Tool calling serializes boolean arguments as strings with Qwen3.5 ("True" instead of true) #750

Description

FastFlowLM v1.0.6 serializes a tool argument defined as JSON Schema
boolean as a JSON string when using qwen3.5:9b.

The tool schema explicitly defines:

"accepted": {
"type": "boolean"
}

However, FastFlowLM returns:

{"accepted":"True"}

instead of:

{"accepted":true}

This causes OpenAI-compatible clients that validate tool arguments
against the supplied JSON Schema to reject the tool call.

Environment

FastFlowLM: v1.0.6

Model: qwen3.5:9b

Linux portable build: fastflowlm_1.0.6_linux.tar.gz

API: /v1/chat/completions

Minimal reproducer

curl -s http://127.0.0.1:8000/v1/chat/completions
-H 'Content-Type: application/json'
-d '{"model":"qwen3.5:9b","stream":false,"messages":[{"role":"user","content":"Call confirm_result with accepted set to true."}],"tools":[{"type":"function","function":{"name":"confirm_result","description":"Confirm the result","parameters":{"type":"object","properties":{"accepted":{"type":"boolean"}},"required":["accepted"]}}}],"tool_choice":{"type":"function","function":{"name":"confirm_result"}}}'

Actual result

FastFlowLM returns the tool arguments as:

"arguments":"{"accepted":"True"}"

Decoded:

{
"accepted": "True"
}

The value is a JSON string.

Expected result

The tool arguments should preserve the type specified by the JSON
Schema:

"arguments":"{"accepted":true}"

Decoded:

{
"accepted": true
}

The value should be a JSON boolean.

Additional information

The model's raw tool output contains:

<parameter=accepted>
True

FastFlowLM then converts this to:

{"accepted":"True"}

So it appears that the tool-call argument conversion/serialization step
does not coerce the generated value according to the parameter type
declared in the supplied tool JSON Schema.

This is reproducible with a minimal request and does not require a large
context or a multi-turn conversation.