<|tool_calls_section_begin|> token appears in `reasoning_delta` causing agent execution failures

[loveofcsdt]

Environment

• Model: kimi-k2-thinking

Description

When using the kimi-k2-thinking model with streaming enabled (stream=True), I've observed that special tool-calling tokens such as <|tool_calls_section_begin|> appear within the reasoning_delta stream. This causes parsing failures in standard agent frameworks that expect reasoning_content to contain only natural language reasoning text.

Expected Behavior

According to the official documentation:

• reasoning_content should contain the model's thinking/reasoning process
• Tool calls should be properly structured in the tool_calls field
• Special tokens like <|tool_calls_section_begin|> should be part of the internal token structure, not exposed in the reasoning stream

Actual Behavior

The reasoning_delta field contains raw special tokens:

stream log from our side

[lsw825]

Thanks a lot for the feedback. Could you please provide more details, including: (1) this problem occurred in Moonshot API or open-source engine such as vllm? (2) What's your original prompt message? We will reproduce and further debug base on the info above.

[M1sTaker]

same problem. i use sglang as the engine .

[lemonteeeeaa]

same problem,openrouter api ,langchain can not use to call tools

[acodercat]

same issue. kimi-k2 0905 instruct

[idoomblast]

same issue, vertex api endpoint