Method createMessage

src/api/providers/zoo-gateway.ts:177–277

(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata,
	)

	}

	override async *createMessage(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata,
	): ApiStream {
		this.ensureAuthenticated()

		const { id: modelId, info } = await this.fetchModel()

		const openAiMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
			{ role: "system", content: systemPrompt },
			...convertToOpenAiMessages(messages),
		]

		// Apply prompt caching for models that support it
		// Zoo Gateway serves the same models as Vercel AI Gateway, so caching support is identical
		if (VERCEL_AI_GATEWAY_PROMPT_CACHING_MODELS.has(modelId) && info.supportsPromptCache) {
			addCacheBreakpoints(systemPrompt, openAiMessages)
		}

		// Build request headers with enrichment metadata
		const requestHeaders: Record<string, string> = {}
		if (metadata?.taskId) {
			requestHeaders["X-Zoo-Task-ID"] = metadata.taskId
		}
		if (metadata?.mode) {
			requestHeaders["X-Zoo-Mode"] = metadata.mode
		}

		const body: OpenAI.Chat.ChatCompletionCreateParams = {
			model: modelId,
			messages: openAiMessages,
			temperature: this.supportsTemperature(modelId)
				? (this.options.modelTemperature ?? ZOO_GATEWAY_DEFAULT_TEMPERATURE)
				: undefined,
			max_completion_tokens: info.maxTokens,
			stream: true,
			stream_options: { include_usage: true },
			tools: this.convertToolsForOpenAI(metadata?.tools),
			tool_choice: metadata?.tool_choice,
			parallel_tool_calls: metadata?.parallelToolCalls ?? true,
		}

		try {
			const completion = await this.client.chat.completions.create(body, {
				headers: requestHeaders,
			})

			for await (const chunk of completion) {
				// Once the gateway starts streaming the HTTP status is already 200, so it
				// reports upstream failures (e.g. provider rate limits) as an in-band error
				// chunk. Surface it so the user sees the real reason instead of an empty reply.
				if ("error" in chunk && chunk.error) {
					throw toGatewayStreamError(chunk.error)
				}

				const delta = chunk.choices[0]?.delta
				if (delta?.content) {