From 8892025ed715186c8b4b7542443633e89484c419 Mon Sep 17 00:00:00 2001
From: Giulio Lovisotto <glovisotto@nvidia.com>
Date: Tue, 6 Oct 2026 23:45:49 +0200
Subject: [PATCH] fix(hermes): report native token usage

Signed-off-by: Giulio Lovisotto <glovisotto@nvidia.com>
---
 responses_api_agents/hermes_agent/app.py      | 24 +++++---
 .../hermes_agent/tests/test_app.py            | 55 +++++++++++++++++++
 2 files changed, 72 insertions(+), 7 deletions(-)

diff --git a/responses_api_agents/hermes_agent/app.py b/responses_api_agents/hermes_agent/app.py
index 45355923db7..990f6fd13b7 100644
--- a/responses_api_agents/hermes_agent/app.py
+++ b/responses_api_agents/hermes_agent/app.py
@@ -82,6 +82,22 @@
 from responses_api_agents.hermes_agent.observability import HermesAgentObserver, normalize_hermes_messages
 
 
+def _usage_from_result(result: dict[str, Any]) -> Optional[NeMoGymResponseUsage]:
+    # Hermes' prompt total already includes cache reads/writes, and its completion
+    # total includes reasoning. Early returns can omit these native aggregates.
+    prompt_tokens = result.get("prompt_tokens")
+    completion_tokens = result.get("completion_tokens")
+    if prompt_tokens is None or completion_tokens is None:
+        return None
+    return NeMoGymResponseUsage(
+        input_tokens=prompt_tokens,
+        input_tokens_details=NeMoGymResponseInputTokensDetails(cached_tokens=result.get("cache_read_tokens")),
+        output_tokens=completion_tokens,
+        output_tokens_details=NeMoGymResponseOutputTokensDetails(reasoning_tokens=result.get("reasoning_tokens")),
+        total_tokens=prompt_tokens + completion_tokens,
+    )
+
+
 def _trajectory_to_output_items(messages, n_input):
     output_items = []
     for item in messages[n_input:]:
@@ -951,13 +967,7 @@ def _response_from_result(
             tool_choice=body.tool_choice,
             tools=body.tools,
             parallel_tool_calls=body.parallel_tool_calls,
-            usage=NeMoGymResponseUsage(
-                input_tokens=0,
-                input_tokens_details=NeMoGymResponseInputTokensDetails(cached_tokens=0),
-                output_tokens=0,
-                output_tokens_details=NeMoGymResponseOutputTokensDetails(reasoning_tokens=0),
-                total_tokens=0,
-            ),
+            usage=_usage_from_result(result),
         )
 
     async def _create_response(
diff --git a/responses_api_agents/hermes_agent/tests/test_app.py b/responses_api_agents/hermes_agent/tests/test_app.py
index be41c6c607a..fcc9a5ebcb4 100644
--- a/responses_api_agents/hermes_agent/tests/test_app.py
+++ b/responses_api_agents/hermes_agent/tests/test_app.py
@@ -679,6 +679,61 @@ def run_conversation(self, *args, **kwargs) -> dict:
         assert hermes.interrupted_agents == set()
 
 
+class TestResultUsage:
+    @pytest.mark.parametrize("partial", [False, True])
+    def test_reports_cache_inclusive_native_totals(self, partial) -> None:
+        hermes = HermesAgent(config=_config(), server_client=MagicMock(spec=ServerClient))
+        response = hermes._response_from_result(
+            body=NeMoGymResponseCreateParamsNonStreaming(input="hi"),
+            result={
+                "messages": [{"role": "assistant", "content": "done"}],
+                "partial": partial,
+                "input_tokens": 100,
+                "cache_read_tokens": 200,
+                "cache_write_tokens": 300,
+                "prompt_tokens": 600,
+                "completion_tokens": 80,
+                "reasoning_tokens": 30,
+            },
+            model_name="model",
+        )
+        assert response.usage.input_tokens == 600
+        assert response.usage.input_tokens_details.cached_tokens == 200
+        assert response.usage.output_tokens == 80
+        assert response.usage.output_tokens_details.reasoning_tokens == 30
+        assert response.usage.total_tokens == 680
+
+    @pytest.mark.parametrize(
+        "usage",
+        [{}, {"prompt_tokens": 10}, {"completion_tokens": 20}, {"prompt_tokens": None, "completion_tokens": 20}],
+    )
+    def test_missing_native_totals_remain_unknown(self, usage) -> None:
+        hermes = HermesAgent(config=_config(), server_client=MagicMock(spec=ServerClient))
+        response = hermes._response_from_result(
+            body=NeMoGymResponseCreateParamsNonStreaming(input="hi"),
+            result={"messages": [{"role": "assistant", "content": "partial answer"}], **usage},
+            model_name="model",
+        )
+        assert response.usage is None
+
+    def test_missing_details_remain_unknown(self) -> None:
+        hermes = HermesAgent(config=_config(), server_client=MagicMock(spec=ServerClient))
+        response = hermes._response_from_result(
+            body=NeMoGymResponseCreateParamsNonStreaming(input="hi"),
+            result={
+                "messages": [{"role": "assistant", "content": "done"}],
+                "prompt_tokens": 10,
+                "completion_tokens": 20,
+            },
+            model_name="model",
+        )
+        assert response.usage.input_tokens == 10
+        assert response.usage.output_tokens == 20
+        assert response.usage.total_tokens == 30
+        assert response.usage.input_tokens_details.cached_tokens is None
+        assert response.usage.output_tokens_details.reasoning_tokens is None
+
+
 class TestResultClassification:
     @pytest.mark.parametrize(
         "error",