feat: fix agent prompt data flow, fix tools in openai endpoint, account for completion usage info

backend/src/inference.py

@@ -11,8 +11,10 @@ from tool import Tool, ToolSpec, ToolCall
 #from llama_cpp import Llama
 from openrouter import OpenRouter
 
+# TODO: Collect usage data
+
 TEST_MOCK = False
-VALID_ROLES = ['user', 'assistant', 'system', 'tool', 'reasoning']
+VALID_ROLES = ['user', 'assistant', 'system', 'tool', 'reasoning', 'usage']
 CONTEXT_WINDOW = 10_000
 
 # Background inference tasks are kept referenced here so they aren't GC'd mid-flight.
@@ -60,14 +62,24 @@ def format_openai(messages):
             formatted.append(msg)
             return msg
 
+    # For debugging
+    _stream_count = 0
+    _badrole_count = 0
+    _badroles = set()
+
     for msg in messages:
+        content = msg.content
+        if isinstance(content, Stream):
+            _stream_count += 1
+            content = content.content
+
         if msg.role == 'tool':
             last = last_assistant_message()
             tool_calls = last.get('tool_calls')
             if tool_calls == None:
                 tool_calls = []
                 last['tool_calls'] = tool_calls
-            tc = msg.content
+            tc = content
             if isinstance(tc, dict):
                 tc = ToolCall(
                     tool_name=tc['tool_name'],
@@ -94,10 +106,19 @@ def format_openai(messages):
         elif msg.role == 'reasoning':
             last = last_assistant_message()
             reasoning = last.get('reasoning', '')
-            last['reasoning'] = reasoning + msg.content
-        else:
-            fmsg = {'role': msg.role, 'content': msg.content}
+            last['reasoning'] = reasoning + content
+        elif msg.role in ['user', 'assistant', 'system']:
+            fmsg = {'role': msg.role, 'content': content}
             formatted.append(fmsg)
+        else:
+            _badrole_count += 1
+            _badroles.add(msg.role)
+
+    _warnings = []
+    if _stream_count: _warnings.append(f"{_stream_count} streams in message list")
+    if _badrole_count: _warnings.append(f"{_badrole_count} messages with bad roles {_badroles}")
+    if _warnings: print(f"[inference] WARNING: " + ", ".join(_warnings))
+
     return formatted
 
 
@@ -350,8 +371,14 @@ async def openrouter_chat_completion(app, messages, model, tools=[], opts={}):
                                     tool_call.tool_name += tc.function.name
                                 if tc.function.arguments:
                                     tool_call.arguments += tc.function.arguments
-                else:
-                    print("[openrouter] no event choices: " + repr(event))
+                if event.usage:
+                    usage = {}
+                    for key in ['completion_tokens', 'prompt_tokens', 'total_tokens', 'cost']:
+                        if hasattr(event.usage, key):
+                            value = getattr(event.usage, key)
+                            if value != None:
+                                usage[key] = value
+                    queue.put_nowait(Message('usage', usage))
 
             if content_stream: content_stream.close()
             if reasoning_stream: reasoning_stream.close()