From 6a32c29a746a2e44de463de647f9f6661eb5086b Mon Sep 17 00:00:00 2001
From: Xuan-Son Nguyen <son@huggingface.co>
Date: Thu, 6 Aug 2026 15:07:22 +0200
Subject: [PATCH] server: fix empty response for /cors-proxy (#26656)

---
 tools/server/server-models.cpp        | 22 +++++++++++++++-----
 tools/server/tests/unit/test_proxy.py | 30 +++++++++++++++++++++++++++
 2 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
index 188a72a374..2fd9519c10 100644
--- a/tools/server/server-models.cpp
+++ b/tools/server/server-models.cpp
@@ -2079,9 +2079,8 @@ server_http_proxy::server_http_proxy(
         return has_next; // false if EOF or pipe broken
     };
 
-    // wire up the HTTP client
-    // note: do NOT capture `this` pointer, as it may be destroyed before the thread ends
-    httplib::ResponseHandler response_handler = [pipe, cli](const httplib::Response & response) {
+    // build the header message forwarded to the reader thread, stripping internal proxy headers
+    auto make_header_msg = [](const httplib::Response & response) {
         msg_t msg;
         msg.status = response.status;
         for (const auto & [key, value] : response.headers) {
@@ -2095,7 +2094,17 @@ server_http_proxy::server_http_proxy(
             }
             msg.headers[key] = value;
         }
-        return pipe->write(std::move(msg)); // send headers first
+        return msg;
+    };
+
+    // true once response_handler has already forwarded the headers
+    auto headers_sent = std::make_shared<std::atomic<bool>>(false);
+
+    // wire up the HTTP client
+    // note: do NOT capture `this` pointer, as it may be destroyed before the thread ends
+    httplib::ResponseHandler response_handler = [pipe, headers_sent, make_header_msg](const httplib::Response & response) {
+        headers_sent->store(true);
+        return pipe->write(make_header_msg(response)); // send headers first
     };
     httplib::ContentReceiverWithProgress content_receiver = [pipe](const char * data, size_t data_length, size_t, size_t) {
         // send data chunks
@@ -2169,13 +2178,16 @@ server_http_proxy::server_http_proxy(
 
     // start the proxy thread
     SRV_DBG("start proxy thread %s %s\n", req.method.c_str(), req.path.c_str());
-    this->thread = std::thread([cli, pipe, req]() {
+    this->thread = std::thread([cli, pipe, req, headers_sent, make_header_msg]() {
         auto result = cli->send(std::move(req));
         if (result.error() != httplib::Error::Success) {
             auto err_str = httplib::to_string(result.error());
             SRV_ERR("http client error: %s\n", err_str.c_str());
             pipe->write({{}, 500, "", ""}); // header
             pipe->write({{}, 0, "proxy error: " + err_str, ""}); // body
+        } else if (!headers_sent->load()) {
+            // httplib skips response_handler for bodyless statuses like 204, send headers here instead
+            pipe->write(make_header_msg(*result));
         }
         pipe->close_write(); // signal EOF to reader
         SRV_DBG("%s", "client request thread ended\n");
diff --git a/tools/server/tests/unit/test_proxy.py b/tools/server/tests/unit/test_proxy.py
index 0fed536e59..cb439b7a45 100644
--- a/tools/server/tests/unit/test_proxy.py
+++ b/tools/server/tests/unit/test_proxy.py
@@ -1,5 +1,7 @@
 import pytest
 from utils import *
+import threading
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 
 server = ServerPreset.tinyllama2()
 
@@ -39,3 +41,31 @@ def test_mcp_proxy_custom_port():
     res = server.make_request("GET", f"/cors-proxy?url=http://{server.server_host}:{server.server_port}/models")
     assert res.status_code == 200
     assert "data" in res.body
+
+
+def test_mcp_proxy_no_content():
+    # note: see issue #26598
+    class NoContentHandler(BaseHTTPRequestHandler):
+        def do_POST(self):
+            self.send_response(204)
+            self.end_headers()
+
+        def log_message(self, format, *args):
+            pass
+
+    target = ThreadingHTTPServer(("127.0.0.1", 0), NoContentHandler)
+    target_thread = threading.Thread(target=target.serve_forever, daemon=True)
+    target_thread.start()
+
+    try:
+        global server
+        server.ui_mcp_proxy = True
+        server.start()
+
+        res = server.make_request("POST", f"/cors-proxy?url=http://127.0.0.1:{target.server_port}/", data={})
+        assert res.status_code == 204
+        assert res.body in (None, b"", "")
+    finally:
+        target.shutdown()
+        target.server_close()
+