Method createStreamWithThinking

src/api/providers/zai.ts:75–132

* Creates a stream with explicit thinking control for GLM thinking-capable models.

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

	 * Creates a stream with explicit thinking control for GLM thinking-capable models.
	 */
	private createStreamWithThinking(
		systemPrompt: string,
		messages: Anthropic.Messages.MessageParam[],
		metadata?: ApiHandlerCreateMessageMetadata,
	) {
		const { id: model, info } = this.getModel()

		// Fall back to the model default when the resolved effort isn't supported by the model.
		const supported = info.supportsReasoningEffort
		const raw =
			this.options.enableReasoningEffort === false
				? undefined
				: (this.options.reasoningEffort ?? info.reasoningEffort)
		const effort =
			raw && raw !== "disable" && Array.isArray(supported) && !supported.includes(raw)
				? info.reasoningEffort
				: raw
		const reasoningEffort = effort && effort !== "disable" ? effort : undefined
		const useReasoning = reasoningEffort !== undefined

		const max_tokens =
			this.options.modelMaxTokens ||
			(getModelMaxOutputTokens({
				modelId: model,
				model: info,
				settings: this.options,
				format: "openai",
			}) ??
				undefined)

		const temperature = this.options.modelTemperature ?? this.defaultTemperature

		// Use Z.ai format to preserve reasoning_content and merge post-tool text into tool messages
		const convertedMessages = convertToZAiFormat(messages, { mergeToolResultText: true })

		const params: ZAiChatCompletionParams = {
			model,
			max_tokens,
			temperature,
			messages: [{ role: "system", content: systemPrompt }, ...convertedMessages],
			stream: true,
			stream_options: { include_usage: true },
			// Thinking is ON by default for these models, so explicitly disable it when needed.
			thinking: useReasoning ? { type: "enabled" } : { type: "disabled" },
			reasoning_effort: reasoningEffort,
			tools: this.convertToolsForOpenAI(metadata?.tools),
			tool_choice: metadata?.tool_choice,
			parallel_tool_calls: metadata?.parallelToolCalls ?? true,
		}

		try {
			return this.client.chat.completions.create(
				params as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming,
			)
		} catch (error) {
			throw handleOpenAIError(error, this.providerName)
		}
	}