Description
structured_outputs_transform creates a single ChatHistory in the factory and closes over it. Each output_transform call add_messages into that same history, so prior orchestration outputs leak into later structured-output prompts and the history grows without bound (lengths 2→3→4…).
Environment
semantic-kernel==1.44.1 (Python)- Python 3.13
- Repro uses a mocked chat service (no API keys)
Minimal repro
import asyncio
from typing import Any
from pydantic import BaseModel, PrivateAttr
from semantic_kernel.agents.orchestration.tools import structured_outputs_transform
from semantic_kernel.connectors.ai.chat_completion_client_base import ChatCompletionClientBase
from semantic_kernel.connectors.ai.prompt_execution_settings import PromptExecutionSettings
from semantic_kernel.contents import ChatMessageContent
from semantic_kernel.contents.utils.author_role import AuthorRole
class Out(BaseModel):
answer: str
class FakeSettings(PromptExecutionSettings):
response_format: Any = None
class FakeService(ChatCompletionClientBase):
_calls: list = PrivateAttr(default_factory=list)
def __init__(self) -> None:
super().__init__(service_id="fake", ai_model_id="fake")
def get_prompt_execution_settings_class(self):
return FakeSettings
async def get_chat_message_contents(self, chat_history, settings, **kwargs):
self._calls.append(len(chat_history.messages))
return [ChatMessageContent(role=AuthorRole.ASSISTANT, content='{"answer":"x"}')]
async def get_chat_message_content(self, chat_history, settings, **kwargs):
self._calls.append(len(chat_history.messages))
return ChatMessageContent(role=AuthorRole.ASSISTANT, content='{"answer":"x"}')
async def main():
svc = FakeService()
transform = structured_outputs_transform(Out, svc)
await transform(ChatMessageContent(role=AuthorRole.USER, content="first"))
await transform(ChatMessageContent(role=AuthorRole.USER, content="second"))
await transform(ChatMessageContent(role=AuthorRole.USER, content="third"))
print(svc._calls) # ACTUAL: [2, 3, 4] EXPECTED: [2, 2, 2]
asyncio.run(main())Expected
Each transform invocation uses a fresh history (system + current message only), so independent orchestration steps do not pollute each other.
Actual
Shared history grows across calls; prior outputs leak into later structured-output prompts.
Root cause (pointer)
semantic_kernel/agents/orchestration/tools.py — structured_outputs_transform: chat_history = ChatHistory(...) created once in the factory and mutated inside output_transform.
Suggested fix direction
Construct a fresh ChatHistory (with the system message) inside output_transform on every call.
Description
structured_outputs_transformcreates a singleChatHistoryin the factory and closes over it. Eachoutput_transformcalladd_messages into that same history, so prior orchestration outputs leak into later structured-output prompts and the history grows without bound (lengths 2→3→4…).Environment
semantic-kernel==1.44.1(Python)Minimal repro
Expected
Each transform invocation uses a fresh history (system + current message only), so independent orchestration steps do not pollute each other.
Actual
Shared history grows across calls; prior outputs leak into later structured-output prompts.
Root cause (pointer)
semantic_kernel/agents/orchestration/tools.py—structured_outputs_transform:chat_history = ChatHistory(...)created once in the factory and mutated insideoutput_transform.Suggested fix direction
Construct a fresh
ChatHistory(with the system message) insideoutput_transformon every call.