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(ai): prevent duplicated system prompts in PydanticAI completions (#8356)

Fabian Preiß844436abe3e2
diff --git a/marimo/_server/ai/providers.py b/marimo/_server/ai/providers.py
index 437093feb5..cda6d1af3e 100644
--- a/marimo/_server/ai/providers.py
+++ b/marimo/_server/ai/providers.py
@@ -225,7 +225,6 @@ class PydanticProvider(ABC, Generic[ProviderT]):
result = await agent.run(
user_prompt=None,
message_history=VercelAIAdapter.load_messages(messages),
- instructions=system_prompt,
)
return str(result.output)
diff --git a/tests/_server/ai/test_providers.py b/tests/_server/ai/test_providers.py
index cbf811dbec..edb5e3ff63 100644
--- a/tests/_server/ai/test_providers.py
+++ b/tests/_server/ai/test_providers.py
@@ -1,5 +1,7 @@
"""Tests for the LLM providers in marimo._server.ai.providers."""
+from unittest.mock import AsyncMock, MagicMock, patch
+
import pytest
from marimo._config.config import AiConfig
@@ -333,3 +335,40 @@ def test_anthropic_is_extended_thinking_model(
config = AnyProviderConfig(api_key="test-key", base_url=None)
provider = AnthropicProvider(model_name, config)
assert provider.is_extended_thinking_model(model_name) == expected
+
+
[email protected]("pydantic_ai")
+async def test_completion_does_not_pass_redundant_instructions() -> None:
+ from pydantic_ai.messages import ModelResponse, TextPart
+ from pydantic_ai.models.openai import OpenAIResponsesModel
+
+ config = AnyProviderConfig(api_key="test-key", base_url="http://test-url")
+ provider = OpenAIProvider("gpt-4", config)
+
+ with (
+ patch("marimo._server.ai.providers.get_tool_manager") as mock_get_tm,
+ patch.object(
+ OpenAIResponsesModel, "request", new_callable=AsyncMock
+ ) as mock_request,
+ ):
+ mock_get_tm.return_value = MagicMock()
+ mock_request.return_value = ModelResponse(
+ parts=[TextPart(content="test")]
+ )
+
+ await provider.completion(
+ messages=[],
+ system_prompt="Test prompt",
+ max_tokens=100,
+ additional_tools=[],
+ )
+
+ mock_request.assert_called_once()
+ request_messages = mock_request.call_args.args[0]
+
+ assert len(request_messages) == 1
+ # The bug caused instructions to be "Test prompt\nTest prompt"
+ instructions = request_messages[0].instructions
+
+ # This asserts the duplication is gone
+ assert instructions == "Test prompt"