wip: refactor tool processing in openai endpoint

backend/src/inference.py

@@ -45,7 +45,8 @@ def format_openai(messages):
     for msg in messages:
         if msg.role == 'tool':
             if not last_assistant_message:
-                raise Exception("tool message before any assistant message")
+                # Create one implicitly
+                last_assistant_message = {'role': 'assistant', 'content': ''}
             tool_calls = last_assistant_message.get('tool_calls')
             if tool_calls == None:
                 tool_calls = []
@@ -314,8 +315,21 @@ async def openrouter_chat_completion(app, messages, model, tools=[], opts={}, ag
     # TODO: Separate as InferenceRun class
     content_stream = None
     reasoning_stream = None
+
+    last_channel = None
+    def open_channel(channel):
+        nonlocal last_channel, content_stream, reasoning_stream
+        if channel != last_channel:
+            if content_stream: content_stream.close()
+            if reasoning_stream: reasoning_stream.close()
+            content_stream = None
+            reasoning_stream = None
+        last_channel = channel
+
+
     async def message_loop(name, value=None):
         nonlocal content_stream, reasoning_stream, agent_run
+        open_channel(name) # Hack: use the message as the channel name
         if name == 'output':
             if content_stream == None:
                 content_stream = Stream(app)