Method createMessage

src/api/providers/bedrock.ts:387–799

(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata & {
			thinking?: {
				enabled: boolean
				maxTokens?: number
				maxThinkingTokens?: number
			}
		},
	)

	}

	override async *createMessage(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata & {
			thinking?: {
				enabled: boolean
				maxTokens?: number
				maxThinkingTokens?: number
			}
		},
	): ApiStream {
		const modelConfig = this.getModel()
		const usePromptCache = Boolean(
			(this.options.awsUsePromptCache ?? true) && this.supportsAwsPromptCache(modelConfig),
		)

		const conversationId =
			messages.length > 0
				? `conv_${messages[0].role}_${
						typeof messages[0].content === "string"
							? messages[0].content.substring(0, 20)
							: "complex_content"
					}`
				: "default_conversation"

		const formatted = this.convertToBedrockConverseMessages(
			messages,
			systemPrompt,
			usePromptCache,
			modelConfig.info,
			conversationId,
		)

		let additionalModelRequestFields: BedrockAdditionalModelFields | undefined
		let thinkingEnabled = false

		// Detect models that require the adaptive-thinking API contract (Opus/Sonnet
		// 4.7 and 4.8). See isAdaptiveThinkingModel for details. The same guard is
		// reused in completePrompt so both request paths stay consistent.
		const baseModelId = this.parseBaseModelId(modelConfig.id)
		const isAdaptiveThinkingModel = this.isAdaptiveThinkingModel(modelConfig.id)

		// Determine if thinking should be enabled
		// metadata?.thinking?.enabled: Explicitly enabled through API metadata (direct request)
		// shouldUseReasoningBudget(): Enabled through user settings (enableReasoningEffort = true)
		const isThinkingExplicitlyEnabled = metadata?.thinking?.enabled
		const isThinkingEnabledBySettings =
			shouldUseReasoningBudget({ model: modelConfig.info, settings: this.options }) &&
			modelConfig.reasoning &&
			modelConfig.reasoningBudget

		if ((isThinkingExplicitlyEnabled || isThinkingEnabledBySettings) && modelConfig.info.supportsReasoningBudget) {
			thinkingEnabled = true
			if (isAdaptiveThinkingModel) {
				// Claude 4.7+ (incl. 4.8 and Fable 5) uses adaptive thinking with effort levels —
				// budget_tokens causes a 400 error.
				// display: "summarized" surfaces thinking content in Zoo Code UI.
				// effort "xhigh" remains the recommended level for agentic coding tasks