Back when not all LLMs support native tool calling, we use this tool as the in-context learning way to make the LLM generate more consistent tool calls. Modern LLMs, even local ones such as GPT-OSS, have been mid-trained and post-trained to generate accurate tool calls in conjunction with generating free-form text replies to the user (in addition to reasoning content), so the talk_to_user tool would actually goes against its original purpose by confusing the models.
Let's remove the internal talk_to_user tool.
Back when not all LLMs support native tool calling, we use this tool as the in-context learning way to make the LLM generate more consistent tool calls. Modern LLMs, even local ones such as GPT-OSS, have been mid-trained and post-trained to generate accurate tool calls in conjunction with generating free-form text replies to the user (in addition to reasoning content), so the talk_to_user tool would actually goes against its original purpose by confusing the models.
Let's remove the internal talk_to_user tool.