Closes MemTensor#2149 Adds enable_thinking configuration parameter for OpenAI-compatible providers (qwen, deepseek, minimax) to control whether the model produces <think> reasoning blocks before the actual response. Changes: - Added enable_thinking field (bool | None) to OpenAILLMConfig - None (default): preserve provider's default behavior - True: explicitly enable thinking mode - False: explicitly disable thinking mode (prevents JSON breaking) - OpenAILLM.generate / generate_stream now pass enable_thinking to API calls when configured, via enable_thinking body param - AzureLLM.generate / generate_stream also support the parameter (gated by getattr for backward compatibility with older configs) - Per-call override supported via kwargs (enable_thinking=True/False) - Added tests covering default, config-level, and kwarg-level enable_thinking behavior in test_enable_thinking.py This is a non-breaking change: when enable_thinking is unset, request bodies are identical to previous versions. Test: python3 -m py_compile src/memos/configs/llm.py Test: python3 -m py_compile src/memos/llms/openai.py Test: python3 -m py_compile tests/llms/test_enable_thinking.py

Uh oh!

Commits

Commits on Aug 5, 2026

feat: add enable_thinking support for OpenAI-compatible providers

style(test): ruff cleanups for enable_thinking tests

style(llm): ruff format openai.py

test(llm): align enable-thinking config coverage