Add enable reasoning flag to llm request

[Neptoos]

I'm using vllm with a qwen3 model with reasoning. When using this with echokit client it also puts out the thinking block. To disable this behavior I changed the mod.rs file:

#[derive(Debug, Clone, serde::Serialize)]
pub struct StableLlmRequest {
    stream: bool,
    #[serde(rename = "chatId")]
    #[serde(skip_serializing_if = "String::is_empty")]
    chat_id: String,
    messages: Vec<llm::Content>,
    #[serde(skip_serializing_if = "String::is_empty")]
    model: String,
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    tools: Vec<llm::Tool>,
    #[serde(skip_serializing_if = "str::is_empty")]
    tool_choice: &'static str,
    
    // --- ADD THIS ---
    #[serde(skip_serializing_if = "Option::is_none")]
    pub chat_template_kwargs: Option<serde_json::Value>,
    // ----------------
}

pub async fn llm_stable<'p, I: IntoIterator<Item = C>, C: AsRef<llm::Content>>(
    llm_url: &str,
    token: &str,
    model: &str,
    chat_id: Option<String>,
    prompts: I,
    tools: Vec<llm::Tool>,
) -> anyhow::Result<StableLlmResponse> {
    let messages = prompts
        .into_iter()
        .map(|c| c.as_ref().clone())
        .collect::<Vec<_>>();
    let mut response_builder = reqwest::Client::new().post(llm_url);
    if !token.is_empty() {
        response_builder = response_builder.bearer_auth(token);
    };
    let tool_choice = if tools.is_empty() { "" } else { "auto" };
    
    let request = StableLlmRequest {
        stream: true,
        chat_id: chat_id.unwrap_or_default(),
        messages,
        model: model.to_string(),
        tools,
        tool_choice,
        
        // --- ADD THIS ---
        chat_template_kwargs: Some(serde_json::json!({
            "enable_thinking": false
        })),
        // ----------------
    };
    
    log::debug!(
        "#### send to llm:\n{}\n#####",
        serde_json::to_string_pretty(&request)?
    );

But it would be cool to have this as a flag in the config.toml

[L-jasmine]

In the recent commit, we've added a new feature that addresses this issue.

ad0f876

[Neptoos]

Pulled the new image but it kinda gets ignored even though it looks like the config has been recognized:

config:

[llm]
llm_chat_url = "http://vllm/v1/chat/completions"
api_key = "vllm"
model = "Qwen/Qwen3-4B-AWQ"
history = 5
extra = { "enable_thinking" = false }

server log:

config: Stable {
            llm: LLMConfig {
                llm_chat_url: "http://vllm/v1/chat/completions",
                api_key: Some(
                    "vllm",
                ),
                model: "Qwen/Qwen3-4B-AWQ",
                prompts_url: "",
                sys_prompts: [
                    Content {
                        role: System,
                        message: "You are world class Voice Agent",
                        tool_calls: None,
                        tool_call_id: None,
                    },
                ],
                dynamic_prompts: [],
                history: 5,
                mcp_server: [],
                extra: Some(
                    Object {
                        "enable_thinking": Bool(false),
                    },
                ),
            },

[Neptoos]

Looks like the culprit is in the src/ai/mod.rs:

 pub fn create_from_config(
        config: &crate::config::AIConfig,
        tools: ToolSet<McpToolAdapter>,
    ) -> Self {
        match config {
            crate::config::AIConfig::Stable { llm, .. } => {
                let mut session = ChatSession::new(
                    llm.llm_chat_url.clone(),
                    llm.api_key.clone().unwrap_or_default(),
                    llm.model.clone(),
                    None,  // <-- should be llm.extra.clone(), 
                    llm.history,
                    tools,
                );

                session.system_prompts = llm.sys_prompts.clone();
                session.messages = llm.dynamic_prompts.clone();

                session
            }