: Exclude<OpenAI.Chat.Completions.ChatCompletionCreateParams["prompt_cache_retention"], null>;
   /**
    * Opt into PDF / document **input**. Default `false` — OpenAI accepts
    * PDF `file` content parts only on specific models (the `gpt-4o`

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

@@ -1 +1 @@
-{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../ai-openai/src/config.type.ts"],"mappings":";;;;;;AAwCA;;;;;;;;;;;;;;;;AAeuC;AAiBvC;;;;AAAgC;AAUhC;;;;;;KA1CY,eAAA,GAAkB,IAAA,CAAK,aAAA;EAgEhB;;;;;;AAqCZ;AAaP;EAzGE,QAAA;;;AAyG+C;AAcjD;;EAjHE,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;AAiHqB;AAahD;;;;AAAkD;AAalD;;;;AAAgE;;;;AA1BhB,KAhGpC,oBAAA;;;;;;;;;KAUA,iBAAA,GAAoB,WAAA;;;;;;;;EAQ9B,MAAA;;;;;;;;;;;;;;EAcA,cAAA,GAAiB,oBAAoB;;;;;;;;;EASrC,gBAAA;;;;;;;;;;;EAWA,SAAA;;;;;;;;;EASA,GAAA;;;;;;;;EAQA,KAAA;AAAA;;;;;;;;;;;KAaU,oBAAA,GAAuB,cAAc;;;;;;;;;;;;;KAcrC,iBAAA,GAAoB,gBAAgB;;;;;;;;;;;;KAapC,kBAAA,GAAqB,iBAAiB;;;;;;;;;;;;KAatC,yBAAA,GAA4B,wBAAwB"}
+{"version":3,"file":"config.type.d.mts","names":[],"sources":["../../../../../../ai-openai/src/config.type.ts"],"mappings":";;;;;AAyCA;;;;;;;;;;;;;;;;AAeuC;AAiBvC;;;;AAAgC;AAahC;;;;AAAqB;AAcrB;;KA3DY,eAAA,GAAkB,IAAA,CAAK,aAAA;EA2DH;;;;;;;;EAlD9B,QAAA;EAwDM;;;;;EAlDN,OAAA,GAAU,MAAA,SAAe,YAAA;AAAA;;;;;;;;;;AA4HpB;AAaP;;;;KAxHY,oBAAA;AAsIZ;;;;AAAgD;AAahD;;;;AAAkD;AAalD;AA1BA,KAzHY,SAAA;;;AAmJoD;;;;;;;;;;KArIpD,iBAAA,GAAoB,WAAA;;;;;;EAM9B,GAAA,GAAM,SAAA;;;;;;;;EAQN,MAAA;;;;;;;;;;;;;;EAcA,cAAA,GAAiB,oBAAA;;;;;;;;;EASjB,gBAAA;;;;;;;;;;;EAWA,SAAA;;;;;;EAMA,cAAA;;;;;;EAMA,oBAAA,GAAuB,OAAA,CACrB,MAAA,CAAO,IAAA,CAAK,WAAA,CAAY,0BAAA;;;;;;;;;EAW1B,GAAA;;;;;;;;EAQA,KAAA;AAAA;;;;;;;;;;;KAaU,oBAAA,GAAuB,cAAc;;;;;;;;;;;;;KAcrC,iBAAA,GAAoB,gBAAgB;;;;;;;;;;;;KAapC,kBAAA,GAAqB,iBAAiB;;;;;;;;;;;;KAatC,yBAAA,GAA4B,wBAAwB"}

package/esm/known-reasoning-models.mjs

@@ -17,11 +17,12 @@ const REASONING_CAPABLE_PREFIXES = [
 	"o1",
 	"o3",
 	"o4",
-	"gpt-5"
+	"gpt-5",
+	"gpt-6"
 ];
 /**
 * 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
+* and the gpt-5 / gpt-6 families) 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).
 *
@@ -29,6 +30,7 @@ const REASONING_CAPABLE_PREFIXES = [
 * inferReasoningCapability("o3-mini");          // → true
 * inferReasoningCapability("o4-mini");          // → true
 * inferReasoningCapability("gpt-5-pro");        // → true
+* inferReasoningCapability("gpt-6-astra");      // → true
 * inferReasoningCapability("gpt-4o");           // → false
 * inferReasoningCapability("custom-llm");       // → false
 */

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

@@ -1 +1 @@
-{"version":3,"file":"known-reasoning-models.mjs","names":[],"sources":["../../../../../../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"}
+{"version":3,"file":"known-reasoning-models.mjs","names":[],"sources":["../../../../../../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\", \"gpt-6\"];\n\n/**\n * Infer whether a given OpenAI model name is a reasoning model (o-series\n * and the gpt-5 / gpt-6 families) 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-6-astra\");      // → 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;CAAS;AAAO;;;;;;;;;;;;;;;AAgBtE,SAAgB,yBAAyB,WAA4B;CACnE,MAAM,aAAa,UAAU,YAAY;CAEzC,OAAO,2BAA2B,MAAM,WAAW,WAAW,WAAW,MAAM,CAAC;AAClF"}

package/esm/known-vision-models.mjs

@@ -16,8 +16,11 @@ const VISION_CAPABLE_PREFIXES = [
 	"gpt-4o",
 	"gpt-4-turbo",
 	"gpt-4.1",
+	"gpt-5",
+	"gpt-6",
 	"o1",
 	"o3",
+	"o4",
 	"chatgpt-4o"
 ];
 /**

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

@@ -1 +1 @@
-{"version":3,"file":"known-vision-models.mjs","names":[],"sources":["../../../../../../ai-openai/src/known-vision-models.ts"],"sourcesContent":["/**\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"],"mappings":";;;;;;;;;;;;;;AAaA,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"}
+{"version":3,"file":"known-vision-models.mjs","names":[],"sources":["../../../../../../ai-openai/src/known-vision-models.ts"],"sourcesContent":["/**\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  \"gpt-5\",\n  \"gpt-6\",\n  \"o1\",\n  \"o3\",\n  \"o4\",\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"],"mappings":";;;;;;;;;;;;;;AAaA,MAAM,0BAA0B;CAC9B;CACA;CACA;CACA;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"}

package/esm/model.mjs

@@ -1,29 +1,20 @@
+import { buildUsage } from "./utils/build-usage.mjs";
+import { wrapOpenAIError } from "./utils/wrap-openai-error.mjs";
 import { mapFinishReason } from "./utils/map-finish-reason.mjs";
+import { buildPromptCacheParams } from "./utils/prompt-cache-params.mjs";
+import { toOpenAIReasoningEffort } from "./utils/reasoning-effort.mjs";
+import { inferStructuredOutput, planStructuredOutput } from "./utils/structured-output.mjs";
 import { 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 { ProviderError, safeJsonParse } from "@warlock.js/ai";
+import { InvalidRequestError, ProviderError, safeJsonParse } from "@warlock.js/ai";
 import { log } from "@warlock.js/logger";
 
 //#region ../ai-openai/src/model.ts
 const LOG_MODULE = "ai.openai";
 /**
-* Map an explicit `responseFormat` override to the default
-* `structuredOutput` capability. Loose wire modes (`"json_object"`,
-* `"text"`) don't enforce shape, so the agent needs to see the soft
-* schema hint in the system prompt — that only happens when the
-* capability is `false`. Default (no override) stays `true` to
-* preserve the prior assumption that OpenAI models support strict
-* structured output.
-*/
-function inferStructuredOutput(responseFormat) {
-	if (responseFormat === "json_object" || responseFormat === "text") return false;
-	return true;
-}
-/**
 * OpenAI-backed implementation of `ModelContract`.
 *
 * **Role.** The provider-facing bridge between the vendor-neutral
@@ -63,6 +54,11 @@ function inferStructuredOutput(responseFormat) {
 var OpenAIModel = class {
 	constructor(client, config, provider = "openai") {
 		this.logger = log;
+		if (config.api === "responses") throw new InvalidRequestError("OpenAIModel speaks Chat Completions only and cannot serve api: \"responses\". Use OpenAISDK.model({ api: \"responses\" }) on the direct OpenAI provider.", { context: {
+			provider,
+			model: config.name,
+			api: config.api
+		} });
 		this.client = client;
 		this.config = config;
 		this.name = config.name;
@@ -72,7 +68,7 @@ var OpenAIModel = class {
 			structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),
 			vision: config.vision ?? inferVisionCapability(config.name),
 			reasoning: config.reasoning ?? inferReasoningCapability(config.name),
-			promptCaching: true,
+			promptCaching: this.provider === "openai",
 			pdf: config.pdf ?? false,
 			audio: config.audio ?? false
 		};
@@ -94,9 +90,9 @@ var OpenAIModel = class {
 		try {
 			response = await this.client.chat.completions.create({
 				model: this.name,
-				messages: toOpenAIMessages(messages),
-				temperature: options?.temperature ?? this.config.temperature,
-				max_tokens: options?.maxTokens ?? this.config.maxTokens,
+				messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),
+				...this.buildSamplingParams(options),
+				...buildPromptCacheParams(this.provider, this.config),
 				tools: toOpenAITools(options?.tools),
 				...this.buildResponseFormat(options?.responseSchema),
 				...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length))
@@ -145,9 +141,9 @@ var OpenAIModel = class {
 		try {
 			stream = await this.client.chat.completions.create({
 				model: this.name,
-				messages: toOpenAIMessages(messages),
-				temperature: options?.temperature ?? this.config.temperature,
-				max_tokens: options?.maxTokens ?? this.config.maxTokens,
+				messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),
+				...this.buildSamplingParams(options),
+				...buildPromptCacheParams(this.provider, this.config),
 				tools: toOpenAITools(options?.tools),
 				stream: true,
 				stream_options: { include_usage: true },
@@ -228,89 +224,31 @@ var OpenAIModel = class {
 			usage
 		};
 	}
+	/** Map the neutral cache hint only for the direct OpenAI provider. */
+	promptCacheBreakpoints(options) {
+		return this.provider === "openai" ? options?.cacheControl?.breakpoints ?? 0 : 0;
+	}
 	/**
 	* Translate the neutral `responseSchema` option into OpenAI's
-	* `response_format` parameter.
-	*
-	* When `config.responseFormat` is set, it wins: `"text"` emits no
-	* `response_format` at all, `"json_object"` always picks the loose
-	* mode, and `"json_schema"` picks strict mode (with the same
-	* `isStrictCompatible` safety check — a malformed schema still
-	* degrades to `json_object` rather than 400). The override exists
-	* because some targets (older OpenAI models, OpenRouter routes,
-	* Ollama OpenAI-compat) reject strict `json_schema` outright.
-	*
-	* When the override is omitted, uses strict `json_schema` mode
-	* (token-level enforcement) only when the schema is a proper
-	* root-object JSON Schema (`{ type: "object", properties: ... }`).
-	* For anything else — malformed extractor output, non-object
-	* schemas, or future shapes we haven't tested — falls back to loose
-	* `json_object` mode, which guarantees *some* valid JSON without
-	* enforcing shape. The agent's soft instruction already embeds the
-	* schema text in the system prompt when the model declares no
-	* native structured-output capability, so shape validation still
-	* runs client-side via the Standard Schema `validate()` call.
+	* `response_format` parameter. The mode decision (override handling,
+	* strict-compatibility downgrade to `json_object`) lives in
+	* `planStructuredOutput`, shared with the Responses adapter.
 	*
 	* Returns an empty spread when no schema was supplied, so the caller
 	* can unconditionally `...buildResponseFormat(...)` into the request.
 	*/
 	buildResponseFormat(responseSchema) {
-		if (!responseSchema) return {};
-		const override = this.config.responseFormat;
-		if (override === "text") return {};
-		if (override === "json_object") return { response_format: { type: "json_object" } };
-		if (this.isStrictCompatible(responseSchema)) return { response_format: {
+		const plan = planStructuredOutput(responseSchema, this.config.responseFormat);
+		if (plan.mode === "json_schema") return { response_format: {
 			type: "json_schema",
 			json_schema: {
 				name: "response",
-				schema: responseSchema,
+				schema: plan.schema,
 				strict: true
 			}
 		} };
-		return { response_format: { type: "json_object" } };
-	}
-	/**
-	* OpenAI strict `json_schema` mode requires the root to be a JSON
-	* Schema object type (`{ type: "object", properties: ... }`). Anything
-	* else (top-level arrays, primitives, unknown shapes) is rejected with
-	* a 400 before a token is sampled. We check structurally here so the
-	* first call doesn't crash on a malformed extraction — loose
-	* `json_object` mode is a safe degradation.
-	*/
-	isStrictCompatible(schema) {
-		return schema.type === "object" && typeof schema.properties === "object" && schema.properties !== null && this.isStrictSafeNode(schema);
-	}
-	/**
-	* Recursively check the one strict-mode rule schemas most often trip on:
-	* every object must list ALL of its `properties` in `required` (OpenAI
-	* strict has no notion of optional — optional fields must be expressed
-	* as nullable, e.g. `type: ["string", "null"]`, and still appear in
-	* `required`). A schema that violates this anywhere in the tree is NOT
-	* sent in strict `json_schema` mode — it degrades to loose
-	* `json_object` so a hand-built or optional-bearing schema can't 400
-	* the call ("'required' ... must include every key in properties").
-	* Client-side `validate()` still enforces the full shape.
-	*/
-	isStrictSafeNode(node) {
-		if (!node || typeof node !== "object") return true;
-		const record = node;
-		if (record.type === "object" && record.properties && typeof record.properties === "object") {
-			const properties = record.properties;
-			const keys = Object.keys(properties);
-			const required = Array.isArray(record.required) ? record.required : [];
-			if (keys.some((key) => !required.includes(key))) return false;
-			for (const key of keys) if (!this.isStrictSafeNode(properties[key])) return false;
-		}
-		if (record.items !== void 0 && !this.isStrictSafeNode(record.items)) return false;
-		for (const branch of [
-			"anyOf",
-			"allOf",
-			"oneOf"
-		]) {
-			const value = record[branch];
-			if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) return false;
-		}
-		return true;
+		if (plan.mode === "json_object") return { response_format: { type: "json_object" } };
+		return {};
 	}
 	/**
 	* Normalize OpenAI's `usage` block (which may be absent on some responses
@@ -332,15 +270,13 @@ var OpenAIModel = class {
 			output: 0,
 			total: 0
 		};
-		const cachedTokens = raw.prompt_tokens_details?.cached_tokens;
-		const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;
-		return {
+		return buildUsage({
 			input: raw.prompt_tokens,
 			output: raw.completion_tokens,
 			total: raw.total_tokens,
-			...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {},
-			...reasoningTokens !== void 0 && reasoningTokens > 0 ? { reasoningTokens } : {}
-		};
+			cachedTokens: raw.prompt_tokens_details?.cached_tokens,
+			reasoningTokens: raw.completion_tokens_details?.reasoning_tokens
+		});
 	}
 	/**
 	* Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's
@@ -377,10 +313,27 @@ var OpenAIModel = class {
 	* Returns an empty spread when nothing applies, so the caller can
 	* unconditionally `...buildReasoningParams(...)` into the request.
 	*/
+	/**
+	* Output-length and sampling params, in current Chat Completions shapes only.
+	*
+	* - `max_completion_tokens`, never the legacy `max_tokens`: current
+	*   models (the gpt-5 family and later) answer `400 unsupported_parameter`
+	*   to `max_tokens`, and `max_completion_tokens` is accepted everywhere.
+	* - No `temperature` for a reasoning-capable model: those accept only the
+	*   default value and reject any other with a 400.
+	*/
+	buildSamplingParams(options) {
+		const maxTokens = options?.maxTokens ?? this.config.maxTokens;
+		const temperature = options?.temperature ?? this.config.temperature;
+		return {
+			max_completion_tokens: maxTokens,
+			...!this.capabilities.reasoning && { temperature }
+		};
+	}
 	buildReasoningParams(reasoning, hasTools) {
 		if (!this.capabilities.reasoning) return {};
 		if (!reasoning?.effort) return hasTools ? { reasoning_effort: "none" } : {};
-		return { reasoning_effort: reasoning.effort };
+		return { reasoning_effort: toOpenAIReasoningEffort(reasoning.effort) };
 	}
 	/**
 	* Reshape OpenAI's `tool_calls` array into the neutral

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../ai-openai/src/model.ts"],"sourcesContent":["import {\n  ProviderError,\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIModelConfig, OpenAIResponseFormat } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toOpenAIMessages, toOpenAITools, wrapOpenAIError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * Map an explicit `responseFormat` override to the default\n * `structuredOutput` capability. Loose wire modes (`\"json_object\"`,\n * `\"text\"`) don't enforce shape, so the agent needs to see the soft\n * schema hint in the system prompt — that only happens when the\n * capability is `false`. Default (no override) stays `true` to\n * preserve the prior assumption that OpenAI models support strict\n * structured output.\n */\nfunction inferStructuredOutput(responseFormat: OpenAIResponseFormat | undefined): boolean {\n  if (responseFormat === \"json_object\" || responseFormat === \"text\") {\n    return false;\n  }\n\n  return true;\n}\n\n/**\n * OpenAI-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `openai` SDK. Agents,\n * workflows, and supervisors never talk to OpenAI directly — they hold a\n * `ModelContract`, and this class is what makes that contract concrete for\n * any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, OpenRouter, local\n * gateways that speak the Chat Completions protocol).\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client + frozen `ModelConfig` (name,\n *   temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` and\n *   `ToolContract[]` into OpenAI wire shapes on the way out, and\n *   translating OpenAI's response (content, finish reason, tool calls,\n *   usage) back into the neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as a\n * class (see §4.2 of code-style.md — \"long-lived state across calls\").\n *\n * @example\n * import OpenAI from \"openai\";\n * const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n * const model = new OpenAIModel(client, { name: \"gpt-4o\", temperature: 0.3 });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class OpenAIModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly config: OpenAIModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIModelConfig, provider: string = \"openai\") {\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),\n      vision: config.vision ?? inferVisionCapability(config.name),\n      // o-series + gpt-5 models surface a reasoning channel and accept\n      // the `reasoning_effort` param. Explicit config wins over the\n      // name-prefix inference.\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      // OpenAI prompt caching is automatic on the Chat Completions API\n      // (no caller-supplied breakpoints — the platform caches long\n      // prompt prefixes server-side and reports the hit count via\n      // `prompt_tokens_details.cached_tokens`). We therefore advertise\n      // the read-side accounting capability as always available while\n      // treating `ModelCallOptions.cacheControl` write breakpoints as a\n      // no-op (see `buildReasoningParams` siblings — there is no cache\n      // param to emit).\n      promptCaching: true,\n      // PDF + audio INPUT are off by default — OpenAI accepts `file`\n      // (PDF) and `input_audio` parts only on specific models, so the\n      // flags are conservative/honest and opt-in via config rather than\n      // name-inferred. When set, the agent admits the attachments and\n      // `toOpenAIMessages` maps them to the real wire parts.\n      pdf: config.pdf ?? false,\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Chat\n   * Completions endpoint, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    // Per-call request/response logs are hot-path in production agents\n    // — keep them at `debug` so `info` stays reserved for lifecycle\n    // events (agent starting/completed, etc.). Operators who need to\n    // audit every LLM call can raise log-level at runtime.\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: OpenAI.Chat.Completions.ChatCompletion;\n\n    try {\n      response = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages),\n          temperature: options?.temperature ?? this.config.temperature,\n          max_tokens: options?.maxTokens ?? this.config.maxTokens,\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const [choice] = response.choices;\n\n    if (choice === undefined) {\n      // A 200 carrying an empty `choices` array. Every line below reads off\n      // `choice`, so without this the caller got a bare `TypeError: Cannot\n      // read properties of undefined` — no provider, no model, no hint that\n      // the response itself was the problem rather than our handling of it.\n      //\n      // `ProviderError` because that is what this is: the provider answered\n      // successfully and returned nothing usable. Callers already branch on\n      // `instanceof ProviderError` for any provider-side failure.\n      throw new ProviderError(\"OpenAI returned a response with no choices.\", {\n        context: { provider: \"openai\", model: this.name, choices: response.choices.length },\n      });\n    }\n\n    const finishReason = mapFinishReason(choice.finish_reason);\n    const usage = this.extractUsage(response.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: choice.message.content ?? \"\",\n      finishReason,\n      usage,\n      toolCalls: this.extractToolCalls(choice.message.tool_calls),\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s —\n   * `delta` for text tokens, `tool-call` when the model requests a tool,\n   * and a terminal `done` carrying the final finish reason + usage totals.\n   * Callers consume it with `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<typeof this.client.chat.completions.create>>;\n\n    try {\n      stream = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages),\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, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    let rawFinishReason: string = \"stop\";\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolCallAccum = new Map<number, { id: string; name: string; arguments: string }>();\n\n    try {\n      for await (const chunk of stream as AsyncIterable<OpenAI.Chat.Completions.ChatCompletionChunk>) {\n        const delta = chunk.choices[0]?.delta;\n        const finish = chunk.choices[0]?.finish_reason;\n\n        if (delta?.content) {\n          yield { type: \"delta\", content: delta.content };\n        }\n\n        if (delta?.tool_calls) {\n          for (const toolCall of delta.tool_calls) {\n            const idx = toolCall.index ?? 0;\n            if (!toolCallAccum.has(idx)) {\n              toolCallAccum.set(idx, { id: \"\", name: \"\", arguments: \"\" });\n            }\n            const acc = toolCallAccum.get(idx)!;\n            if (toolCall.id) acc.id = toolCall.id;\n            if (toolCall.function?.name) acc.name = toolCall.function.name;\n            if (toolCall.function?.arguments) acc.arguments += toolCall.function.arguments;\n          }\n        }\n\n        if (finish) {\n          rawFinishReason = finish;\n        }\n\n        if (chunk.usage) {\n          usage.input = chunk.usage.prompt_tokens ?? 0;\n          usage.output = chunk.usage.completion_tokens ?? 0;\n          usage.total = chunk.usage.total_tokens ?? 0;\n          const cached = chunk.usage.prompt_tokens_details?.cached_tokens;\n          if (cached !== undefined && cached > 0) {\n            usage.cachedTokens = cached;\n          }\n          const reasoning = chunk.usage.completion_tokens_details?.reasoning_tokens;\n          if (reasoning !== undefined && reasoning > 0) {\n            usage.reasoningTokens = reasoning;\n          }\n        }\n      }\n\n      for (const acc of toolCallAccum.values()) {\n        // Skip accumulators that never received a function name — those\n        // are partial fragments the model started but never identified\n        // (e.g. arguments-only deltas with no originating `id`/`name`).\n        // Yielding them produces nameless tool-calls the agent runtime\n        // can't dispatch and would mis-attribute as a registered tool.\n        if (!acc.name) continue;\n\n        yield {\n          type: \"tool-call\",\n          id: acc.id,\n          name: acc.name,\n          input: safeJsonParse<Record<string, unknown>>(acc.arguments, {}),\n        };\n      }\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const finishReason = mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter.\n   *\n   * When `config.responseFormat` is set, it wins: `\"text\"` emits no\n   * `response_format` at all, `\"json_object\"` always picks the loose\n   * mode, and `\"json_schema\"` picks strict mode (with the same\n   * `isStrictCompatible` safety check — a malformed schema still\n   * degrades to `json_object` rather than 400). The override exists\n   * because some targets (older OpenAI models, OpenRouter routes,\n   * Ollama OpenAI-compat) reject strict `json_schema` outright.\n   *\n   * When the override is omitted, uses strict `json_schema` mode\n   * (token-level enforcement) only when the schema is a proper\n   * root-object JSON Schema (`{ type: \"object\", properties: ... }`).\n   * For anything else — malformed extractor output, non-object\n   * schemas, or future shapes we haven't tested — falls back to loose\n   * `json_object` mode, which guarantees *some* valid JSON without\n   * enforcing shape. The agent's soft instruction already embeds the\n   * schema text in the system prompt when the model declares no\n   * native structured-output capability, so shape validation still\n   * runs client-side via the Standard Schema `validate()` call.\n   *\n   * Returns an empty spread when no schema was supplied, so the caller\n   * can unconditionally `...buildResponseFormat(...)` into the request.\n   */\n  private buildResponseFormat(responseSchema: Record<string, unknown> | undefined): {\n    response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"response_format\"];\n  } {\n    if (!responseSchema) {\n      return {};\n    }\n\n    const override = this.config.responseFormat;\n\n    if (override === \"text\") {\n      return {};\n    }\n\n    if (override === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    // Either auto-select (no override) or explicit `\"json_schema\"`.\n    // The strict-compat check still applies in the explicit case —\n    // a malformed / non-object schema would 400 before sampling, so\n    // we degrade to `json_object` rather than crash.\n    if (this.isStrictCompatible(responseSchema)) {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: {\n            name: \"response\",\n            schema: responseSchema,\n            strict: true,\n          },\n        },\n      };\n    }\n\n    return { response_format: { type: \"json_object\" } };\n  }\n\n  /**\n   * OpenAI strict `json_schema` mode requires the root to be a JSON\n   * Schema object type (`{ type: \"object\", properties: ... }`). Anything\n   * else (top-level arrays, primitives, unknown shapes) is rejected with\n   * a 400 before a token is sampled. We check structurally here so the\n   * first call doesn't crash on a malformed extraction — loose\n   * `json_object` mode is a safe degradation.\n   */\n  private isStrictCompatible(schema: Record<string, unknown>): boolean {\n    return (\n      schema.type === \"object\" &&\n      typeof schema.properties === \"object\" &&\n      schema.properties !== null &&\n      this.isStrictSafeNode(schema)\n    );\n  }\n\n  /**\n   * Recursively check the one strict-mode rule schemas most often trip on:\n   * every object must list ALL of its `properties` in `required` (OpenAI\n   * strict has no notion of optional — optional fields must be expressed\n   * as nullable, e.g. `type: [\"string\", \"null\"]`, and still appear in\n   * `required`). A schema that violates this anywhere in the tree is NOT\n   * sent in strict `json_schema` mode — it degrades to loose\n   * `json_object` so a hand-built or optional-bearing schema can't 400\n   * the call (\"'required' ... must include every key in properties\").\n   * Client-side `validate()` still enforces the full shape.\n   */\n  private isStrictSafeNode(node: unknown): boolean {\n    if (!node || typeof node !== \"object\") {\n      return true;\n    }\n\n    const record = node as Record<string, unknown>;\n\n    if (record.type === \"object\" && record.properties && typeof record.properties === \"object\") {\n      const properties = record.properties as Record<string, unknown>;\n      const keys = Object.keys(properties);\n      const required = Array.isArray(record.required) ? (record.required as unknown[]) : [];\n\n      if (keys.some((key) => !required.includes(key))) {\n        return false;\n      }\n\n      for (const key of keys) {\n        if (!this.isStrictSafeNode(properties[key])) {\n          return false;\n        }\n      }\n    }\n\n    if (record.items !== undefined && !this.isStrictSafeNode(record.items)) {\n      return false;\n    }\n\n    for (const branch of [\"anyOf\", \"allOf\", \"oneOf\"] as const) {\n      const value = record[branch];\n      if (Array.isArray(value) && value.some((sub) => !this.isStrictSafeNode(sub))) {\n        return false;\n      }\n    }\n\n    return true;\n  }\n\n  /**\n   * Normalize OpenAI's `usage` block (which may be absent on some responses\n   * or partials) into the neutral `Usage` shape. Missing usage collapses to\n   * zeros rather than propagating `undefined`, so downstream aggregation\n   * math stays safe.\n   *\n   * `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the\n   * subset of the prompt served from OpenAI's automatic prompt cache);\n   * `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`\n   * (the hidden reasoning channel on o-series / gpt-5 models, already\n   * counted within `output`). Both are emitted only when the provider\n   * reports a positive value, so non-reasoning / uncached calls keep the\n   * lean `{ input, output, total }` shape.\n   */\n  private extractUsage(raw: OpenAI.Completions.CompletionUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    const cachedTokens = raw.prompt_tokens_details?.cached_tokens;\n    const reasoningTokens = raw.completion_tokens_details?.reasoning_tokens;\n\n    return {\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n      ...(reasoningTokens !== undefined && reasoningTokens > 0 ? { reasoningTokens } : {}),\n    };\n  }\n\n  /**\n   * Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's\n   * `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat\n   * Completions API exposes a discrete effort knob, not a token budget,\n   * so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no\n   * wire equivalent here and is silently ignored.\n   *\n   * The neutral `ReasoningEffort` (`\"low\" | \"medium\" | \"high\" | \"none\"`)\n   * is a subset of OpenAI's accepted values, so it forwards verbatim —\n   * `\"none\"` included. `\"none\"` is load-bearing: gpt-5 / o-series models\n   * **reject function tools** on Chat Completions while reasoning is\n   * active, and the endpoint accepts tools only when `reasoning_effort`\n   * is `\"none\"` (the alternative is the Responses API). This is why the\n   * param is EMITTED for `\"none\"` rather than omitted — omitting it\n   * leaves the model reasoning server-side by default, so tools would\n   * still be rejected.\n   *\n   * No-ops when the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`).\n   *\n   * When the caller gave no explicit `effort`: defaults to `\"none\"` IF\n   * `hasTools` is true, else omits the param (provider default — the\n   * pre-existing behavior for a tool-less call). The default exists\n   * because, on a reasoning-capable model, there is no working alternative\n   * to `\"none\"` when tools are attached — omitting `reasoning_effort`\n   * leaves reasoning on server-side, which Chat Completions rejects tools\n   * for (empty replies on some model generations, a hard 400 — \"Function\n   * tools with reasoning_effort are not supported ... in\n   * /v1/chat/completions\" — on newer ones). An explicit `effort` from the\n   * caller always wins over this default, in either direction.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  private buildReasoningParams(\n    reasoning: ModelCallOptions[\"reasoning\"],\n    hasTools: boolean,\n  ): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning) {\n      return {};\n    }\n\n    if (!reasoning?.effort) {\n      return hasTools ? { reasoning_effort: \"none\" } : {};\n    }\n\n    return { reasoning_effort: 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":";;;;;;;;;;;AAoBA,MAAM,aAAa;;;;;;;;;;AAWnB,SAAS,sBAAsB,gBAA2D;CACxF,IAAI,mBAAmB,iBAAiB,mBAAmB,QACzD,OAAO;CAGT,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAGhC,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB,sBAAsB,OAAO,cAAc;GACxF,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;GAI1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GASnE,eAAe;GAMf,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,KAAK,OAAO,MAAM,YAAY,WAAW,qCAAqC;GAC5E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAAK,YAAY,OAC5C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,QAAQ;IACnC,aAAa,SAAS,eAAe,KAAK,OAAO;IACjD,YAAY,SAAS,aAAa,KAAK,OAAO;IAC9C,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,CAAC,UAAU,SAAS;EAE1B,IAAI,WAAW,QASb,MAAM,IAAI,cAAc,+CAA+C,EACrE,SAAS;GAAE,UAAU;GAAU,OAAO,KAAK;GAAM,SAAS,SAAS,QAAQ;EAAO,EACpF,CAAC;EAGH,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,sCAAsC;GAC9E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,OAAO,QAAQ,WAAW;GACnC;GACA;GACA,WAAW,KAAK,iBAAiB,OAAO,QAAQ,UAAU;EAC5D;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,+CAA+C;GACtF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY,OAC1C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,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,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,IAAI,kBAA0B;EAC9B,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,gCAAgB,IAAI,IAA6D;EAEvF,IAAI;GACF,WAAW,MAAM,SAAS,QAAsE;IAC9F,MAAM,QAAQ,MAAM,QAAQ,EAAE,EAAE;IAChC,MAAM,SAAS,MAAM,QAAQ,EAAE,EAAE;IAEjC,IAAI,OAAO,SACT,MAAM;KAAE,MAAM;KAAS,SAAS,MAAM;IAAQ;IAGhD,IAAI,OAAO,YACT,KAAK,MAAM,YAAY,MAAM,YAAY;KACvC,MAAM,MAAM,SAAS,SAAS;KAC9B,IAAI,CAAC,cAAc,IAAI,GAAG,GACxB,cAAc,IAAI,KAAK;MAAE,IAAI;MAAI,MAAM;MAAI,WAAW;KAAG,CAAC;KAE5D,MAAM,MAAM,cAAc,IAAI,GAAG;KACjC,IAAI,SAAS,IAAI,IAAI,KAAK,SAAS;KACnC,IAAI,SAAS,UAAU,MAAM,IAAI,OAAO,SAAS,SAAS;KAC1D,IAAI,SAAS,UAAU,WAAW,IAAI,aAAa,SAAS,SAAS;IACvE;IAGF,IAAI,QACF,kBAAkB;IAGpB,IAAI,MAAM,OAAO;KACf,MAAM,QAAQ,MAAM,MAAM,iBAAiB;KAC3C,MAAM,SAAS,MAAM,MAAM,qBAAqB;KAChD,MAAM,QAAQ,MAAM,MAAM,gBAAgB;KAC1C,MAAM,SAAS,MAAM,MAAM,uBAAuB;KAClD,IAAI,WAAW,UAAa,SAAS,GACnC,MAAM,eAAe;KAEvB,MAAM,YAAY,MAAM,MAAM,2BAA2B;KACzD,IAAI,cAAc,UAAa,YAAY,GACzC,MAAM,kBAAkB;IAE5B;GACF;GAEA,KAAK,MAAM,OAAO,cAAc,OAAO,GAAG;IAMxC,IAAI,CAAC,IAAI,MAAM;IAEf,MAAM;KACJ,MAAM;KACN,IAAI,IAAI;KACR,MAAM,IAAI;KACV,OAAO,cAAuC,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAM,YAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4BA,AAAQ,oBAAoB,gBAE1B;EACA,IAAI,CAAC,gBACH,OAAO,CAAC;EAGV,MAAM,WAAW,KAAK,OAAO;EAE7B,IAAI,aAAa,QACf,OAAO,CAAC;EAGV,IAAI,aAAa,eACf,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAOpD,IAAI,KAAK,mBAAmB,cAAc,GACxC,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IACX,MAAM;IACN,QAAQ;IACR,QAAQ;GACV;EACF,EACF;EAGF,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;CACpD;;;;;;;;;CAUA,AAAQ,mBAAmB,QAA0C;EACnE,OACE,OAAO,SAAS,YAChB,OAAO,OAAO,eAAe,YAC7B,OAAO,eAAe,QACtB,KAAK,iBAAiB,MAAM;CAEhC;;;;;;;;;;;;CAaA,AAAQ,iBAAiB,MAAwB;EAC/C,IAAI,CAAC,QAAQ,OAAO,SAAS,UAC3B,OAAO;EAGT,MAAM,SAAS;EAEf,IAAI,OAAO,SAAS,YAAY,OAAO,cAAc,OAAO,OAAO,eAAe,UAAU;GAC1F,MAAM,aAAa,OAAO;GAC1B,MAAM,OAAO,OAAO,KAAK,UAAU;GACnC,MAAM,WAAW,MAAM,QAAQ,OAAO,QAAQ,IAAK,OAAO,WAAyB,CAAC;GAEpF,IAAI,KAAK,MAAM,QAAQ,CAAC,SAAS,SAAS,GAAG,CAAC,GAC5C,OAAO;GAGT,KAAK,MAAM,OAAO,MAChB,IAAI,CAAC,KAAK,iBAAiB,WAAW,IAAI,GACxC,OAAO;EAGb;EAEA,IAAI,OAAO,UAAU,UAAa,CAAC,KAAK,iBAAiB,OAAO,KAAK,GACnE,OAAO;EAGT,KAAK,MAAM,UAAU;GAAC;GAAS;GAAS;EAAO,GAAY;GACzD,MAAM,QAAQ,OAAO;GACrB,IAAI,MAAM,QAAQ,KAAK,KAAK,MAAM,MAAM,QAAQ,CAAC,KAAK,iBAAiB,GAAG,CAAC,GACzE,OAAO;EAEX;EAEA,OAAO;CACT;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,MAAM,eAAe,IAAI,uBAAuB;EAChD,MAAM,kBAAkB,IAAI,2BAA2B;EAEvD,OAAO;GACL,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;GACzE,GAAI,oBAAoB,UAAa,kBAAkB,IAAI,EAAE,gBAAgB,IAAI,CAAC;EACpF;CACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAqCA,AAAQ,qBACN,WACA,UAGA;EACA,IAAI,CAAC,KAAK,aAAa,WACrB,OAAO,CAAC;EAGV,IAAI,CAAC,WAAW,QACd,OAAO,WAAW,EAAE,kBAAkB,OAAO,IAAI,CAAC;EAGpD,OAAO,EAAE,kBAAkB,UAAU,OAAO;CAC9C;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,OAAO,cAAwC,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF"}
+{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../ai-openai/src/model.ts"],"sourcesContent":["import {\n  InvalidRequestError,\n  ProviderError,\n  safeJsonParse,\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type OpenAI from \"openai\";\nimport type { OpenAIModelConfig } from \"./config.type\";\nimport { inferReasoningCapability } from \"./known-reasoning-models\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport {\n  buildPromptCacheParams,\n  buildUsage,\n  inferStructuredOutput,\n  mapFinishReason,\n  planStructuredOutput,\n  toOpenAIMessages,\n  toOpenAIReasoningEffort,\n  toOpenAITools,\n  wrapOpenAIError,\n} from \"./utils\";\n\nconst LOG_MODULE = \"ai.openai\";\n\n/**\n * OpenAI-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the official `openai` SDK. Agents,\n * workflows, and supervisors never talk to OpenAI directly — they hold a\n * `ModelContract`, and this class is what makes that contract concrete for\n * any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, OpenRouter, local\n * gateways that speak the Chat Completions protocol).\n *\n * **Responsibility.**\n * - Owns: a long-lived `OpenAI` client + frozen `ModelConfig` (name,\n *   temperature, maxTokens) used as defaults for every call.\n * - Owns: translating vendor-neutral `Message[]` and\n *   `ToolContract[]` into OpenAI wire shapes on the way out, and\n *   translating OpenAI's response (content, finish reason, tool calls,\n *   usage) back into the neutral shapes on the way in.\n * - Does NOT own: dispatching tools, deciding whether to loop, tracking\n *   conversation history, or retrying on failure — those are agent\n *   concerns. The model is a stateless (per-call) protocol adapter.\n *\n * Because it holds a live client and shared defaults, it is modeled as a\n * class (see §4.2 of code-style.md — \"long-lived state across calls\").\n *\n * @example\n * import OpenAI from \"openai\";\n * const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n * const model = new OpenAIModel(client, { name: \"gpt-4o\", temperature: 0.3 });\n *\n * const myAgent = agent({\n *   model,\n *   systemPrompt: \"You are a helpful assistant.\",\n *   tools: [searchTool],\n * });\n *\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class OpenAIModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly client: OpenAI;\n  private readonly config: OpenAIModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(client: OpenAI, config: OpenAIModelConfig, provider: string = \"openai\") {\n    // `api: \"responses\"` belongs to `OpenAIResponsesModel`, which only the\n    // `OpenAISDK.model()` factory constructs. This class speaks Chat\n    // Completions only; accepting the flag here would silently ignore it.\n    if (config.api === \"responses\") {\n      throw new InvalidRequestError(\n        'OpenAIModel speaks Chat Completions only and cannot serve api: \"responses\". Use OpenAISDK.model({ api: \"responses\" }) on the direct OpenAI provider.',\n        { context: { provider, model: config.name, api: config.api } },\n      );\n    }\n\n    this.client = client;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? inferStructuredOutput(config.responseFormat),\n      vision: config.vision ?? inferVisionCapability(config.name),\n      // o-series + gpt-5 / gpt-6 models surface a reasoning channel and accept\n      // the `reasoning_effort` param. Explicit config wins over the\n      // name-prefix inference.\n      reasoning: config.reasoning ?? inferReasoningCapability(config.name),\n      // Direct OpenAI requests support automatic cache hits and explicit\n      // content-part breakpoints. Wrapper endpoints have distinct provider\n      // labels and may reject these OpenAI-only request fields.\n      promptCaching: this.provider === \"openai\",\n      // PDF + audio INPUT are off by default — OpenAI accepts `file`\n      // (PDF) and `input_audio` parts only on specific models, so the\n      // flags are conservative/honest and opt-in via config rather than\n      // name-inferred. When set, the agent admits the attachments and\n      // `toOpenAIMessages` maps them to the real wire parts.\n      pdf: config.pdf ?? false,\n      audio: config.audio ?? false,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to the Chat\n   * Completions endpoint, waits for the terminal response, and reshapes it\n   * into a vendor-neutral `ModelResponse`. Per-call `options` override the\n   * instance's `ModelConfig` defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    // Per-call request/response logs are hot-path in production agents\n    // — keep them at `debug` so `info` stays reserved for lifecycle\n    // events (agent starting/completed, etc.). Operators who need to\n    // audit every LLM call can raise log-level at runtime.\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let response: OpenAI.Chat.Completions.ChatCompletion;\n\n    try {\n      response = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n          ...this.buildSamplingParams(options),\n          ...buildPromptCacheParams(this.provider, this.config),\n          tools: toOpenAITools(options?.tools),\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const [choice] = response.choices;\n\n    if (choice === undefined) {\n      // A 200 carrying an empty `choices` array. Every line below reads off\n      // `choice`, so without this the caller got a bare `TypeError: Cannot\n      // read properties of undefined` — no provider, no model, no hint that\n      // the response itself was the problem rather than our handling of it.\n      //\n      // `ProviderError` because that is what this is: the provider answered\n      // successfully and returned nothing usable. Callers already branch on\n      // `instanceof ProviderError` for any provider-side failure.\n      throw new ProviderError(\"OpenAI returned a response with no choices.\", {\n        context: { provider: \"openai\", model: this.name, choices: response.choices.length },\n      });\n    }\n\n    const finishReason = mapFinishReason(choice.finish_reason);\n    const usage = this.extractUsage(response.usage);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: choice.message.content ?? \"\",\n      finishReason,\n      usage,\n      toolCalls: this.extractToolCalls(choice.message.tool_calls),\n    };\n  }\n\n  /**\n   * Incremental streaming completion. Yields neutral `ModelStreamChunk`s —\n   * `delta` for text tokens, `tool-call` when the model requests a tool,\n   * and a terminal `done` carrying the final finish reason + usage totals.\n   * Callers consume it with `for await`.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting streaming call to chat.completions\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    let stream: Awaited<ReturnType<typeof this.client.chat.completions.create>>;\n\n    try {\n      stream = await this.client.chat.completions.create(\n        {\n          model: this.name,\n          messages: toOpenAIMessages(messages, this.promptCacheBreakpoints(options)),\n          ...this.buildSamplingParams(options),\n          ...buildPromptCacheParams(this.provider, this.config),\n          tools: toOpenAITools(options?.tools),\n          stream: true,\n          stream_options: { include_usage: true },\n          ...this.buildResponseFormat(options?.responseSchema),\n          ...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length)),\n        },\n        options?.signal ? { signal: options.signal } : undefined,\n      );\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    let rawFinishReason: string = \"stop\";\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n    const toolCallAccum = new Map<number, { id: string; name: string; arguments: string }>();\n\n    try {\n      for await (const chunk of stream as AsyncIterable<OpenAI.Chat.Completions.ChatCompletionChunk>) {\n        const delta = chunk.choices[0]?.delta;\n        const finish = chunk.choices[0]?.finish_reason;\n\n        if (delta?.content) {\n          yield { type: \"delta\", content: delta.content };\n        }\n\n        if (delta?.tool_calls) {\n          for (const toolCall of delta.tool_calls) {\n            const idx = toolCall.index ?? 0;\n            if (!toolCallAccum.has(idx)) {\n              toolCallAccum.set(idx, { id: \"\", name: \"\", arguments: \"\" });\n            }\n            const acc = toolCallAccum.get(idx)!;\n            if (toolCall.id) acc.id = toolCall.id;\n            if (toolCall.function?.name) acc.name = toolCall.function.name;\n            if (toolCall.function?.arguments) acc.arguments += toolCall.function.arguments;\n          }\n        }\n\n        if (finish) {\n          rawFinishReason = finish;\n        }\n\n        if (chunk.usage) {\n          usage.input = chunk.usage.prompt_tokens ?? 0;\n          usage.output = chunk.usage.completion_tokens ?? 0;\n          usage.total = chunk.usage.total_tokens ?? 0;\n          const cached = chunk.usage.prompt_tokens_details?.cached_tokens;\n          if (cached !== undefined && cached > 0) {\n            usage.cachedTokens = cached;\n          }\n          const reasoning = chunk.usage.completion_tokens_details?.reasoning_tokens;\n          if (reasoning !== undefined && reasoning > 0) {\n            usage.reasoningTokens = reasoning;\n          }\n        }\n      }\n\n      for (const acc of toolCallAccum.values()) {\n        // Skip accumulators that never received a function name — those\n        // are partial fragments the model started but never identified\n        // (e.g. arguments-only deltas with no originating `id`/`name`).\n        // Yielding them produces nameless tool-calls the agent runtime\n        // can't dispatch and would mis-attribute as a registered tool.\n        if (!acc.name) continue;\n\n        yield {\n          type: \"tool-call\",\n          id: acc.id,\n          name: acc.name,\n          input: safeJsonParse<Record<string, unknown>>(acc.arguments, {}),\n        };\n      }\n    } catch (thrown) {\n      const wrapped = wrapOpenAIError(thrown);\n\n      this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const finishReason = mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"Streaming call to chat.completions succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /** Map the neutral cache hint only for the direct OpenAI provider. */\n  private promptCacheBreakpoints(options: ModelCallOptions | undefined): number {\n    return this.provider === \"openai\" ? (options?.cacheControl?.breakpoints ?? 0) : 0;\n  }\n\n  /**\n   * Translate the neutral `responseSchema` option into OpenAI's\n   * `response_format` parameter. The mode decision (override handling,\n   * strict-compatibility downgrade to `json_object`) lives in\n   * `planStructuredOutput`, shared with the Responses adapter.\n   *\n   * Returns an empty spread when no schema was supplied, so the caller\n   * can unconditionally `...buildResponseFormat(...)` into the request.\n   */\n  private buildResponseFormat(responseSchema: Record<string, unknown> | undefined): {\n    response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"response_format\"];\n  } {\n    const plan = planStructuredOutput(responseSchema, this.config.responseFormat);\n\n    if (plan.mode === \"json_schema\") {\n      return {\n        response_format: {\n          type: \"json_schema\",\n          json_schema: { name: \"response\", schema: plan.schema, strict: true },\n        },\n      };\n    }\n\n    if (plan.mode === \"json_object\") {\n      return { response_format: { type: \"json_object\" } };\n    }\n\n    return {};\n  }\n\n  /**\n   * Normalize OpenAI's `usage` block (which may be absent on some responses\n   * or partials) into the neutral `Usage` shape. Missing usage collapses to\n   * zeros rather than propagating `undefined`, so downstream aggregation\n   * math stays safe.\n   *\n   * `cachedTokens` mirrors `prompt_tokens_details.cached_tokens` (the\n   * subset of the prompt served from OpenAI's automatic prompt cache);\n   * `reasoningTokens` mirrors `completion_tokens_details.reasoning_tokens`\n   * (the hidden reasoning channel on o-series / gpt-5 models, already\n   * counted within `output`). Both are emitted only when the provider\n   * reports a positive value, so non-reasoning / uncached calls keep the\n   * lean `{ input, output, total }` shape.\n   */\n  private extractUsage(raw: OpenAI.Completions.CompletionUsage | undefined): Usage {\n    if (!raw) {\n      return { input: 0, output: 0, total: 0 };\n    }\n\n    return buildUsage({\n      input: raw.prompt_tokens,\n      output: raw.completion_tokens,\n      total: raw.total_tokens,\n      cachedTokens: raw.prompt_tokens_details?.cached_tokens,\n      reasoningTokens: raw.completion_tokens_details?.reasoning_tokens,\n    });\n  }\n\n  /**\n   * Translate the neutral `ModelCallOptions.reasoning` hint into OpenAI's\n   * `reasoning_effort` request param. Only `effort` maps — OpenAI's Chat\n   * Completions API exposes a discrete effort knob, not a token budget,\n   * so `reasoning.maxTokens` (the Anthropic extended-thinking cap) has no\n   * wire equivalent here and is silently ignored.\n   *\n   * The neutral `ReasoningEffort` (`\"low\" | \"medium\" | \"high\" | \"none\"`)\n   * is a subset of OpenAI's accepted values, so it forwards verbatim —\n   * `\"none\"` included. `\"none\"` is load-bearing: gpt-5 / o-series models\n   * **reject function tools** on Chat Completions while reasoning is\n   * active, and the endpoint accepts tools only when `reasoning_effort`\n   * is `\"none\"` (the alternative is the Responses API). This is why the\n   * param is EMITTED for `\"none\"` rather than omitted — omitting it\n   * leaves the model reasoning server-side by default, so tools would\n   * still be rejected.\n   *\n   * No-ops when the model is not reasoning-capable\n   * (`capabilities.reasoning` is false — e.g. `gpt-4o`, which 400s on any\n   * `reasoning_effort`).\n   *\n   * When the caller gave no explicit `effort`: defaults to `\"none\"` IF\n   * `hasTools` is true, else omits the param (provider default — the\n   * pre-existing behavior for a tool-less call). The default exists\n   * because, on a reasoning-capable model, there is no working alternative\n   * to `\"none\"` when tools are attached — omitting `reasoning_effort`\n   * leaves reasoning on server-side, which Chat Completions rejects tools\n   * for (empty replies on some model generations, a hard 400 — \"Function\n   * tools with reasoning_effort are not supported ... in\n   * /v1/chat/completions\" — on newer ones). An explicit `effort` from the\n   * caller always wins over this default, in either direction.\n   *\n   * Returns an empty spread when nothing applies, so the caller can\n   * unconditionally `...buildReasoningParams(...)` into the request.\n   */\n  /**\n   * Output-length and sampling params, in current Chat Completions shapes only.\n   *\n   * - `max_completion_tokens`, never the legacy `max_tokens`: current\n   *   models (the gpt-5 family and later) answer `400 unsupported_parameter`\n   *   to `max_tokens`, and `max_completion_tokens` is accepted everywhere.\n   * - No `temperature` for a reasoning-capable model: those accept only the\n   *   default value and reject any other with a 400.\n   */\n  private buildSamplingParams(options: ModelCallOptions | undefined): {\n    max_completion_tokens?: number;\n    temperature?: number;\n  } {\n    const maxTokens = options?.maxTokens ?? this.config.maxTokens;\n    const temperature = options?.temperature ?? this.config.temperature;\n\n    return {\n      max_completion_tokens: maxTokens,\n      ...(!this.capabilities.reasoning && { temperature }),\n    };\n  }\n\n  private buildReasoningParams(\n    reasoning: ModelCallOptions[\"reasoning\"],\n    hasTools: boolean,\n  ): {\n    reasoning_effort?: OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"];\n  } {\n    if (!this.capabilities.reasoning) {\n      return {};\n    }\n\n    if (!reasoning?.effort) {\n      return hasTools ? { reasoning_effort: \"none\" } : {};\n    }\n\n    return { reasoning_effort: toOpenAIReasoningEffort(reasoning.effort) };\n  }\n\n  /**\n   * Reshape OpenAI's `tool_calls` array into the neutral\n   * `ModelToolCallRequest[]`. The raw `arguments` field is a JSON string\n   * per OpenAI's protocol — we parse it defensively via `safeJsonParse` so\n   * malformed or empty arguments yield an empty object instead of crashing\n   * the trip. Returns `undefined` when no tools were requested so callers\n   * can branch on presence.\n   */\n  private extractToolCalls(\n    rawToolCalls: OpenAI.Chat.Completions.ChatCompletionMessageToolCall[] | undefined,\n  ): ModelToolCallRequest[] | undefined {\n    if (!rawToolCalls || rawToolCalls.length === 0) {\n      return undefined;\n    }\n\n    return rawToolCalls.map((toolCall) => ({\n      id: toolCall.id,\n      name: (toolCall as any).function.name,\n      input: safeJsonParse<Record<string, unknown>>((toolCall as any).function.arguments, {}),\n    }));\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;AA+BA,MAAM,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCnB,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,QAAgB,QAA2B,WAAmB,UAAU;gBAFzD;EAMhC,IAAI,OAAO,QAAQ,aACjB,MAAM,IAAI,oBACR,4JACA,EAAE,SAAS;GAAE;GAAU,OAAO,OAAO;GAAM,KAAK,OAAO;EAAI,EAAE,CAC/D;EAGF,KAAK,SAAS;EACd,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB,sBAAsB,OAAO,cAAc;GACxF,QAAQ,OAAO,UAAU,sBAAsB,OAAO,IAAI;GAI1D,WAAW,OAAO,aAAa,yBAAyB,OAAO,IAAI;GAInE,eAAe,KAAK,aAAa;GAMjC,KAAK,OAAO,OAAO;GACnB,OAAO,OAAO,SAAS;EACzB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAK7F,KAAK,OAAO,MAAM,YAAY,WAAW,qCAAqC;GAC5E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,OAAO,KAAK,YAAY,OAC5C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,OAAO,cAAc,SAAS,KAAK;IACnC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,CAAC,UAAU,SAAS;EAE1B,IAAI,WAAW,QASb,MAAM,IAAI,cAAc,+CAA+C,EACrE,SAAS;GAAE,UAAU;GAAU,OAAO,KAAK;GAAM,SAAS,SAAS,QAAQ;EAAO,EACpF,CAAC;EAGH,MAAM,eAAe,gBAAgB,OAAO,aAAa;EACzD,MAAM,QAAQ,KAAK,aAAa,SAAS,KAAK;EAE9C,KAAK,OAAO,MAAM,YAAY,YAAY,sCAAsC;GAC9E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,OAAO,QAAQ,WAAW;GACnC;GACA;GACA,WAAW,KAAK,iBAAiB,OAAO,QAAQ,UAAU;EAC5D;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,+CAA+C;GACtF,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,IAAI;EAEJ,IAAI;GACF,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY,OAC1C;IACE,OAAO,KAAK;IACZ,UAAU,iBAAiB,UAAU,KAAK,uBAAuB,OAAO,CAAC;IACzE,GAAG,KAAK,oBAAoB,OAAO;IACnC,GAAG,uBAAuB,KAAK,UAAU,KAAK,MAAM;IACpD,OAAO,cAAc,SAAS,KAAK;IACnC,QAAQ;IACR,gBAAgB,EAAE,eAAe,KAAK;IACtC,GAAG,KAAK,oBAAoB,SAAS,cAAc;IACnD,GAAG,KAAK,qBAAqB,SAAS,WAAW,QAAQ,SAAS,OAAO,MAAM,CAAC;GAClF,GACA,SAAS,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,MACjD;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,IAAI,kBAA0B;EAC9B,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EACrD,MAAM,gCAAgB,IAAI,IAA6D;EAEvF,IAAI;GACF,WAAW,MAAM,SAAS,QAAsE;IAC9F,MAAM,QAAQ,MAAM,QAAQ,EAAE,EAAE;IAChC,MAAM,SAAS,MAAM,QAAQ,EAAE,EAAE;IAEjC,IAAI,OAAO,SACT,MAAM;KAAE,MAAM;KAAS,SAAS,MAAM;IAAQ;IAGhD,IAAI,OAAO,YACT,KAAK,MAAM,YAAY,MAAM,YAAY;KACvC,MAAM,MAAM,SAAS,SAAS;KAC9B,IAAI,CAAC,cAAc,IAAI,GAAG,GACxB,cAAc,IAAI,KAAK;MAAE,IAAI;MAAI,MAAM;MAAI,WAAW;KAAG,CAAC;KAE5D,MAAM,MAAM,cAAc,IAAI,GAAG;KACjC,IAAI,SAAS,IAAI,IAAI,KAAK,SAAS;KACnC,IAAI,SAAS,UAAU,MAAM,IAAI,OAAO,SAAS,SAAS;KAC1D,IAAI,SAAS,UAAU,WAAW,IAAI,aAAa,SAAS,SAAS;IACvE;IAGF,IAAI,QACF,kBAAkB;IAGpB,IAAI,MAAM,OAAO;KACf,MAAM,QAAQ,MAAM,MAAM,iBAAiB;KAC3C,MAAM,SAAS,MAAM,MAAM,qBAAqB;KAChD,MAAM,QAAQ,MAAM,MAAM,gBAAgB;KAC1C,MAAM,SAAS,MAAM,MAAM,uBAAuB;KAClD,IAAI,WAAW,UAAa,SAAS,GACnC,MAAM,eAAe;KAEvB,MAAM,YAAY,MAAM,MAAM,2BAA2B;KACzD,IAAI,cAAc,UAAa,YAAY,GACzC,MAAM,kBAAkB;IAE5B;GACF;GAEA,KAAK,MAAM,OAAO,cAAc,OAAO,GAAG;IAMxC,IAAI,CAAC,IAAI,MAAM;IAEf,MAAM;KACJ,MAAM;KACN,IAAI,IAAI;KACR,MAAM,IAAI;KACV,OAAO,cAAuC,IAAI,WAAW,CAAC,CAAC;IACjE;GACF;EACF,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;IACtD,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,eAAe,gBAAgB,eAAe;EAEpD,KAAK,OAAO,MAAM,YAAY,YAAY,gDAAgD;GACxF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;CAGA,AAAQ,uBAAuB,SAA+C;EAC5E,OAAO,KAAK,aAAa,WAAY,SAAS,cAAc,eAAe,IAAK;CAClF;;;;;;;;;;CAWA,AAAQ,oBAAoB,gBAE1B;EACA,MAAM,OAAO,qBAAqB,gBAAgB,KAAK,OAAO,cAAc;EAE5E,IAAI,KAAK,SAAS,eAChB,OAAO,EACL,iBAAiB;GACf,MAAM;GACN,aAAa;IAAE,MAAM;IAAY,QAAQ,KAAK;IAAQ,QAAQ;GAAK;EACrE,EACF;EAGF,IAAI,KAAK,SAAS,eAChB,OAAO,EAAE,iBAAiB,EAAE,MAAM,cAAc,EAAE;EAGpD,OAAO,CAAC;CACV;;;;;;;;;;;;;;;CAgBA,AAAQ,aAAa,KAA4D;EAC/E,IAAI,CAAC,KACH,OAAO;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAGzC,OAAO,WAAW;GAChB,OAAO,IAAI;GACX,QAAQ,IAAI;GACZ,OAAO,IAAI;GACX,cAAc,IAAI,uBAAuB;GACzC,iBAAiB,IAAI,2BAA2B;EAClD,CAAC;CACH;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA8CA,AAAQ,oBAAoB,SAG1B;EACA,MAAM,YAAY,SAAS,aAAa,KAAK,OAAO;EACpD,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EAExD,OAAO;GACL,uBAAuB;GACvB,GAAI,CAAC,KAAK,aAAa,aAAa,EAAE,YAAY;EACpD;CACF;CAEA,AAAQ,qBACN,WACA,UAGA;EACA,IAAI,CAAC,KAAK,aAAa,WACrB,OAAO,CAAC;EAGV,IAAI,CAAC,WAAW,QACd,OAAO,WAAW,EAAE,kBAAkB,OAAO,IAAI,CAAC;EAGpD,OAAO,EAAE,kBAAkB,wBAAwB,UAAU,MAAM,EAAE;CACvE;;;;;;;;;CAUA,AAAQ,iBACN,cACoC;EACpC,IAAI,CAAC,gBAAgB,aAAa,WAAW,GAC3C;EAGF,OAAO,aAAa,KAAK,cAAc;GACrC,IAAI,SAAS;GACb,MAAO,SAAiB,SAAS;GACjC,OAAO,cAAwC,SAAiB,SAAS,WAAW,CAAC,CAAC;EACxF,EAAE;CACJ;AACF"}