From c53e89b45c94884d248ef8fc7a49b64c8e33e49e Mon Sep 17 00:00:00 2001
From: Charles Cheng <charlescheng@rezona.ai>
Date: Sun, 28 Jun 2026 10:16:30 +0800
Subject: [PATCH] Python: convert Pydantic model class response_format to JSON
 schema in OllamaChatClient

Ollama's `format` param only accepts '', 'json', or a JSON-schema dict, so
passing a Pydantic model class (the form OpenAIChatClient/FoundryChatClient and
create_harness_agent plan mode use) raised a ValidationError while building the
request. Convert a model class to its JSON schema when mapping response_format
-> format, keeping the original class for typed response parsing.
---
 .../agent_framework_ollama/_chat_client.py    | 10 ++++-
 .../ollama/tests/test_ollama_chat_client.py   | 38 +++++++++++++++++++
 2 files changed, 47 insertions(+), 1 deletion(-)

diff --git a/python/packages/ollama/agent_framework_ollama/_chat_client.py b/python/packages/ollama/agent_framework_ollama/_chat_client.py
index 931011d74c1..f5820d46e09 100644
--- a/python/packages/ollama/agent_framework_ollama/_chat_client.py
+++ b/python/packages/ollama/agent_framework_ollama/_chat_client.py
@@ -87,7 +87,8 @@ class OllamaChatOptions(ChatOptions[ResponseModelT], Generic[ResponseModelT], to
         presence_penalty: Presence penalty, translates to ``options.presence_penalty``.
         tools: List of function tools.
         response_format: Output format, translates to ``format``.
-            Use 'json' for JSON mode or a JSON schema dict for structured output.
+            Use 'json' for JSON mode, a JSON schema dict, or a Pydantic model class
+            (converted to its JSON schema) for structured output.
 
         # Options not supported in Ollama:
         tool_choice: Ollama only supports auto tool choice.
@@ -415,6 +416,13 @@ def _prepare_options(self, messages: Sequence[Message], options: Mapping[str, An
             else:
                 # Apply top-level translations (e.g., response_format -> format)
                 translated_key = OLLAMA_OPTION_TRANSLATIONS.get(key, key)
+                if translated_key == "format" and isinstance(value, type) and issubclass(value, BaseModel):
+                    # Ollama's `format` accepts '', 'json', or a JSON-schema dict, not a
+                    # Pydantic model class. Convert the class to its JSON schema, matching
+                    # OpenAIChatClient/FoundryChatClient and Ollama's documented usage
+                    # (https://ollama.com/blog/structured-outputs). The original class is
+                    # kept in `options` for typed parsing of the response.
+                    value = value.model_json_schema()
                 run_options[translated_key] = value
 
         # Add model options to run_options if any
diff --git a/python/packages/ollama/tests/test_ollama_chat_client.py b/python/packages/ollama/tests/test_ollama_chat_client.py
index f3061f60369..2be736df79f 100644
--- a/python/packages/ollama/tests/test_ollama_chat_client.py
+++ b/python/packages/ollama/tests/test_ollama_chat_client.py
@@ -19,6 +19,7 @@
 from ollama._types import ChatResponse as OllamaChatResponse
 from ollama._types import Message as OllamaMessage
 from openai import AsyncStream
+from pydantic import BaseModel
 from pytest import fixture
 
 from agent_framework_ollama import OllamaChatClient
@@ -281,6 +282,43 @@ async def test_cmc_response_format_dict(
     assert result.value["answer"] == "test"
 
 
+@patch.object(AsyncClient, "chat", new_callable=AsyncMock)
+async def test_cmc_response_format_pydantic_model(
+    mock_chat: AsyncMock,
+    ollama_unit_test_env: dict[str, str],
+    chat_history: list[Message],
+) -> None:
+    """A Pydantic model class is converted to a JSON schema dict for Ollama's ``format``.
+
+    Ollama only accepts ``''``, ``'json'``, or a JSON-schema dict for ``format``; a model
+    class would fail request construction. The class is still kept for typed parsing of
+    the response, matching OpenAI/Foundry behavior.
+    """
+
+    class Answer(BaseModel):
+        answer: str
+
+    mock_chat.return_value = OllamaChatResponse(
+        message=OllamaMessage(content='{"answer": "test"}', role="assistant"),
+        model="test",
+        eval_count=1,
+        prompt_eval_count=1,
+        created_at="2024-01-01T00:00:00Z",
+    )
+    chat_history.append(Message(contents=["hello world"], role="user"))
+
+    ollama_client = OllamaChatClient()
+    result = await ollama_client.get_response(messages=chat_history, options={"response_format": Answer})
+
+    # Outgoing ``format`` must be the JSON schema dict, not the model class.
+    assert mock_chat.await_args is not None
+    assert mock_chat.await_args.kwargs["format"] == Answer.model_json_schema()
+
+    # Typed parsing still works because the original model class is preserved.
+    assert isinstance(result.value, Answer)
+    assert result.value.answer == "test"
+
+
 @patch.object(AsyncClient, "chat", new_callable=AsyncMock)
 async def test_cmc_reasoning(
     mock_chat: AsyncMock,