Method createMessage

src/api/providers/minimax.ts:80–237

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

	}

	async *createMessage(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata,
	): ApiStream {
		const cacheControl: CacheControlEphemeral = { type: "ephemeral" }
		const { id: modelId, info, maxTokens, temperature } = this.getModel()

		// MiniMax M2 models support prompt caching
		const supportsPromptCache = info.supportsPromptCache ?? false

		// Merge environment_details from messages that follow tool_result blocks
		// into the tool_result content. This preserves reasoning continuity for
		// thinking models by preventing user messages from interrupting the
		// reasoning context after tool use (similar to r1-format's mergeToolResultText).
		const processedMessages = mergeEnvironmentDetailsForMiniMax(messages)

		// Build the system blocks array
		const systemBlocks: Anthropic.Messages.TextBlockParam[] = [
			supportsPromptCache
				? { text: systemPrompt, type: "text", cache_control: cacheControl }
				: { text: systemPrompt, type: "text" },
		]

		// Prepare request parameters
		const requestParams: Anthropic.Messages.MessageCreateParams = {
			model: modelId,
			max_tokens: maxTokens ?? 16_384,
			temperature: temperature ?? 1.0,
			system: systemBlocks,
			messages: supportsPromptCache ? this.addCacheControl(processedMessages, cacheControl) : processedMessages,
			stream: true,
			tools: convertOpenAIToolsToAnthropic(metadata?.tools ?? []),
			tool_choice: convertOpenAIToolChoice(metadata?.tool_choice),
		}

		const stream = await this.client.messages.create(requestParams)

		let inputTokens = 0
		let outputTokens = 0
		let cacheWriteTokens = 0
		let cacheReadTokens = 0

		for await (const chunk of stream) {
			switch (chunk.type) {
				case "message_start": {
					// Tells us cache reads/writes/input/output.
					const {
						input_tokens = 0,
						output_tokens = 0,
						cache_creation_input_tokens,
						cache_read_input_tokens,
					} = chunk.message.usage

					yield {
						type: "usage",
						inputTokens: input_tokens,
						outputTokens: output_tokens,