fix: force HTTP/1.1 for LLM API requests

src/provider.rs

@@ -8,6 +8,9 @@ use crate::utils::{
 ///
 /// Routes `"anthropic"` to [`AnthropicProvider`] (Messages API) and everything else to
 /// [`OpenAIProvider`] (OpenAI-compatible chat completions). Temperature is set to 0.3.
+///
+/// DeepSeek providers get an HTTP/1.1-only reqwest client to avoid HTTP/2 body-framing
+/// issues with CloudFront on long-running reasoning responses.
 pub fn create_provider(
     provider_name: &str,
     base_url: &str,
@@ -17,8 +20,13 @@ pub fn create_provider(
     if provider_name == "anthropic" {
         Box::new(AnthropicProvider::new(base_url, api_key, model_name).with_temperature(0.3))
     } else {
-        let client =
-            LLMClient::new_openai_compatible(base_url, api_key, model_name).with_temperature(0.3);
+        let is_deepseek = provider_name.to_lowercase().contains("deepseek");
+        let client = if is_deepseek {
+            LLMClient::new_openai_compatible_http1(base_url, api_key, model_name)
+        } else {
+            LLMClient::new_openai_compatible(base_url, api_key, model_name)
+        }
+        .with_temperature(0.3);
         Box::new(OpenAIProvider::new(client))
     }
 }

src/utils.rs

@@ -334,23 +334,48 @@ pub fn build_reqwest_client() -> reqwest::Client {
         .expect("failed to build reqwest client")
 }
 
+/// Build a reqwest client forced to HTTP/1.1.
+///
+/// Some providers (notably DeepSeek behind CloudFront) have HTTP/2 body-framing
+/// issues on long-running responses. Use this builder only for providers known
+/// to require HTTP/1.1.
+pub fn build_reqwest_client_http1() -> reqwest::Client {
+    reqwest::Client::builder()
+        .no_proxy()
+        .http1_only()
+        .build()
+        .expect("failed to build reqwest client")
+}
+
 impl LLMClient {
     /// Create a new client with default settings (temp=0.7, timeout=600s).
     pub fn new_openai_compatible(base_url: &str, api_key: &str, model: &str) -> Self {
-        // Ensure base_url ends with slash if needed, or handle path joining correctly
-        // For simplicity, we assume user gives enough of the path or we append standardized paths.
-        // However, OpenAI-compatible APIs often vary in suffix.
-        // We'll treat `base_url` as the full endpoint URL for completions for maximum flexibility,
-        // OR we can default to adding "/chat/completions" if the user provided base domain.
-        // Let's go with robust: User provides full endpoint or we construct.
-        // Actually, simplest usage: base_url is the helper.
+        Self::new_openai_compatible_with_http1(base_url, api_key, model, false)
+    }
+
+    /// Create a new client with HTTP/1.1 forced (for providers with HTTP/2 framing issues).
+    pub fn new_openai_compatible_http1(base_url: &str, api_key: &str, model: &str) -> Self {
+        Self::new_openai_compatible_with_http1(base_url, api_key, model, true)
+    }
+
+    fn new_openai_compatible_with_http1(
+        base_url: &str,
+        api_key: &str,
+        model: &str,
+        http1_only: bool,
+    ) -> Self {
+        let client = if http1_only {
+            build_reqwest_client_http1()
+        } else {
+            build_reqwest_client()
+        };
         Self {
             base_url: base_url.to_string(),
             api_key: api_key.to_string(),
             model: model.to_string(),
             temperature: 0.7,
             timeout: Duration::from_secs(600),
-            client: build_reqwest_client(),
+            client,
         }
     }