@warlock.js/ai-openai 5.24.0 → 5.26.0

package/esm/model.mjs

@@ -11,6 +11,23 @@ import { log } from "@warlock.js/logger";
 //#region ../ai-openai/src/model.ts
 const LOG_MODULE = "ai.openai";
 /**
+* The installed OpenAI SDK 7.23.0 declares `reasoning_effort` as
+* `none | minimal | low | medium | high | xhigh | max` in
+* `openai/src/resources/shared.ts:367`. Its type notes that model support
+* varies but does not provide a model-family-specific accepted set, so map by
+* that union alone. This exhaustive record makes a future core level a
+* typecheck failure until its OpenAI clamp is consciously selected.
+*/
+const EFFORT_TO_OPENAI_EFFORT = {
+	none: "none",
+	minimal: "minimal",
+	low: "low",
+	medium: "medium",
+	high: "high",
+	xhigh: "xhigh",
+	max: "max"
+};
+/**
 * 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
@@ -72,7 +89,7 @@ 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 +111,9 @@ 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),
+				...this.buildPromptCacheParams(),
 				tools: toOpenAITools(options?.tools),
 				...this.buildResponseFormat(options?.responseSchema),
 				...this.buildReasoningParams(options?.reasoning, Boolean(options?.tools?.length))
@@ -145,9 +162,9 @@ 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),
+				...this.buildPromptCacheParams(),
 				tools: toOpenAITools(options?.tools),
 				stream: true,
 				stream_options: { include_usage: true },
@@ -228,6 +245,18 @@ var OpenAIModel = class {
 			usage
 		};
 	}
+	/** OpenAI-only request fields; wrappers use the same transport but not this wire extension. */
+	buildPromptCacheParams() {
+		if (this.provider !== "openai") return {};
+		return {
+			...this.config.promptCacheKey !== void 0 ? { prompt_cache_key: this.config.promptCacheKey } : {},
+			...this.config.promptCacheRetention !== void 0 ? { prompt_cache_retention: this.config.promptCacheRetention } : {}
+		};
+	}
+	/** 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.
@@ -377,10 +406,27 @@ 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: EFFORT_TO_OPENAI_EFFORT[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  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 ReasoningEffort,\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\ntype OpenAIReasoningEffort = NonNullable<\n  OpenAI.Chat.Completions.ChatCompletionCreateParams[\"reasoning_effort\"]\n>;\n\n/**\n * The installed OpenAI SDK 7.23.0 declares `reasoning_effort` as\n * `none | minimal | low | medium | high | xhigh | max` in\n * `openai/src/resources/shared.ts:367`. Its type notes that model support\n * varies but does not provide a model-family-specific accepted set, so map by\n * that union alone. This exhaustive record makes a future core level a\n * typecheck failure until its OpenAI clamp is consciously selected.\n */\nconst EFFORT_TO_OPENAI_EFFORT: Record<ReasoningEffort, OpenAIReasoningEffort> = {\n  none: \"none\",\n  minimal: \"minimal\",\n  low: \"low\",\n  medium: \"medium\",\n  high: \"high\",\n  xhigh: \"xhigh\",\n  max: \"max\",\n};\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 / 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          ...this.buildPromptCacheParams(),\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          ...this.buildPromptCacheParams(),\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  /** OpenAI-only request fields; wrappers use the same transport but not this wire extension. */\n  private buildPromptCacheParams(): Pick<\n    OpenAI.Chat.Completions.ChatCompletionCreateParams,\n    \"prompt_cache_key\" | \"prompt_cache_retention\"\n  > {\n    if (this.provider !== \"openai\") {\n      return {};\n    }\n\n    return {\n      ...(this.config.promptCacheKey !== undefined\n        ? { prompt_cache_key: this.config.promptCacheKey }\n        : {}),\n      ...(this.config.promptCacheRetention !== undefined\n        ? { prompt_cache_retention: this.config.promptCacheRetention }\n        : {}),\n    };\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.\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  /**\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: EFFORT_TO_OPENAI_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":";;;;;;;;;;;AAqBA,MAAM,aAAa;;;;;;;;;AAcnB,MAAM,0BAA0E;CAC9E,MAAM;CACN,SAAS;CACT,KAAK;CACL,QAAQ;CACR,MAAM;CACN,OAAO;CACP,KAAK;AACP;;;;;;;;;;AAWA,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;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,KAAK,uBAAuB;IAC/B,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,KAAK,uBAAuB;IAC/B,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,yBAGN;EACA,IAAI,KAAK,aAAa,UACpB,OAAO,CAAC;EAGV,OAAO;GACL,GAAI,KAAK,OAAO,mBAAmB,SAC/B,EAAE,kBAAkB,KAAK,OAAO,eAAe,IAC/C,CAAC;GACL,GAAI,KAAK,OAAO,yBAAyB,SACrC,EAAE,wBAAwB,KAAK,OAAO,qBAAqB,IAC3D,CAAC;EACP;CACF;;CAGA,AAAQ,uBAAuB,SAA+C;EAC5E,OAAO,KAAK,aAAa,WAAY,SAAS,cAAc,eAAe,IAAK;CAClF;;;;;;;;;;;;;;;;;;;;;;;;;;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;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,QAAQ;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"}

package/esm/utils/to-openai-messages.mjs

@@ -25,8 +25,8 @@ import { InvalidRequestError } from "@warlock.js/ai";
 *   ]},
 * ]);
 */
-function toOpenAIMessages(messages) {
-	return messages.map((m) => {
+function toOpenAIMessages(messages, cacheBreakpoints = 0) {
+	return addPromptCacheBreakpoints(messages.map((m) => {
 		if (m.role === "tool") return {
 			role: "tool",
 			content: stringifyContent(m.content),
@@ -52,7 +52,44 @@ function toOpenAIMessages(messages) {
 			role: m.role,
 			content: stringifyContent(m.content)
 		};
-	});
+	}), cacheBreakpoints);
+}
+/** Place explicit markers on the last requested user-content parts (maximum four). */
+function addPromptCacheBreakpoints(messages, requested) {
+	let remaining = Math.min(4, Math.max(0, Math.floor(requested)));
+	if (remaining === 0) return messages;
+	const marked = [...messages];
+	for (let index = marked.length - 1; index >= 0 && remaining > 0; index--) {
+		const message = marked[index];
+		if (message === void 0 || message.role !== "user") continue;
+		if (typeof message.content === "string") {
+			marked[index] = {
+				...message,
+				content: [{
+					type: "text",
+					text: message.content,
+					prompt_cache_breakpoint: { mode: "explicit" }
+				}]
+			};
+			remaining--;
+			continue;
+		}
+		const content = [...message.content];
+		for (let partIndex = content.length - 1; partIndex >= 0 && remaining > 0; partIndex--) {
+			const part = content[partIndex];
+			if (part === void 0) continue;
+			content[partIndex] = {
+				...part,
+				prompt_cache_breakpoint: { mode: "explicit" }
+			};
+			remaining--;
+		}
+		marked[index] = {
+			...message,
+			content
+		};
+	}
+	return marked;
 }
 /**
 * Multipart content is only meaningful on user messages — for any other

package/esm/utils/to-openai-messages.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"to-openai-messages.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/to-openai-messages.ts"],"sourcesContent":["import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral Message[] to OpenAI's chat message shape.\n * Handles the `tool` role (requires `tool_call_id`) and assistant messages\n * that carry `toolCalls` from a prior model response.\n *\n * Multipart `content` (a `ContentPart[]`) is mapped into OpenAI's user-message\n * content-parts shape: text becomes `{ type: \"text\", text }`, images become\n * `{ type: \"image_url\", image_url: { url } }` — with base64 sources rendered\n * as `data:` URLs inline.\n *\n * @example\n * const openaiMessages = toOpenAIMessages([\n *   { role: \"user\", content: \"Hi\" },\n *   { role: \"tool\", toolCallId: \"call_1\", content: '{\"ok\":true}' },\n * ]);\n *\n * @example\n * toOpenAIMessages([\n *   { role: \"user\", content: [\n *     { type: \"text\", text: \"What is this?\" },\n *     { type: \"image\", source: { url: \"https://example.com/cat.jpg\" } },\n *   ]},\n * ]);\n */\nexport function toOpenAIMessages(\n  messages: Message[],\n): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {\n  return messages.map((m) => {\n    if (m.role === \"tool\") {\n      return {\n        role: \"tool\",\n        content: stringifyContent(m.content),\n        tool_call_id: m.toolCallId ?? \"\",\n      };\n    }\n    if (m.role === \"assistant\" && m.toolCalls && m.toolCalls.length > 0) {\n      return {\n        role: \"assistant\",\n        content: stringifyContent(m.content),\n        tool_calls: m.toolCalls.map((tc) => ({\n          id: tc.id,\n          type: \"function\" as const,\n          function: { name: tc.name, arguments: JSON.stringify(tc.input ?? {}) },\n        })),\n      };\n    }\n\n    if (m.role === \"user\" && Array.isArray(m.content)) {\n      return {\n        role: \"user\",\n        content: m.content.map(toOpenAIContentPart),\n      };\n    }\n\n    return { role: m.role, content: stringifyContent(m.content) } as\n      | OpenAI.Chat.Completions.ChatCompletionUserMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionSystemMessageParam\n      | OpenAI.Chat.Completions.ChatCompletionAssistantMessageParam;\n  });\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role (system / assistant text / tool), collapse a `ContentPart[]` to\n * its concatenated text so OpenAI's wire format stays valid. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to an OpenAI chat content part — one\n * branch per modality, each to its real wire shape:\n *\n * - `text` → `{ type: \"text\" }`.\n * - `image` → `{ type: \"image_url\" }` (remote URL, or a `data:` URL for\n *   inlined base64 bytes).\n * - `pdf` → `{ type: \"file\", file: { file_data } }` (OpenAI document\n *   input; base64 only — there is no remote-URL file source).\n * - `audio` → `{ type: \"input_audio\", input_audio: { data, format } }`\n *   (base64 only; `wav` / `mp3` are the only formats OpenAI accepts).\n *\n * PDF and audio reach this point ONLY when the model declared the\n * matching capability (`openai.model({ name, pdf: true })` /\n * `{ audio: true }`) — the agent's modality gate throws upfront\n * otherwise, so capability and behavior stay in lockstep. A remote-URL\n * pdf/audio source raises a typed `InvalidRequestError` here rather\n * than a downstream provider fault.\n */\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (part.type === \"image\") {\n    const url =\n      \"url\" in part.source\n        ? part.source.url\n        : `data:${part.source.mediaType};base64,${part.source.base64}`;\n\n    return { type: \"image_url\", image_url: { url } };\n  }\n\n  if (part.type === \"pdf\") {\n    if (\"url\" in part.source) {\n      throw new InvalidRequestError(\n        \"OpenAI chat completions cannot fetch a remote-URL PDF; supply base64 document bytes instead.\",\n      );\n    }\n\n    return {\n      type: \"file\",\n      file: {\n        filename: \"document.pdf\",\n        file_data: `data:${part.source.mediaType};base64,${part.source.base64}`,\n      },\n    };\n  }\n\n  // Audio — the remaining `ContentPart` variant.\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"OpenAI chat completions cannot fetch remote-URL audio; supply base64 audio bytes instead.\",\n    );\n  }\n\n  return {\n    type: \"input_audio\",\n    input_audio: {\n      data: part.source.base64,\n      format: toOpenAIAudioFormat(part.source.mediaType),\n    },\n  };\n}\n\n/**\n * Narrow a neutral audio media type to the two formats OpenAI's\n * `input_audio` accepts (`wav` / `mp3`). An unsupported type raises a\n * typed `InvalidRequestError` up front rather than a provider 400.\n */\nfunction toOpenAIAudioFormat(mediaType: string): \"wav\" | \"mp3\" {\n  if (mediaType === \"audio/wav\" || mediaType === \"audio/x-wav\" || mediaType === \"audio/wave\") {\n    return \"wav\";\n  }\n\n  if (mediaType === \"audio/mp3\" || mediaType === \"audio/mpeg\" || mediaType === \"audio/mpga\") {\n    return \"mp3\";\n  }\n\n  throw new InvalidRequestError(\n    `OpenAI input_audio supports only \"wav\" and \"mp3\"; got \"${mediaType}\".`,\n  );\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,iBACd,UACsD;CACtD,OAAO,SAAS,KAAK,MAAM;EACzB,IAAI,EAAE,SAAS,QACb,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,cAAc,EAAE,cAAc;EAChC;EAEF,IAAI,EAAE,SAAS,eAAe,EAAE,aAAa,EAAE,UAAU,SAAS,GAChE,OAAO;GACL,MAAM;GACN,SAAS,iBAAiB,EAAE,OAAO;GACnC,YAAY,EAAE,UAAU,KAAK,QAAQ;IACnC,IAAI,GAAG;IACP,MAAM;IACN,UAAU;KAAE,MAAM,GAAG;KAAM,WAAW,KAAK,UAAU,GAAG,SAAS,CAAC,CAAC;IAAE;GACvE,EAAE;EACJ;EAGF,IAAI,EAAE,SAAS,UAAU,MAAM,QAAQ,EAAE,OAAO,GAC9C,OAAO;GACL,MAAM;GACN,SAAS,EAAE,QAAQ,IAAI,mBAAmB;EAC5C;EAGF,OAAO;GAAE,MAAM,EAAE;GAAM,SAAS,iBAAiB,EAAE,OAAO;EAAE;CAI9D,CAAC;AACH;;;;;;;AAQA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,KAAK,SAAS,SAMhB,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;CAGjD,IAAI,KAAK,SAAS,OAAO;EACvB,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,8FACF;EAGF,OAAO;GACL,MAAM;GACN,MAAM;IACJ,UAAU;IACV,WAAW,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO;GACjE;EACF;CACF;CAGA,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,2FACF;CAGF,OAAO;EACL,MAAM;EACN,aAAa;GACX,MAAM,KAAK,OAAO;GAClB,QAAQ,oBAAoB,KAAK,OAAO,SAAS;EACnD;CACF;AACF;;;;;;AAOA,SAAS,oBAAoB,WAAkC;CAC7D,IAAI,cAAc,eAAe,cAAc,iBAAiB,cAAc,cAC5E,OAAO;CAGT,IAAI,cAAc,eAAe,cAAc,gBAAgB,cAAc,cAC3E,OAAO;CAGT,MAAM,IAAI,oBACR,0DAA0D,UAAU,GACtE;AACF"}
+{"version":3,"file":"to-openai-messages.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/to-openai-messages.ts"],"sourcesContent":["import { InvalidRequestError, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type OpenAI from \"openai\";\n\n/**\n * Convert vendor-neutral Message[] to OpenAI's chat message shape.\n * Handles the `tool` role (requires `tool_call_id`) and assistant messages\n * that carry `toolCalls` from a prior model response.\n *\n * Multipart `content` (a `ContentPart[]`) is mapped into OpenAI's user-message\n * content-parts shape: text becomes `{ type: \"text\", text }`, images become\n * `{ type: \"image_url\", image_url: { url } }` — with base64 sources rendered\n * as `data:` URLs inline.\n *\n * @example\n * const openaiMessages = toOpenAIMessages([\n *   { role: \"user\", content: \"Hi\" },\n *   { role: \"tool\", toolCallId: \"call_1\", content: '{\"ok\":true}' },\n * ]);\n *\n * @example\n * toOpenAIMessages([\n *   { role: \"user\", content: [\n *     { type: \"text\", text: \"What is this?\" },\n *     { type: \"image\", source: { url: \"https://example.com/cat.jpg\" } },\n *   ]},\n * ]);\n */\nexport function toOpenAIMessages(\n  messages: Message[],\n  cacheBreakpoints: number = 0,\n): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {\n  const mapped = 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      } satisfies OpenAI.Chat.Completions.ChatCompletionToolMessageParam;\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      } satisfies OpenAI.Chat.Completions.ChatCompletionAssistantMessageParam;\n    }\n\n    if (m.role === \"user\" && Array.isArray(m.content)) {\n      return {\n        role: \"user\",\n        content: m.content.map(toOpenAIContentPart),\n      } satisfies OpenAI.Chat.Completions.ChatCompletionUserMessageParam;\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  return addPromptCacheBreakpoints(mapped, cacheBreakpoints);\n}\n\n/** Place explicit markers on the last requested user-content parts (maximum four). */\nfunction addPromptCacheBreakpoints(\n  messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[],\n  requested: number,\n): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {\n  let remaining = Math.min(4, Math.max(0, Math.floor(requested)));\n  if (remaining === 0) return messages;\n\n  const marked = [...messages];\n  for (let index = marked.length - 1; index >= 0 && remaining > 0; index--) {\n    const message = marked[index];\n    if (message === undefined || message.role !== \"user\") continue;\n\n    if (typeof message.content === \"string\") {\n      marked[index] = {\n        ...message,\n        content: [\n          { type: \"text\", text: message.content, prompt_cache_breakpoint: { mode: \"explicit\" } },\n        ],\n      };\n      remaining--;\n      continue;\n    }\n\n    const content = [...message.content];\n    for (let partIndex = content.length - 1; partIndex >= 0 && remaining > 0; partIndex--) {\n      const part = content[partIndex];\n      if (part === undefined) continue;\n      content[partIndex] = { ...part, prompt_cache_breakpoint: { mode: \"explicit\" } };\n      remaining--;\n    }\n    marked[index] = { ...message, content };\n  }\n\n  return marked;\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any other\n * role (system / assistant text / tool), collapse a `ContentPart[]` to\n * its concatenated text so OpenAI's wire format stays valid. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Map a resolved `ContentPart` to an OpenAI chat content part — one\n * branch per modality, each to its real wire shape:\n *\n * - `text` → `{ type: \"text\" }`.\n * - `image` → `{ type: \"image_url\" }` (remote URL, or a `data:` URL for\n *   inlined base64 bytes).\n * - `pdf` → `{ type: \"file\", file: { file_data } }` (OpenAI document\n *   input; base64 only — there is no remote-URL file source).\n * - `audio` → `{ type: \"input_audio\", input_audio: { data, format } }`\n *   (base64 only; `wav` / `mp3` are the only formats OpenAI accepts).\n *\n * PDF and audio reach this point ONLY when the model declared the\n * matching capability (`openai.model({ name, pdf: true })` /\n * `{ audio: true }`) — the agent's modality gate throws upfront\n * otherwise, so capability and behavior stay in lockstep. A remote-URL\n * pdf/audio source raises a typed `InvalidRequestError` here rather\n * than a downstream provider fault.\n */\nfunction toOpenAIContentPart(part: ContentPart): OpenAI.Chat.Completions.ChatCompletionContentPart {\n  if (part.type === \"text\") {\n    return { type: \"text\", text: part.text };\n  }\n\n  if (part.type === \"image\") {\n    const url =\n      \"url\" in part.source\n        ? part.source.url\n        : `data:${part.source.mediaType};base64,${part.source.base64}`;\n\n    return { type: \"image_url\", image_url: { url } };\n  }\n\n  if (part.type === \"pdf\") {\n    if (\"url\" in part.source) {\n      throw new InvalidRequestError(\n        \"OpenAI chat completions cannot fetch a remote-URL PDF; supply base64 document bytes instead.\",\n      );\n    }\n\n    return {\n      type: \"file\",\n      file: {\n        filename: \"document.pdf\",\n        file_data: `data:${part.source.mediaType};base64,${part.source.base64}`,\n      },\n    };\n  }\n\n  // Audio — the remaining `ContentPart` variant.\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      \"OpenAI chat completions cannot fetch remote-URL audio; supply base64 audio bytes instead.\",\n    );\n  }\n\n  return {\n    type: \"input_audio\",\n    input_audio: {\n      data: part.source.base64,\n      format: toOpenAIAudioFormat(part.source.mediaType),\n    },\n  };\n}\n\n/**\n * Narrow a neutral audio media type to the two formats OpenAI's\n * `input_audio` accepts (`wav` / `mp3`). An unsupported type raises a\n * typed `InvalidRequestError` up front rather than a provider 400.\n */\nfunction toOpenAIAudioFormat(mediaType: string): \"wav\" | \"mp3\" {\n  if (mediaType === \"audio/wav\" || mediaType === \"audio/x-wav\" || mediaType === \"audio/wave\") {\n    return \"wav\";\n  }\n\n  if (mediaType === \"audio/mp3\" || mediaType === \"audio/mpeg\" || mediaType === \"audio/mpga\") {\n    return \"mp3\";\n  }\n\n  throw new InvalidRequestError(\n    `OpenAI input_audio supports only \"wav\" and \"mp3\"; got \"${mediaType}\".`,\n  );\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA2BA,SAAgB,iBACd,UACA,mBAA2B,GAC2B;CAkCtD,OAAO,0BAjCQ,SAAS,KAAK,MAAM;EACjC,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,CAEsC,GAAG,gBAAgB;AAC3D;;AAGA,SAAS,0BACP,UACA,WACsD;CACtD,IAAI,YAAY,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,CAAC,CAAC;CAC9D,IAAI,cAAc,GAAG,OAAO;CAE5B,MAAM,SAAS,CAAC,GAAG,QAAQ;CAC3B,KAAK,IAAI,QAAQ,OAAO,SAAS,GAAG,SAAS,KAAK,YAAY,GAAG,SAAS;EACxE,MAAM,UAAU,OAAO;EACvB,IAAI,YAAY,UAAa,QAAQ,SAAS,QAAQ;EAEtD,IAAI,OAAO,QAAQ,YAAY,UAAU;GACvC,OAAO,SAAS;IACd,GAAG;IACH,SAAS,CACP;KAAE,MAAM;KAAQ,MAAM,QAAQ;KAAS,yBAAyB,EAAE,MAAM,WAAW;IAAE,CACvF;GACF;GACA;GACA;EACF;EAEA,MAAM,UAAU,CAAC,GAAG,QAAQ,OAAO;EACnC,KAAK,IAAI,YAAY,QAAQ,SAAS,GAAG,aAAa,KAAK,YAAY,GAAG,aAAa;GACrF,MAAM,OAAO,QAAQ;GACrB,IAAI,SAAS,QAAW;GACxB,QAAQ,aAAa;IAAE,GAAG;IAAM,yBAAyB,EAAE,MAAM,WAAW;GAAE;GAC9E;EACF;EACA,OAAO,SAAS;GAAE,GAAG;GAAS;EAAQ;CACxC;CAEA,OAAO;AACT;;;;;;;AAQA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,oBAAoB,MAAsE;CACjG,IAAI,KAAK,SAAS,QAChB,OAAO;EAAE,MAAM;EAAQ,MAAM,KAAK;CAAK;CAGzC,IAAI,KAAK,SAAS,SAMhB,OAAO;EAAE,MAAM;EAAa,WAAW,EAAE,KAJvC,SAAS,KAAK,SACV,KAAK,OAAO,MACZ,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO,SAEb;CAAE;CAGjD,IAAI,KAAK,SAAS,OAAO;EACvB,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,8FACF;EAGF,OAAO;GACL,MAAM;GACN,MAAM;IACJ,UAAU;IACV,WAAW,QAAQ,KAAK,OAAO,UAAU,UAAU,KAAK,OAAO;GACjE;EACF;CACF;CAGA,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,2FACF;CAGF,OAAO;EACL,MAAM;EACN,aAAa;GACX,MAAM,KAAK,OAAO;GAClB,QAAQ,oBAAoB,KAAK,OAAO,SAAS;EACnD;CACF;AACF;;;;;;AAOA,SAAS,oBAAoB,WAAkC;CAC7D,IAAI,cAAc,eAAe,cAAc,iBAAiB,cAAc,cAC5E,OAAO;CAGT,IAAI,cAAc,eAAe,cAAc,gBAAgB,cAAc,cAC3E,OAAO;CAGT,MAAM,IAAI,oBACR,0DAA0D,UAAU,GACtE;AACF"}

package/esm/utils/wrap-openai-error.mjs

@@ -75,6 +75,7 @@ function toShape(thrown) {
 		code: thrown.code,
 		message: thrown.message,
 		type: thrown.type,
+		param: thrown.param,
 		headers: thrown.headers,
 		name: thrown.name
 	};
@@ -85,6 +86,7 @@ function toShape(thrown) {
 			code: typeof raw.code === "string" ? raw.code : void 0,
 			message: typeof raw.message === "string" ? raw.message : void 0,
 			type: typeof raw.type === "string" ? raw.type : void 0,
+			param: typeof raw.param === "string" ? raw.param : void 0,
 			headers: typeof raw.headers === "object" && raw.headers !== null ? raw.headers : void 0,
 			name: typeof raw.name === "string" ? raw.name : void 0
 		};
@@ -114,19 +116,22 @@ function buildContext(thrown, shape) {
 	if (shape.status !== void 0) context.status = shape.status;
 	if (shape.code) context.code = shape.code;
 	if (shape.type) context.type = shape.type;
+	if (shape.param) context.param = shape.param;
 	const requestId = readRequestId(thrown);
 	if (requestId) context.requestId = requestId;
 	return context;
 }
 /**
-* OpenAI puts the request id on `APIError.request_id`. Extract
-* defensively — both camel and snake keys exist across SDK versions.
+* The request id OpenAI returned. Current SDKs expose it as
+* `APIError.requestID`; older ones and hand-rolled errors use
+* `request_id` / `requestId`. Extract defensively.
 */
 function readRequestId(thrown) {
 	if (typeof thrown !== "object" || thrown === null) return;
 	const raw = thrown;
 	if (typeof raw.request_id === "string") return raw.request_id;
 	if (typeof raw.requestId === "string") return raw.requestId;
+	if (typeof raw.requestID === "string") return raw.requestID;
 }
 /**
 * Parse the `Retry-After` response header (seconds per HTTP spec)

package/esm/utils/wrap-openai-error.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"wrap-openai-error.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/wrap-openai-error.ts"],"sourcesContent":["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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AAkDA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkB,SACpB,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,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,mBACzC,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,sBACjB,OAAO,IAAI,mBAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,uBACzC,OAAO,IAAI,uBAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BACjB,OAAO,IAAI,2BAA2B,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAG3E,IAAI,MAAM,SAAS,kBACjB,OAAO,IAAI,mBAAmB,SAAS;EACrC,OAAO;EACP;EACA,QAAQ;CACV,CAAC;CAGH,IAAI,OAAO,MAAM,WAAW,YAAY,MAAM,UAAU,OAAO,MAAM,SAAS,KAC5E,OAAO,IAAI,oBAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;;AAQA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkB,OAAO,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,kBAAkB,OAAO,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"}
+{"version":3,"file":"wrap-openai-error.mjs","names":[],"sources":["../../../../../../../ai-openai/src/utils/wrap-openai-error.ts"],"sourcesContent":["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  /** The request parameter the provider rejected, e.g. `max_tokens`. */\n  param?: 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      param: thrown.param,\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      param: typeof raw.param === \"string\" ? raw.param : 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  if (shape.param) {\n    context.param = shape.param;\n  }\n\n  const requestId = readRequestId(thrown);\n\n  if (requestId) {\n    context.requestId = requestId;\n  }\n\n  return context;\n}\n\n/**\n * The request id OpenAI returned. Current SDKs expose it as\n * `APIError.requestID`; older ones and hand-rolled errors use\n * `request_id` / `requestId`. Extract defensively.\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  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"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AAoDA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkB,SACpB,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,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,mBACzC,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,SAAS,sBACjB,OAAO,IAAI,mBAAmB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGnE,IAAI,MAAM,WAAW,OAAO,MAAM,SAAS,uBACzC,OAAO,IAAI,uBAAuB,SAAS;EACzC,OAAO;EACP;EACA,YAAY,gBAAgB,MAAM,OAAO;CAC3C,CAAC;CAGH,IAAI,MAAM,SAAS,2BACjB,OAAO,IAAI,2BAA2B,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAG3E,IAAI,MAAM,SAAS,kBACjB,OAAO,IAAI,mBAAmB,SAAS;EACrC,OAAO;EACP;EACA,QAAQ;CACV,CAAC;CAGH,IAAI,OAAO,MAAM,WAAW,YAAY,MAAM,UAAU,OAAO,MAAM,SAAS,KAC5E,OAAO,IAAI,oBAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;;AAQA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkB,OAAO,UAC3B,OAAO;EACL,QAAQ,OAAO;EACf,MAAM,OAAO;EACb,SAAS,OAAO;EAChB,MAAM,OAAO;EACb,OAAO,OAAO;EACd,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,OAAO,OAAO,IAAI,UAAU,WAAW,IAAI,QAAQ;GACnD,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,kBAAkB,OAAO,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,IAAI,MAAM,OACR,QAAQ,QAAQ,MAAM;CAGxB,MAAM,YAAY,cAAc,MAAM;CAEtC,IAAI,WACF,QAAQ,YAAY;CAGtB,OAAO;AACT;;;;;;AAOA,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;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"}

package/llms-full.txt

@@ -53,7 +53,7 @@ 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. |
+| `promptCaching` | `true` for the direct OpenAI provider. It supports automatic cache hits plus explicit content-part write breakpoints; OpenAI-compatible wrapper providers leave the OpenAI-only fields off. |
 | `pdf` | `false` by default (opt-in `.model({ pdf: true })`). OpenAI accepts PDF `file` parts only on specific models (the `gpt-4o` family) — honest off until you set it. |
 | `audio` | `false` by default (opt-in `.model({ audio: true })`). Only the `gpt-4o-audio-preview` family accepts `input_audio`. |
 
@@ -157,6 +157,14 @@ await model.complete(messages, { tools, reasoning: { effort: "high" } }); // you
 await model.complete(messages, { tools, reasoning: { effort: "none" } }); // same effect as the default, explicit
 ```
 
+The gpt-6 models (`gpt-6-astra` / `gpt-6-sol` / `gpt-6-luna`) are hard-400 cases, verified live on `gpt-6-sol`:
+
+```
+400 Function tools with reasoning_effort are not supported for gpt-6-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.
+```
+
+Tools work with reasoning off (the adapter's default when tools are present and no reasoning is set). Responses API support, planned for 5.27, lifts this.
+
 Calling **without** `tools` is unaffected — `reasoning_effort` is still omitted by default in that case (provider default reasoning, nothing to unblock). Trade-off of the `"none"` default: you lose reasoning on that call. For tool-heavy agent work (function calls + good replies rather than deep analysis) this is the right call — it's the difference between empty replies and working ones. When you need reasoning **and** tools together, use the Responses API (planned).
 
 ## Token usage — what's reported
@@ -171,7 +179,21 @@ Calling **without** `tools` is unaffected — `reasoning_effort` is still omitte
 
 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.
+### Explicit prompt caching (direct OpenAI only)
+
+Chat Completions exposes OpenAI-specific cache controls on current SDK types. Configure a stable cache bucket and (when needed) its maximum retention policy per model:
+
+```ts
+const model = openai.model({
+  name: "gpt-5.6",
+  promptCacheKey: "tenant:acme",
+  promptCacheRetention: "24h", // or "in_memory"
+});
+```
+
+`promptCacheKey` improves matching for related requests. `promptCacheRetention` is OpenAI's maximum policy (`"in_memory" | "24h"`). `cacheControl.breakpoints` maps to `prompt_cache_breakpoint: { mode: "explicit" }` on the final user content parts (up to OpenAI's four-request limit); use it when you need to pin a reusable prefix rather than rely only on OpenAI's implicit breakpoint.
+
+These fields are sent only when the SDK provider label is exactly `"openai"`. `DeepSeekSDK`, `GroqSDK`, `XaiSDK`, and `MistralSDK` label their wrapped endpoint differently, so their requests omit every OpenAI-only cache field even though they reuse this transport. Read-side `usage.cachedTokens` accounting remains available whenever the upstream reports it.
 
 ## Token counting
 

package/package.json

@@ -13,13 +13,13 @@
     "url": "https://github.com/warlockjs/ai-openai"
   },
   "dependencies": {
-    "openai": "^7.4.0",
-    "@warlock.js/logger": "5.24.0"
+    "openai": "^7.23.0",
+    "@warlock.js/logger": "5.26.0"
   },
   "peerDependencies": {
-    "@warlock.js/ai": "5.24.0"
+    "@warlock.js/ai": "5.26.0"
   },
-  "version": "5.24.0",
+  "version": "5.26.0",
   "main": "./cjs/index.cjs",
   "module": "./esm/index.mjs",
   "types": "./esm/index.d.mts",

package/skills/setup-openai/SKILL.md

@@ -45,7 +45,7 @@ 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. |
+| `promptCaching` | `true` for the direct OpenAI provider. It supports automatic cache hits plus explicit content-part write breakpoints; OpenAI-compatible wrapper providers leave the OpenAI-only fields off. |
 | `pdf` | `false` by default (opt-in `.model({ pdf: true })`). OpenAI accepts PDF `file` parts only on specific models (the `gpt-4o` family) — honest off until you set it. |
 | `audio` | `false` by default (opt-in `.model({ audio: true })`). Only the `gpt-4o-audio-preview` family accepts `input_audio`. |
 
@@ -149,6 +149,14 @@ await model.complete(messages, { tools, reasoning: { effort: "high" } }); // you
 await model.complete(messages, { tools, reasoning: { effort: "none" } }); // same effect as the default, explicit
 ```
 
+The gpt-6 models (`gpt-6-astra` / `gpt-6-sol` / `gpt-6-luna`) are hard-400 cases, verified live on `gpt-6-sol`:
+
+```
+400 Function tools with reasoning_effort are not supported for gpt-6-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.
+```
+
+Tools work with reasoning off (the adapter's default when tools are present and no reasoning is set). Responses API support, planned for 5.27, lifts this.
+
 Calling **without** `tools` is unaffected — `reasoning_effort` is still omitted by default in that case (provider default reasoning, nothing to unblock). Trade-off of the `"none"` default: you lose reasoning on that call. For tool-heavy agent work (function calls + good replies rather than deep analysis) this is the right call — it's the difference between empty replies and working ones. When you need reasoning **and** tools together, use the Responses API (planned).
 
 ## Token usage — what's reported
@@ -163,7 +171,21 @@ Calling **without** `tools` is unaffected — `reasoning_effort` is still omitte
 
 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.
+### Explicit prompt caching (direct OpenAI only)
+
+Chat Completions exposes OpenAI-specific cache controls on current SDK types. Configure a stable cache bucket and (when needed) its maximum retention policy per model:
+
+```ts
+const model = openai.model({
+  name: "gpt-5.6",
+  promptCacheKey: "tenant:acme",
+  promptCacheRetention: "24h", // or "in_memory"
+});
+```
+
+`promptCacheKey` improves matching for related requests. `promptCacheRetention` is OpenAI's maximum policy (`"in_memory" | "24h"`). `cacheControl.breakpoints` maps to `prompt_cache_breakpoint: { mode: "explicit" }` on the final user content parts (up to OpenAI's four-request limit); use it when you need to pin a reusable prefix rather than rely only on OpenAI's implicit breakpoint.
+
+These fields are sent only when the SDK provider label is exactly `"openai"`. `DeepSeekSDK`, `GroqSDK`, `XaiSDK`, and `MistralSDK` label their wrapped endpoint differently, so their requests omit every OpenAI-only cache field even though they reuse this transport. Read-side `usage.cachedTokens` accounting remains available whenever the upstream reports it.
 
 ## Token counting