Skip to content

Build and run agents you can see, understand and trust.

fix(model): read Moonshot cached prompt tokens from prompt_tokens_details (#3269)

zgy2d8eab428346
diff --git a/src/agentscope/model/_moonshot/_model.py b/src/agentscope/model/_moonshot/_model.py
index dd93f7dbc..e2ce68d23 100644
--- a/src/agentscope/model/_moonshot/_model.py
+++ b/src/agentscope/model/_moonshot/_model.py
@@ -280,15 +280,18 @@ class MoonshotChatModel(ChatModelBase):
if chunk.usage:
u = chunk.usage
+ details = getattr(u, "prompt_tokens_details", None)
usage = ChatUsage(
input_tokens=u.prompt_tokens,
output_tokens=u.completion_tokens,
time=(datetime.now() - start_datetime).total_seconds(),
cache_input_tokens=getattr(
- u,
+ details,
"cached_tokens",
0,
- ),
+ )
+ if details
+ else 0,
)
if not chunk.choices:
@@ -386,11 +389,18 @@ class MoonshotChatModel(ChatModelBase):
usage = None
if response.usage:
u = response.usage
+ details = getattr(u, "prompt_tokens_details", None)
usage = ChatUsage(
input_tokens=u.prompt_tokens,
output_tokens=u.completion_tokens,
time=(datetime.now() - start_datetime).total_seconds(),
- cache_input_tokens=getattr(u, "cached_tokens", 0),
+ cache_input_tokens=getattr(
+ details,
+ "cached_tokens",
+ 0,
+ )
+ if details
+ else 0,
)
resp_kwargs: dict[str, Any] = {
diff --git a/tests/model_moonshot_test.py b/tests/model_moonshot_test.py
index 864f6530b..f7f7119fb 100644
--- a/tests/model_moonshot_test.py
+++ b/tests/model_moonshot_test.py
@@ -63,7 +63,7 @@ def _mock_completion(
resp.choices = [choice]
resp.usage.prompt_tokens = 10
resp.usage.completion_tokens = 5
- resp.usage.cached_tokens = 0
+ resp.usage.prompt_tokens_details.cached_tokens = 0
return resp
@@ -83,7 +83,10 @@ def _make_stream_chunk(
chunk.usage = MagicMock()
chunk.usage.prompt_tokens = usage.get("prompt_tokens", 0)
chunk.usage.completion_tokens = usage.get("completion_tokens", 0)
- chunk.usage.cached_tokens = usage.get("cached_tokens", 0)
+ chunk.usage.prompt_tokens_details.cached_tokens = usage.get(
+ "cached_tokens",
+ 0,
+ )
else:
chunk.usage = None
@@ -170,6 +173,52 @@ class TestMoonshotNonStream(IsolatedAsyncioTestCase):
)
self.assertEqual(result.id, "kimi-1")
+ async def test_usage_cache_input_tokens_from_nested_details(self) -> None:
+ """Read cached tokens from nested usage details."""
+ resp = _mock_completion(text="Hello!")
+ resp.usage.prompt_tokens = 1000
+ resp.usage.completion_tokens = 20
+ resp.usage.prompt_tokens_details.cached_tokens = 832
+ mock_create = AsyncMock(return_value=resp)
+ self.mock_client.chat.completions.create = mock_create
+
+ result = await self.model([])
+
+ self.assertEqual(
+ dict(result.usage),
+ {
+ "input_tokens": 1000,
+ "output_tokens": 20,
+ "time": result.usage.time,
+ "cache_creation_input_tokens": 0,
+ "cache_input_tokens": 832,
+ "type": "chat",
+ "metadata": None,
+ },
+ )
+
+ async def test_usage_cache_input_tokens_absent_details(self) -> None:
+ """Default cached tokens to zero without usage details."""
+ resp = _mock_completion(text="Hello!")
+ del resp.usage.prompt_tokens_details
+ mock_create = AsyncMock(return_value=resp)
+ self.mock_client.chat.completions.create = mock_create
+
+ result = await self.model([])
+
+ self.assertEqual(
+ dict(result.usage),
+ {
+ "input_tokens": 10,
+ "output_tokens": 5,
+ "time": result.usage.time,
+ "cache_creation_input_tokens": 0,
+ "cache_input_tokens": 0,
+ "type": "chat",
+ "metadata": None,
+ },
+ )
+
async def test_extra_body_is_not_mutated(self) -> None:
"""Defaulting thinking.type leaves the caller's extra_body intact."""
mock_create = AsyncMock(
@@ -331,6 +380,40 @@ class TestMoonshotStream(IsolatedAsyncioTestCase):
],
)
+ async def test_stream_usage_cache_input_tokens_from_nested_details(
+ self,
+ ) -> None:
+ """Read cached tokens from a streaming usage chunk."""
+ chunks = [
+ _make_stream_chunk(delta_text="Hi"),
+ _make_stream_chunk(
+ has_choices=False,
+ usage={
+ "prompt_tokens": 1000,
+ "completion_tokens": 20,
+ "cached_tokens": 832,
+ },
+ ),
+ ]
+ mock_create = AsyncMock(return_value=_MockAsyncStream(chunks))
+ self.mock_client.chat.completions.create = mock_create
+
+ gen = await self.model([])
+ responses = [r async for r in gen]
+
+ self.assertEqual(
+ dict(responses[-1].usage),
+ {
+ "input_tokens": 1000,
+ "output_tokens": 20,
+ "time": responses[-1].usage.time,
+ "cache_creation_input_tokens": 0,
+ "cache_input_tokens": 832,
+ "type": "chat",
+ "metadata": None,
+ },
+ )
+
async def test_stream_thinking_then_text(
self,
) -> None: