Populate Codex usage metadata for token accounting (#2585)

backend/packages/harness/deerflow/models/openai_codex_provider.py

@@ -21,6 +21,7 @@ from langchain_core.callbacks import CallbackManagerForLLMRun
 from langchain_core.language_models.chat_models import BaseChatModel
 from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage, ToolMessage
 from langchain_core.outputs import ChatGeneration, ChatResult
+from langchain_openai.chat_models.base import _create_usage_metadata_responses
 
 from deerflow.models.credential_loader import CodexCliCredential, load_codex_cli_credential
 
@@ -346,6 +347,7 @@ class CodexChatModel(BaseChatModel):
                 )
 
         usage = response.get("usage", {})
+        usage_metadata = _create_usage_metadata_responses(usage) if usage else None
         additional_kwargs = {}
         if reasoning_content:
             additional_kwargs["reasoning_content"] = reasoning_content
@@ -355,6 +357,7 @@ class CodexChatModel(BaseChatModel):
             tool_calls=tool_calls if tool_calls else [],
             invalid_tool_calls=invalid_tool_calls,
             additional_kwargs=additional_kwargs,
+            usage_metadata=usage_metadata,
             response_metadata={
                 "model": response.get("model", self.model),
                 "usage": usage,

backend/tests/test_codex_provider.py

@@ -82,6 +82,36 @@ def test_parse_response_text_content():
     assert result.generations[0].message.content == "Hello world"
 
 
+def test_parse_response_populates_usage_metadata():
+    model = _make_model()
+    response = {
+        "output": [
+            {
+                "type": "message",
+                "content": [{"type": "output_text", "text": "Hello world"}],
+            }
+        ],
+        "usage": {
+            "input_tokens": 10,
+            "output_tokens": 5,
+            "total_tokens": 15,
+            "input_tokens_details": {"cached_tokens": 3},
+            "output_tokens_details": {"reasoning_tokens": 2},
+        },
+        "model": "gpt-5.4",
+    }
+
+    result = model._parse_response(response)
+
+    assert result.generations[0].message.usage_metadata == {
+        "input_tokens": 10,
+        "output_tokens": 5,
+        "total_tokens": 15,
+        "input_token_details": {"cache_read": 3},
+        "output_token_details": {"reasoning": 2},
+    }
+
+
 def test_parse_response_reasoning_content():
     model = _make_model()
     response = {

backend/tests/test_token_usage_middleware.py

@@ -0,0 +1,32 @@
+from unittest.mock import MagicMock, patch
+
+from langchain_core.messages import AIMessage
+
+from deerflow.agents.middlewares.token_usage_middleware import TokenUsageMiddleware
+
+
+def test_after_model_logs_usage_metadata_counts():
+    middleware = TokenUsageMiddleware()
+    state = {
+        "messages": [
+            AIMessage(
+                content="done",
+                usage_metadata={
+                    "input_tokens": 10,
+                    "output_tokens": 5,
+                    "total_tokens": 15,
+                },
+            )
+        ]
+    }
+
+    with patch("deerflow.agents.middlewares.token_usage_middleware.logger.info") as info_mock:
+        result = middleware.after_model(state=state, runtime=MagicMock())
+
+    assert result is None
+    info_mock.assert_called_once_with(
+        "LLM token usage: input=%s output=%s total=%s",
+        10,
+        5,
+        15,
+    )