It looks like the StructureFixingParser does not w...

# koog-agentic-framework

[Sanga]

It looks like the StructureFixingParser does not work when it comes to thinking models.

I simply connected nodeSaveToMemory after nodeCallLLM onAssistantMesage { true }

I manually even hacked around and gave my own save node implementation but since it does requestLLMStructured - The parsing fails in think thinking tokens /think validJson.

No matter how many retries we give the thinking prompt is going to be there. Any idea on how to implement this?

Koog ver: 0.6.2 Platform: JVM

I run qwen 8B and use OpenAI client (because its in a different machine)

Looks like there is some specific handling for Qwq models?

[Jakub Amanowicz]

Hey @Sanga, thanks for the report!

I created a ticket to address this - we will try to tackle this in the near future 🙂

The issue seems to be that when using thinking models via an OpenAI-compatible endpoint, the <think>...</think> block ends up embedded in the text content rather than being parsed out separately — so structured parsing chokes on it.

In the meantime, you could try to work around it by wrapping your executor and overriding execute in a way where you strip <think> block from the Message.Assistant messages

[Sanga]

Yep, I just did that and its fine.

[Sanga]

I wanted to report back here but got distracted. Thanks for your help