Skip to content

Maybe I found a bug in cache that caused <tool_call> in <think> in model qwen 3.5 4b and 9b and omnicoder 9b #20614

@x4080

Description

@x4080

Here's the output after some agentic iteration, so it wont output correct tool calling instead it outputs xml style tool_call

<think>
.....
<tool_call> <function=execute_bash_command> <parameter=command> node --check export.js </parameter> </function> </tool_call> </think>

After some frustrating time looking for solution, I thought maybe its the cache ? Since I reset the messages sent to the LLM if it call certain tool_name, and thats the answer

After I add just --no-cache_prompt, it works fine - but very slowly, understandable since so cache exists

Maybe my resetting the messages data, still have some similarity but different with the cached messages or something I'm not sure

9b is q6, 4b is q8

4b is less frequent (thats why I tried using 9b)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions