Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(bot): preserve token usage for no-tool responses (#3535)

DuTaoa7c1065f01df
diff --git a/bot/vikingbot/providers/vlm_adapter.py b/bot/vikingbot/providers/vlm_adapter.py
index cf830c56d..de903eada 100644
--- a/bot/vikingbot/providers/vlm_adapter.py
+++ b/bot/vikingbot/providers/vlm_adapter.py
@@ -82,13 +82,16 @@ class VLMProviderAdapter(LLMProvider):
# --- Call VLM backend ---
attempt = 1
+ # An explicit empty list asks VLM backends for a structured response
+ # without exposing tools, so per-response usage is preserved.
+ response_tools = tools if tools is not None else []
while True:
try:
result = await self._vlm.get_completion_async(
messages=messages,
thinking=getattr(self._vlm, "thinking", None),
- tools=tools,
- tool_choice="auto" if tools else None,
+ tools=response_tools,
+ tool_choice="auto" if response_tools else None,
)
break
except Exception as e:
diff --git a/openviking/models/vlm/backends/openai_vlm.py b/openviking/models/vlm/backends/openai_vlm.py
index e06de8fbe..7c629209f 100644
--- a/openviking/models/vlm/backends/openai_vlm.py
+++ b/openviking/models/vlm/backends/openai_vlm.py
@@ -308,7 +308,7 @@ class OpenAIVLM(VLMBase):
t0 = time.perf_counter()
response = client.chat.completions.create(**kwargs)
elapsed = time.perf_counter() - t0
- if tools:
+ if tools is not None:
self._update_token_usage_from_response(response, duration_seconds=elapsed)
return self._build_vlm_response(response, has_tools=True)
return self._extract_completion_content(response, elapsed)
@@ -338,7 +338,7 @@ class OpenAIVLM(VLMBase):
t0 = time.perf_counter()
response = await client.chat.completions.create(**kwargs)
elapsed = time.perf_counter() - t0
- if tools:
+ if tools is not None:
self._update_token_usage_from_response(response, duration_seconds=elapsed)
return self._build_vlm_response(response, has_tools=True)
return await self._extract_completion_content_async(response, elapsed)
@@ -424,7 +424,7 @@ class OpenAIVLM(VLMBase):
t0 = time.perf_counter()
response = client.chat.completions.create(**kwargs)
elapsed = time.perf_counter() - t0
- if tools:
+ if tools is not None:
self._update_token_usage_from_response(response, duration_seconds=elapsed)
return self._build_vlm_response(response, has_tools=True)
return self._extract_completion_content(response, elapsed)
@@ -456,7 +456,7 @@ class OpenAIVLM(VLMBase):
t0 = time.perf_counter()
response = await client.chat.completions.create(**kwargs)
elapsed = time.perf_counter() - t0
- if tools:
+ if tools is not None:
self._update_token_usage_from_response(response, duration_seconds=elapsed)
return self._build_vlm_response(response, has_tools=True)
return await self._extract_completion_content_async(response, elapsed)
diff --git a/tests/unit/test_vikingbot_vlm_adapter_retry.py b/tests/unit/test_vikingbot_vlm_adapter_retry.py
index e51f8927e..44ee4a918 100644
--- a/tests/unit/test_vikingbot_vlm_adapter_retry.py
+++ b/tests/unit/test_vikingbot_vlm_adapter_retry.py
@@ -148,6 +148,48 @@ async def test_chat_accepts_string_response_from_openai_backend_with_tools(monke
assert response.finish_reason == "stop"
+async def test_chat_without_tools_preserves_usage_from_openai_backend(monkeypatch):
+ raw_response = SimpleNamespace(
+ choices=[
+ SimpleNamespace(
+ message=SimpleNamespace(content="plain response", tool_calls=None),
+ finish_reason="stop",
+ )
+ ],
+ usage=SimpleNamespace(
+ prompt_tokens=13,
+ completion_tokens=5,
+ total_tokens=18,
+ prompt_tokens_details=None,
+ completion_tokens_details=None,
+ ),
+ )
+
+ async def create(**kwargs):
+ assert "tools" not in kwargs
+ return raw_response
+
+ client = SimpleNamespace(
+ chat=SimpleNamespace(completions=SimpleNamespace(create=create)),
+ )
+ vlm = OpenAIVLM({"provider": "openai", "model": "gpt-5.6-terra"})
+ monkeypatch.setattr(vlm, "get_async_client", lambda: client)
+ adapter = VLMProviderAdapter(vlm, "gpt-5.6-terra", langfuse_client=_DisabledLangfuse())
+
+ response = await adapter.chat(messages=[{"role": "user", "content": "hello"}])
+
+ assert response.content == "plain response"
+ assert response.tool_calls == []
+ assert response.finish_reason == "stop"
+ assert response.usage == {
+ "prompt_tokens": 13,
+ "completion_tokens": 5,
+ "total_tokens": 18,
+ "prompt_tokens_details": None,
+ }
+
+
@pytest.mark.asyncio
async def test_chat_stream_retries_rate_limit_until_success(monkeypatch):
sleep_delays: list[float] = []
diff --git a/tests/unit/test_vlm_response_formats.py b/tests/unit/test_vlm_response_formats.py
index 8ff35ef58..3c9c768b4 100644
--- a/tests/unit/test_vlm_response_formats.py
+++ b/tests/unit/test_vlm_response_formats.py
@@ -111,3 +111,49 @@ async def test_openai_async_completion_from_str_with_tools(monkeypatch):
assert response.content == "plain string response"
assert response.tool_calls == []
assert response.finish_reason == "stop"
+
+
+async def test_openai_async_completion_with_empty_tools_preserves_usage(monkeypatch):
+ usage = SimpleNamespace(
+ prompt_tokens=11,
+ completion_tokens=4,
+ total_tokens=15,
+ prompt_tokens_details=None,
+ completion_tokens_details=None,
+ )
+ raw_response = SimpleNamespace(
+ choices=[
+ SimpleNamespace(
+ message=SimpleNamespace(content="plain response", tool_calls=None),
+ finish_reason="stop",
+ )
+ ],
+ usage=usage,
+ )
+
+ async def create(**kwargs):
+ assert "tools" not in kwargs
+ return raw_response
+
+ client = SimpleNamespace(
+ chat=SimpleNamespace(completions=SimpleNamespace(create=create)),
+ )
+ vlm = OpenAIVLM({"provider": "openai", "model": "gpt-5.6-terra"})
+ monkeypatch.setattr(vlm, "get_async_client", lambda: client)
+
+ response = await vlm.get_completion_async(
+ messages=[{"role": "user", "content": "hello"}],
+ tools=[],
+ )
+
+ assert isinstance(response, VLMResponse)
+ assert response.content == "plain response"
+ assert response.tool_calls == []
+ assert response.finish_reason == "stop"
+ assert response.usage == {
+ "prompt_tokens": 11,
+ "completion_tokens": 4,
+ "total_tokens": 15,
+ "prompt_tokens_details": None,
+ }