fix: Anthropic proxy timeout — native SSE streaming + hardened HTTP client

src/llm/anthropic.rs

@@ -7,5 +7,5 @@ pub mod tools;
 
 pub use auth::{AnthropicAuthPath, apply_auth_headers, detect_auth_path};
 pub use cache::{CacheRetention, get_cache_control, resolve_cache_retention};
-pub use params::build_anthropic_request;
+pub use params::{AnthropicRequest, anthropic_messages_url, build_anthropic_request};
 pub use tools::{from_claude_code_name, to_claude_code_name};

src/llm/anthropic/params.rs

@@ -17,6 +17,8 @@ pub struct AnthropicRequest {
     pub auth_path: AnthropicAuthPath,
     /// Original tool (name, description) pairs for reverse-mapping response tool calls.
     pub original_tools: Vec<(String, String)>,
+    /// The JSON body sent to the API, exposed for streaming variants to add `"stream": true`.
+    pub body: serde_json::Value,
 }
 
 /// Adaptive thinking is only available on 4.6-generation models.
@@ -44,6 +46,11 @@ fn messages_url(base_url: &str) -> String {
     }
 }
 
+/// The Anthropic messages endpoint URL, exposed for streaming request construction.
+pub fn anthropic_messages_url(base_url: &str) -> String {
+    messages_url(base_url)
+}
+
 /// Build a fully configured Anthropic API request from a CompletionRequest.
 ///
 /// `base_url` is the provider's configured base URL (e.g. `https://api.anthropic.com`
@@ -109,6 +116,7 @@ pub fn build_anthropic_request(
         builder,
         auth_path,
         original_tools,
+        body,
     }
 }
 

src/llm/manager.rs

@@ -50,7 +50,10 @@ impl LlmManager {
     /// Create a new LLM manager with the given configuration.
     pub async fn new(config: LlmConfig) -> Result<Self> {
         let http_client = reqwest::Client::builder()
-            .timeout(std::time::Duration::from_secs(120))
+            .timeout(std::time::Duration::from_secs(300))
+            .connect_timeout(std::time::Duration::from_secs(30))
+            .tcp_keepalive(std::time::Duration::from_secs(30))
+            .pool_idle_timeout(std::time::Duration::from_secs(90))
             .build()
             .with_context(|| "failed to build HTTP client")?;
 
@@ -94,7 +97,10 @@ impl LlmManager {
     /// Initialize with an instance directory (for use at construction time).
     pub async fn with_instance_dir(config: LlmConfig, instance_dir: PathBuf) -> Result<Self> {
         let http_client = reqwest::Client::builder()
-            .timeout(std::time::Duration::from_secs(120))
+            .timeout(std::time::Duration::from_secs(300))
+            .connect_timeout(std::time::Duration::from_secs(30))
+            .tcp_keepalive(std::time::Duration::from_secs(30))
+            .pool_idle_timeout(std::time::Duration::from_secs(90))
             .build()
             .with_context(|| "failed to build HTTP client")?;