common : preserve original Gemma 4 tool responses even when JSON-like

[kiwixz]

Overview

Before this PR, llama.cpp Gemma 4 specialized parser (#21418) would attempt to parse the string-typed tool responses as JSON; and convert it to an object if successful. It would then be minified by the template:

{%- if tool_response['response'] is mapping -%}
    {{- 'response:' + tool_response['name'] | default('unknown') + '{' -}}
    {%- for key, value in tool_response['response'] | dictsort -%}
        {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
        {%- if not loop.last %},{% endif -%}
    {%- endfor -%}
    {{- '}' -}}
{%- else -%}
    {{- 'response:' + tool_response['name'] | default('unknown') + '{value:' + format_argument(tool_response['response'], escape_keys=False) + '}' -}}
{%- endif -%}

This means that anything that looks like JSON would get unexpectedly altered.
I encountered this with Gemma 4 failing to edit a JSON file; it was getting confused as the inconsistent line numbers and indentations between what it read and what it attempted to write/edit.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[aldehir]

Which coding agent are you using? This might be the best thing to do, but it does have negative implications for clients that send structured tool output.

[kiwixz]

I'm developing my own thing. I found that text works better than structured things at least for simple tools.
The only difference for structured output would be keeping the quotes around JSON properties and wrapping the response in a {value: } object.

It's too bad that OpenAI API doesn't support just passing a JSON object directly in there; would be much more elegant than forcing it to a string and then guessing what it was...

[aldehir]

Many read tools prepend the line number to each line, which would make this get handled as a string.

Here is my concern: the model may in-context learn to start producing JSON instead of its native structure. We see this happen in other models where we manipulate the input in a way it wasn't trained.