Skip to content

Commit 1ac6d82

Browse files
fix: normalize nullable Chat Completions token counts (#5238)
1 parent 7b27134 commit 1ac6d82

4 files changed

Lines changed: 116 additions & 6 deletions

File tree

‎src/agents/extensions/models/any_llm_model.py‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -631,9 +631,9 @@ async def _get_response_via_chat(
631631
usage = (
632632
Usage(
633633
requests=1,
634-
input_tokens=response.usage.prompt_tokens,
635-
output_tokens=response.usage.completion_tokens,
636-
total_tokens=response.usage.total_tokens,
634+
input_tokens=response.usage.prompt_tokens or 0,
635+
output_tokens=response.usage.completion_tokens or 0,
636+
total_tokens=response.usage.total_tokens or 0,
637637
input_tokens_details=response.usage.prompt_tokens_details, # type: ignore[arg-type]
638638
output_tokens_details=response.usage.completion_tokens_details, # type: ignore[arg-type]
639639
)

‎src/agents/models/openai_chatcompletions.py‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -288,9 +288,9 @@ async def get_response(
288288
usage = (
289289
Usage(
290290
requests=1,
291-
input_tokens=response.usage.prompt_tokens,
292-
output_tokens=response.usage.completion_tokens,
293-
total_tokens=response.usage.total_tokens,
291+
input_tokens=response.usage.prompt_tokens or 0,
292+
output_tokens=response.usage.completion_tokens or 0,
293+
total_tokens=response.usage.total_tokens or 0,
294294
# BeforeValidator in Usage normalizes these from Chat Completions types
295295
input_tokens_details=response.usage.prompt_tokens_details, # type: ignore[arg-type]
296296
output_tokens_details=response.usage.completion_tokens_details, # type: ignore[arg-type]

‎tests/models/test_any_llm_model.py‎

Lines changed: 54 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -305,6 +305,60 @@ async def _empty_chat_stream() -> AsyncIterator[ChatCompletionChunk]:
305305
)
306306

307307

308+
@pytest.mark.allow_call_model_methods
309+
@pytest.mark.asyncio
310+
@pytest.mark.parametrize(
311+
("counts", "expected"),
312+
[
313+
((None, 5, 12), (0, 5, 12)),
314+
((7, None, 12), (7, 0, 12)),
315+
((7, 5, None), (7, 5, 0)),
316+
((None, None, None), (0, 0, 0)),
317+
((7, 5, 12), (7, 5, 12)),
318+
(None, (0, 0, 0)),
319+
],
320+
ids=["null-input", "null-output", "null-total", "all-null", "valid", "no-usage"],
321+
)
322+
async def test_any_llm_runner_normalizes_nullable_chat_usage(
323+
monkeypatch: pytest.MonkeyPatch,
324+
counts: tuple[int | None, int | None, int | None] | None,
325+
expected: tuple[int, int, int],
326+
) -> None:
327+
usage = (
328+
{
329+
"prompt_tokens": counts[0],
330+
"completion_tokens": counts[1],
331+
"total_tokens": counts[2],
332+
"prompt_tokens_details": {"cached_tokens": 3},
333+
"completion_tokens_details": {"reasoning_tokens": 2},
334+
}
335+
if counts is not None
336+
else None
337+
)
338+
chat = _chat_completion("Hello")
339+
# Provider SDK parsing can construct typed responses without validating null counts.
340+
chat.usage = CompletionUsage.model_construct(**usage) if usage is not None else None
341+
if chat.usage is not None:
342+
chat.usage.prompt_tokens_details = PromptTokensDetails(cached_tokens=3)
343+
chat.usage.completion_tokens_details = CompletionTokensDetails(reasoning_tokens=2)
344+
provider = FakeAnyLLMProvider(supports_responses=False, chat_response=chat)
345+
module, _ = _import_any_llm_module(monkeypatch, provider)
346+
agent = Agent(
347+
name="test",
348+
model=module.AnyLLMModel(model="openai/fake", api="chat_completions"),
349+
model_settings=ModelSettings(preserve_raw_usage=True),
350+
)
351+
result = await Runner.run(agent, "hi")
352+
353+
assert result.final_output == "Hello"
354+
normalized = result.context_wrapper.usage
355+
assert normalized.requests == 1
356+
assert (normalized.input_tokens, normalized.output_tokens, normalized.total_tokens) == expected
357+
assert normalized.input_tokens_details.cached_tokens == (3 if counts is not None else 0)
358+
assert normalized.output_tokens_details.reasoning_tokens == (2 if counts is not None else 0)
359+
assert result.raw_responses[0].raw_usage == usage
360+
361+
308362
@pytest.mark.allow_call_model_methods
309363
@pytest.mark.asyncio
310364
@pytest.mark.parametrize("override_ua", [None, "test_user_agent"])

‎tests/models/test_openai_chatcompletions.py‎

Lines changed: 56 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -120,6 +120,62 @@ def __init__(self, completions: DummyCompletions) -> None:
120120
return completions.kwargs
121121

122122

123+
@pytest.mark.allow_call_model_methods
124+
@pytest.mark.asyncio
125+
@pytest.mark.parametrize(
126+
("counts", "expected"),
127+
[
128+
((None, 5, 12), (0, 5, 12)),
129+
((7, None, 12), (7, 0, 12)),
130+
((7, 5, None), (7, 5, 0)),
131+
((None, None, None), (0, 0, 0)),
132+
((7, 5, 12), (7, 5, 12)),
133+
(None, (0, 0, 0)),
134+
],
135+
ids=["null-input", "null-output", "null-total", "all-null", "valid", "no-usage"],
136+
)
137+
async def test_runner_normalizes_nullable_chat_usage(
138+
counts: tuple[int | None, int | None, int | None] | None,
139+
expected: tuple[int, int, int],
140+
) -> None:
141+
usage = (
142+
{
143+
"prompt_tokens": counts[0],
144+
"completion_tokens": counts[1],
145+
"total_tokens": counts[2],
146+
"prompt_tokens_details": {"cached_tokens": 3},
147+
"completion_tokens_details": {"reasoning_tokens": 2},
148+
}
149+
if counts is not None
150+
else None
151+
)
152+
153+
async def handler(request: httpx2.Request) -> httpx2.Response:
154+
payload = _minimal_chat_completion("Hello").model_dump()
155+
payload["usage"] = usage
156+
return httpx2.Response(200, json=payload)
157+
158+
async with AsyncOpenAI(
159+
api_key="test-key",
160+
base_url="https://provider.example/v1",
161+
http_client=httpx2.AsyncClient(transport=httpx2.MockTransport(handler)),
162+
) as client:
163+
agent = Agent(
164+
name="test",
165+
model=OpenAIChatCompletionsModel(model="fake", openai_client=client),
166+
model_settings=ModelSettings(preserve_raw_usage=True),
167+
)
168+
result = await Runner.run(agent, "hi")
169+
170+
assert result.final_output == "Hello"
171+
normalized = result.context_wrapper.usage
172+
assert normalized.requests == 1
173+
assert (normalized.input_tokens, normalized.output_tokens, normalized.total_tokens) == expected
174+
assert normalized.input_tokens_details.cached_tokens == (3 if counts is not None else 0)
175+
assert normalized.output_tokens_details.reasoning_tokens == (2 if counts is not None else 0)
176+
assert result.raw_responses[0].raw_usage == usage
177+
178+
123179
@pytest.mark.allow_call_model_methods
124180
@pytest.mark.asyncio
125181
async def test_falsy_reasoning_is_forwarded() -> None:

0 commit comments

Comments
 (0)