Skip to content

A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.

fix: disable thinking for inline autocomplete (#10890)

Shahmir Varqhab3c010464fb9
diff --git a/marimo/_server/ai/providers.py b/marimo/_server/ai/providers.py
index 01bd0924c5..e6ddcb85ed 100644
--- a/marimo/_server/ai/providers.py
+++ b/marimo/_server/ai/providers.py
@@ -407,6 +407,7 @@ class PydanticProvider(ABC, Generic[ProviderT_co]):
additional_tools: list[ToolDefinition],
span_info: SpanInfo,
enable_capabilities: bool = True,
+ thinking: ThinkingLevel | None = None,
) -> str:
"""Return a string response from the given messages."""
@@ -428,6 +429,9 @@ class PydanticProvider(ABC, Generic[ProviderT_co]):
result = await agent.run(
user_prompt=None,
message_history=VercelAIAdapter.load_messages(messages),
+ model_settings={"thinking": thinking}
+ if thinking is not None
+ else None,
)
return str(result.output)
diff --git a/marimo/_server/api/endpoints/ai.py b/marimo/_server/api/endpoints/ai.py
index 7d4c96c60d..1ed4241398 100644
--- a/marimo/_server/api/endpoints/ai.py
+++ b/marimo/_server/api/endpoints/ai.py
@@ -345,6 +345,7 @@ async def ai_inline_completion(
messages=[create_simple_prompt(prompt)],
system_prompt=system_prompt,
max_tokens=INLINE_COMPLETION_MAX_TOKENS,
+ thinking=False,
additional_tools=[],
enable_capabilities=False,
span_info=SpanInfo(
diff --git a/tests/_server/ai/test_providers.py b/tests/_server/ai/test_providers.py
index 2988171c52..a9f79bfa6d 100644
--- a/tests/_server/ai/test_providers.py
+++ b/tests/_server/ai/test_providers.py
@@ -860,6 +860,42 @@ async def test_completion_does_not_pass_redundant_instructions() -> None:
assert instructions == "Test prompt"
[email protected]("pydantic_ai")
[email protected]("thinking", [None, False])
+async def test_completion_thinking_override(thinking: bool | None) -> None:
+ from pydantic_ai.messages import ModelResponse, TextPart
+ from pydantic_ai.models.openai import OpenAIResponsesModel
+
+ config = AnyProviderConfig(api_key="test-key", base_url=None)
+ provider = OpenAIProvider("gpt-5.1", config)
+
+ with patch.object(
+ OpenAIResponsesModel, "request", new_callable=AsyncMock
+ ) as mock_request:
+ mock_request.return_value = ModelResponse(
+ parts=[TextPart(content="print(1)")]
+ )
+ result = await provider.completion(
+ messages=[],
+ system_prompt="Complete the code.",
+ max_tokens=1024,
+ additional_tools=[],
+ enable_capabilities=False,
+ thinking=thinking,
+ span_info=SpanInfo(
+ endpoint="inline_completion", model="openai/gpt-5.1"
+ ),
+ )
+
+ assert result == "print(1)"
+ mock_request.assert_called_once()
+ assert mock_request.call_args.args[1] == {
+ "max_tokens": 1024,
+ "thinking": True if thinking is None else thinking,
+ "openai_reasoning_summary": "auto",
+ }
+
+
@pytest.mark.requires("pydantic_ai")
async def test_completion_tool_count_includes_capabilities() -> None:
"""`completion` reports tools plus the agent's native capabilities, so its
diff --git a/tests/_server/api/endpoints/test_ai.py b/tests/_server/api/endpoints/test_ai.py
index 0a46038ff1..0e20c4ed02 100644
--- a/tests/_server/api/endpoints/test_ai.py
+++ b/tests/_server/api/endpoints/test_ai.py
@@ -402,6 +402,7 @@ class TestOpenAiEndpoints:
mock_completion.assert_called_once()
# Assert the messages contain FIM format
call_kwargs = mock_completion.call_args.kwargs
+ assert call_kwargs["thinking"] is False
messages = call_kwargs["messages"]
assert len(messages) == 1
# Verify FIM format is used