Kimi-K2 via OpenRouter fails with tool call tokens inside thinking block

[graham33]

Description

When using Kimi-K2 via OpenRouter with a recent OpenCode (1.1.21), I see tool call tokens in thinking sections, and the tool stops before I get a useful response.

kimi-k2-session.txt

Plugins

No response

OpenCode version

1.1.21

Steps to reproduce

Given the following prompt:

what is the tech stack of this project?

I get this thinking output in the UI:

Thinking: I'll analyze the tech stack of this project by examining the configuration files and codebase structure. <|tool_calls_section_begin|> <|tool_call_begin|> functions.bash:0 <|tool_call_argument_begin|> {"command": "ls -la", "description": "List files in current directory"} <|tool_call_end|> <|tool_call_begin|> functions.glob:1 <|tool_call_argument_begin|> {"pattern": "/.json"} <|tool_call_end|> <|tool_call_begin|> functions.glob:2 <|tool_call_argument_begin|> {"pattern": "/.{js,ts,jsx,tsx}"} <|tool_call_end|> <|tool_call_begin|> functions.glob:3 <|tool_call_argument_begin|> {"pattern": "/*.{py,java,go,rs,c,cpp}"} <|tool_call_end|> <|tool_calls_section_end|

Screenshot and/or share link

Operating System

NixOS unstable

Terminal

kitty

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Expected thinking or redacted_thinking but found tool_use #8010: Expected thinking or redacted_thinking but found tool_use - Similar issue with Claude models where tool calls appear outside thinking blocks
• GLM 4.7 on Zai coding plan puts tool calls inside the thinking/reasoning tag. #6708: GLM 4.7 on Zai coding plan puts tool calls inside the thinking/reasoning tag - Same core problem with tool call tokens appearing in thinking blocks
• Amazon Bedrock Kimi2 Thinking output #5630: Amazon Bedrock Kimi2 Thinking output - Related issue with Kimi2 thinking model displaying tool call tokens in output
• Qwen3 as well as Kimi K2 keep stopping mid-chat #1522: Qwen3 as well as Kimi K2 keep stopping mid-chat - Related issue with Kimi K2 via OpenRouter

Feel free to ignore if none of these address your specific case.

[juanfadev]

Same with KimiK2 Thinking in nvidia provider:

Thinking: <|tool_calls_section_begin|> <|tool_call_begin|> functions.bash:1 <|tool_call_argument_begin|> {"command": "ls -la /home/user/repo", "description": "List directory"} <|tool_call_end|> <|tool_calls_section_end|>

[graham33]

Looking at this in a bit more detail, it looks like there is an open bug with the Kimi-K2 model itself: MoonshotAI/Kimi-K2#89 and vLLM implements a workaround for it: vllm-project/vllm#28543

This also seems similar to this issue with GLM 4.7: #6708, although there it seems more sporadic whereas with Kimi-K2 it happens a lot.

[qingchunnh]

I solved this issue following #6040. Hope this helps.

[pascalandy]

Provider returned error

All other models from OpenRouter are stable on my setup.
v1.1.39

[wwnbb]

I solved this issue following #6040. Hope this helps.

can you describe how you manage to fix the problem?