SGLang uses the qwen3_coder tool-call parser for Qwen 3.5: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.5
In SGLang, Qwen3CoderDetector._convert_param_value does lightweight schema-aware coercion for parameter values. In particular, for bool arguments it lowercases the raw value before interpreting it: https://github.com/sgl-project/sglang/blob/c67b2870569a1a639459389b0355641074ea9a74/python/sglang/srt/function_call/qwen3_coder_detector.py#L144
Renderers’ Qwen 3.5 parser is stricter. If the model generates True / False for boolean parameters, Renderers marks the tool call as ToolCallParseStatus.INVALID_JSON, because json.loads("True") / json.loads("False") fails: https://github.com/PrimeIntellect-ai/renderers/blob/main/renderers/parsing.py#L89
For example, with a tool schema like:
{
"type": "object",
"properties": {
"path": {"type": "string"},
"max_depth": {"type": "integer"},
"include_files": {"type": "boolean"},
"show_size": {"type": "boolean"}
}
}and model output like:
<function=filesystem_server_get_directory_tree>
<parameter=path>/</parameter>
<parameter=max_depth>2</parameter>
<parameter=include_files>True</parameter>
<parameter=show_size>True</parameter>
</function>
SGLang's Qwen3CoderDetector finds a valid tool call, but Renderers reports INVALID_JSON.
From Renderers' perspective, what's the source of truth for the tool parsing behavior? Should Renderers have integration tests against the known SGLang/vLLM tool call parsers, or should downstream framework integrators handle these semantic differences themselves?
SGLang uses the qwen3_coder tool-call parser for Qwen 3.5: https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.5
In SGLang, Qwen3CoderDetector._convert_param_value does lightweight schema-aware coercion for parameter values. In particular, for bool arguments it lowercases the raw value before interpreting it: https://github.com/sgl-project/sglang/blob/c67b2870569a1a639459389b0355641074ea9a74/python/sglang/srt/function_call/qwen3_coder_detector.py#L144
Renderers’ Qwen 3.5 parser is stricter. If the model generates True / False for boolean parameters, Renderers marks the tool call as ToolCallParseStatus.INVALID_JSON, because json.loads("True") / json.loads("False") fails: https://github.com/PrimeIntellect-ai/renderers/blob/main/renderers/parsing.py#L89
For example, with a tool schema like:
{ "type": "object", "properties": { "path": {"type": "string"}, "max_depth": {"type": "integer"}, "include_files": {"type": "boolean"}, "show_size": {"type": "boolean"} } }and model output like:
SGLang's Qwen3CoderDetector finds a valid tool call, but Renderers reports INVALID_JSON.
From Renderers' perspective, what's the source of truth for the tool parsing behavior? Should Renderers have integration tests against the known SGLang/vLLM tool call parsers, or should downstream framework integrators handle these semantic differences themselves?