Skip to content

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

feat(model): omit Gemini thinking_config unless thinking is explicitly configured (#3523)

Kiko03012599410f4625
diff --git a/src/agentscope/model/_gemini/_model.py b/src/agentscope/model/_gemini/_model.py
index f4739d887..1163c2c7f 100644
--- a/src/agentscope/model/_gemini/_model.py
+++ b/src/agentscope/model/_gemini/_model.py
@@ -153,10 +153,12 @@ class GeminiChatModel(ChatModelBase):
gt=0,
)
- thinking_enable: bool = Field(
- default=False,
+ thinking_enable: bool | None = Field(
+ default=None,
title="Thinking",
- description="Whether to enable thinking output.",
+ description="Whether to enable thinking output."
+ " When None, no ``thinking_config`` is sent and the"
+ " model's default behavior applies.",
)
thinking_budget: int | None = Field(
@@ -296,16 +298,17 @@ class GeminiChatModel(ChatModelBase):
if self.parameters.top_p is not None:
config["top_p"] = self.parameters.top_p
- if self.parameters.thinking_enable:
- config["thinking_config"] = {
- "include_thoughts": True,
- "thinking_budget": self.parameters.thinking_budget or 1024,
- }
- else:
- config["thinking_config"] = {
- "include_thoughts": False,
- "thinking_budget": 0,
- }
+ if self.parameters.thinking_enable is not None:
+ if self.parameters.thinking_enable:
+ config["thinking_config"] = {
+ "include_thoughts": True,
+ "thinking_budget": self.parameters.thinking_budget or 1024,
+ }
+ else:
+ config["thinking_config"] = {
+ "include_thoughts": False,
+ "thinking_budget": 0,
+ }
# Per-call config takes precedence over model-level defaults.
config.update(config_kwargs)
diff --git a/tests/model_gemini_test.py b/tests/model_gemini_test.py
index 1013aa715..4e7eae268 100644
--- a/tests/model_gemini_test.py
+++ b/tests/model_gemini_test.py
@@ -166,6 +166,7 @@ class TestGeminiNonStream(IsolatedAsyncioTestCase):
self.model.parameters.max_tokens = 1024
self.model.parameters.temperature = 0.2
self.model.parameters.top_p = 0.3
+ self.model.parameters.thinking_enable = True
self.mock_client.aio.models.generate_content = AsyncMock(
return_value=_mock_completion([_make_part(text="ok")]),
)
@@ -175,8 +176,8 @@ class TestGeminiNonStream(IsolatedAsyncioTestCase):
"temperature": 0.2,
"top_p": 0.3,
"thinking_config": {
- "include_thoughts": False,
- "thinking_budget": 0,
+ "include_thoughts": True,
+ "thinking_budget": 1024,
},
}
cases = {
@@ -202,6 +203,78 @@ class TestGeminiNonStream(IsolatedAsyncioTestCase):
{**default_config, key: value},
)
+ async def test_thinking_config_omitted_by_default(self) -> None:
+ """Default parameters send no ``thinking_config`` so the model's
+ default behavior applies (e.g. Pro models think by default and
+ reject ``thinking_budget: 0``)."""
+ self.mock_client.aio.models.generate_content = AsyncMock(
+ return_value=_mock_completion([_make_part(text="ok")]),
+ )
+ self.model.parameters = GeminiChatModel.Parameters()
+
+ await self.model([])
+
+ sent_config = (
+ self.mock_client.aio.models.generate_content.call_args.kwargs[
+ "config"
+ ]
+ )
+ self.assertEqual(sent_config, {})
+
+ async def test_thinking_config_false_forwarded(self) -> None:
+ """``thinking_enable=False`` sends the disable configuration."""
+ self.mock_client.aio.models.generate_content = AsyncMock(
+ return_value=_mock_completion([_make_part(text="ok")]),
+ )
+ self.model.parameters = GeminiChatModel.Parameters(
+ thinking_enable=False,
+ )
+
+ await self.model([])
+
+ sent_config = (
+ self.mock_client.aio.models.generate_content.call_args.kwargs[
+ "config"
+ ]
+ )
+ self.assertEqual(
+ sent_config,
+ {
+ "thinking_config": {
+ "include_thoughts": False,
+ "thinking_budget": 0,
+ },
+ },
+ )
+
+ async def test_thinking_config_true_forwarded(self) -> None:
+ """Enabled thinking uses the default or explicitly provided budget."""
+ self.mock_client.aio.models.generate_content = AsyncMock(
+ return_value=_mock_completion([_make_part(text="ok")]),
+ )
+
+ for budget, expected_budget in [(None, 1024), (8192, 8192)]:
+ with self.subTest(thinking_budget=budget):
+ self.model.parameters = GeminiChatModel.Parameters(
+ thinking_enable=True,
+ thinking_budget=budget,
+ )
+
+ await self.model([])
+
+ call_args = (
+ self.mock_client.aio.models.generate_content.call_args
+ )
+ self.assertEqual(
+ call_args.kwargs["config"],
+ {
+ "thinking_config": {
+ "include_thoughts": True,
+ "thinking_budget": expected_budget,
+ },
+ },
+ )
+
async def test_tool_call_response(self) -> None:
"""Non-stream tool call response creates ToolCallBlocks."""
parts = [