Fix Kimi-K2 tool-call parsing issues

[hksdpc255]

Kimi-K2 still appears to have issues with tool-call parsing.

Closes #17155.

[fernandaspets]

seems to work. i've been using the copy of this PR in ik_llama.cpp and it solved the problems I was having with Roo Code not working after they hardcoded openai compatible endpoints to use native tool calling instead of XML style

[hksdpc255]

@ggerganov Could we take a look for this?

[aldehir]

I'm looking at the tool calling done by Kimi K2, and I don't understand why it was implemented with the XML parsing.

The output looks more akin to GPT-OSS, with the function name separated from the arguments and the arguments provided in a single JSON object. I don't see any resemblance to the separate XML arguments that the other models exhibit (MiniMax M2, Qwen3-Coder).

What's the rationale for rolling this into the XML parsing?

[hksdpc255]

@aldehir You’re correct that Kimi-K2 is more like a GPT-OSS style tool call. However, the XML tool-call parser can still handle it with roughly 30 lines of code. (no one has implemented a generic GPT-OSS style parser yet)

Additionally, another purpose of this PR is to make the previously untested allow_toolcall_in_think feature fully tested and usable.

[hksdpc255]

I think it is ready for merge. @ggerganov @ngxson @CISC

[pwilkin]

Some questions, some nitpicks, but the important part is we need to support the singular form <|tool_call_section_begin|> in addition to <|tool_calls_section_begin|>

[pwilkin]

Looks fine to me, the tests mirror those in vLLM so I guess they should be pretty comprehensive. @bartowski1182 do you potentially have any box where you could run a Kimi K2-Thinking with some tool-calling tests?

[20jeka08]

Looks fine to me, the tests mirror those in vLLM so I guess they should be pretty comprehensive. @bartowski1182 do you potentially have any box where you could run a Kimi K2-Thinking with some tool-calling tests?

Hello, I've tried some tests of this PR on my local PC with 512gb RAM + 32gb VRAM (unsloth, kimi-k2-thinking Q3 quant), it looks good with Roo Code + Native tool calls, it can create, edit, perform terminal cmd. Also it works with Anthropic API without issues for me.

[pwilkin]

Merged into master.