fix(vllm-router): fix routing oddities early

src/tests/test_request_auth_headers.py

@@ -33,7 +33,7 @@ def _build_request(headers=None):
     state.callbacks = None
     state.request_stats_monitor.get_request_stats.return_value = {}
     state.engine_stats_scraper.get_engine_stats.return_value = {}
-    state.router.route_request.return_value = "http://whisper-engine"
+    state.router.route_request = AsyncMock(return_value="http://whisper-engine")
 
     request = MagicMock()
     request.headers = headers or {}

src/vllm_router/services/request_service/rewriter.py

@@ -19,6 +19,7 @@
 """
 
 import abc
+import json
 
 from vllm_router.log import init_logger
 from vllm_router.utils import SingletonABCMeta
@@ -70,6 +71,92 @@ def rewrite_request(self, request_body: str, model: str, endpoint: str) -> str:
         return request_body
 
 
+class MessagesRewriter(RequestRewriter):
+    """
+    A request rewriter for Anthropic Messages API and OpenAI Chat Completions API
+    requests that normalizes messages before forwarding to the backend.
+
+    Normalizations:
+    - Filters out messages with empty/null content (some backends reject them).
+    - For ``/v1/messages``, promotes ``role: "system"`` entries in the messages
+      array to the top-level ``system`` parameter (handles the ``mid-conversation-system``
+      beta format sent by e.g. Claude Code).
+    """
+
+    def rewrite_request(self, request_body: str, model: str, endpoint: str) -> str:
+        try:
+            body = json.loads(request_body)
+        except json.JSONDecodeError:
+            return request_body
+
+        messages = body.get("messages")
+        if not messages or not isinstance(messages, list):
+            return request_body
+
+        # Guard: skip messages with empty content (some backends reject them).
+        messages = [m for m in messages if _message_has_content(m)]
+
+        if not messages:
+            return request_body
+
+        # For Anthropic Messages API, also promote role: "system" to top-level system param.
+        if endpoint == "/v1/messages":
+            system_messages = [m for m in messages if m.get("role") == "system"]
+            if system_messages:
+                system_content = _join_system_content(system_messages)
+                body["messages"] = [m for m in messages if m.get("role") != "system"]
+                if body.get("system") is not None:
+                    existing = body["system"]
+                    if isinstance(existing, str):
+                        body["system"] = existing + "\n" + system_content
+                    elif isinstance(existing, list):
+                        body["system"].append({"type": "text", "text": system_content})
+                    else:
+                        body["system"] = system_content
+                else:
+                    body["system"] = system_content
+
+                logger.info(
+                    "Promoted %d system message(s) from messages array to top-level system param",
+                    len(system_messages),
+                )
+                return json.dumps(body)
+
+            body["messages"] = messages
+            return json.dumps(body)
+
+        # For chat completions, just apply the empty-content guard.
+        if endpoint in ("/v1/chat/completions", "/chat/completions"):
+            body["messages"] = messages
+            return json.dumps(body)
+
+        return request_body
+
+
+def _message_has_content(message: dict) -> bool:
+    content = message.get("content")
+    if content is None:
+        return False
+    if isinstance(content, str):
+        return content.strip() != ""
+    if isinstance(content, list):
+        return len(content) > 0
+    return bool(content)
+
+
+def _join_system_content(system_messages: list[dict]) -> str:
+    parts = []
+    for msg in system_messages:
+        content = msg.get("content")
+        if isinstance(content, str):
+            parts.append(content)
+        elif isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict) and block.get("type") == "text":
+                    parts.append(block.get("text", ""))
+    return "\n".join(parts)
+
+
 # Singleton instance
 _request_rewriter_instance = None
 
@@ -87,10 +174,12 @@ def initialize_request_rewriter(rewriter_type: str, **kwargs) -> RequestRewriter
     """
     global _request_rewriter_instance
 
-    # TODO: Implement different rewriter types
-    # For now, just use the NoopRequestRewriter
-    _request_rewriter_instance = NoopRequestRewriter()
-    logger.info(f"Initialized placeholder request rewriter (type: {rewriter_type})")
+    if rewriter_type == "messages":
+        _request_rewriter_instance = MessagesRewriter()
+        logger.info("Initialized MessagesRewriter")
+    else:
+        _request_rewriter_instance = NoopRequestRewriter()
+        logger.info(f"Initialized placeholder request rewriter (type: {rewriter_type})")
 
     return _request_rewriter_instance
 
@@ -111,9 +200,10 @@ def get_request_rewriter() -> RequestRewriter:
     Get the request rewriter singleton instance.
 
     Returns:
-        The request rewriter instance or NoopRequestRewriter if not initialized
+        The request rewriter instance or MessagesRewriter if not initialized
     """
     global _request_rewriter_instance
     if _request_rewriter_instance is None:
-        _request_rewriter_instance = NoopRequestRewriter()
+        _request_rewriter_instance = MessagesRewriter()
+        logger.info("Initialized default MessagesRewriter")
     return _request_rewriter_instance