feat: add http_headers support to LlmProvider for custom upstream headers (fixes #592)

Allows users to specify custom HTTP request headers per LLM provider in
plano_config.yaml, useful for providers requiring non-standard auth or
telemetry headers:

  model_providers:
    - model: provider/model
      access_key: $API_KEY
      http_headers:
        X-Custom-Header: "custom-value"
        X-API-Version: "v1"

crates/common/src/configuration.rs

@@ -445,6 +445,7 @@ pub struct LlmProvider {
     pub base_url_path_prefix: Option<String>,
     pub internal: Option<bool>,
     pub passthrough_auth: Option<bool>,
+    pub http_headers: Option<HashMap<String, String>>,
 }
 
 pub trait IntoModels {
@@ -488,6 +489,7 @@ impl Default for LlmProvider {
             base_url_path_prefix: None,
             internal: None,
             passthrough_auth: None,
+            http_headers: None,
         }
     }
 }

crates/common/src/llm_providers.rs

@@ -277,6 +277,7 @@ mod tests {
             internal: None,
             stream: None,
             passthrough_auth: None,
+            http_headers: None,
         }
     }
 

crates/llm_gateway/src/stream_context.rs

@@ -232,6 +232,14 @@ impl StreamContext {
         Ok(())
     }
 
+    fn set_custom_provider_headers(&mut self) {
+        if let Some(http_headers) = self.llm_provider().http_headers.clone() {
+            for (key, value) in &http_headers {
+                self.set_http_request_header(key.as_str(), Some(value.as_str()));
+            }
+        }
+    }
+
     fn delete_content_length_header(&mut self) {
         // Remove the Content-Length header because further body manipulations in the gateway logic will invalidate it.
         // Server's generally throw away requests whose body length do not match the Content-Length header.
@@ -879,6 +887,7 @@ impl HttpContext for StreamContext {
                     self.send_server_error(error, Some(StatusCode::BAD_REQUEST));
                 }
             }
+            self.set_custom_provider_headers();
         }
 
         self.delete_content_length_header();