Method createMessage

src/api/providers/openrouter.ts:209–534

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

	}

	override async *createMessage(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata,
	): AsyncGenerator<ApiStreamChunk> {
		const model = await this.fetchModel()

		let { id: modelId, maxTokens, temperature, topP, reasoning } = model

		// Reset reasoning_details accumulator for this request
		this.currentReasoningDetails = []

		// OpenRouter sends reasoning tokens by default for Gemini 2.5 Pro models
		// even if you don't request them. This is not the default for
		// other providers (including Gemini), so we need to explicitly disable
		// them unless the user has explicitly configured reasoning.
		// Note: Gemini 3 models use reasoning_details format with thought signatures,
		// but we handle this via skip_thought_signature_validator injection below.
		if (
			(modelId === "google/gemini-2.5-pro-preview" || modelId === "google/gemini-2.5-pro") &&
			typeof reasoning === "undefined"
		) {
			reasoning = { exclude: true }
		}

		// Convert Anthropic messages to OpenAI format.
		// Pass normalization function for Mistral compatibility (requires 9-char alphanumeric IDs)
		const isMistral = modelId.toLowerCase().includes("mistral")
		let openAiMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
			{ role: "system", content: systemPrompt },
			...convertToOpenAiMessages(
				messages,
				isMistral ? { normalizeToolCallId: normalizeMistralToolCallId } : undefined,
			),
		]

		// DeepSeek highly recommends using user instead of system role.
		if (modelId.startsWith("deepseek/deepseek-r1") || modelId === "perplexity/sonar-reasoning") {
			openAiMessages = convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
		}

		// Process reasoning_details when switching models to Gemini.
		const isGemini = modelId.startsWith("google/gemini")

		// For Gemini models with native protocol:
		// 1. Sanitize messages to handle thought signature validation issues.
		//    This must happen BEFORE fake encrypted block injection to avoid injecting for
		//    tool calls that will be dropped due to missing/mismatched reasoning_details.
		// 2. Inject fake reasoning.encrypted block for tool calls without existing encrypted reasoning.
		//    This is required when switching from other models to Gemini to satisfy API validation.
		//    Per OpenRouter documentation (conversation with Toven, Nov 2025):
		//    - Create ONE reasoning_details entry per assistant message with tool calls
		//    - Set `id` to the FIRST tool call's ID from the tool_calls array
		//    - Set `data` to "skip_thought_signature_validator" to bypass signature validation
		//    - Set `index` to 0
		// See: https://github.com/cline/cline/issues/8214
		if (isGemini) {
			// Step 1: Sanitize messages - filter out tool calls with missing/mismatched reasoning_details