Add configurable LM Studio flushing and LLM retries.

test/test_llm_session_mocked.py

@@ -1,5 +1,7 @@
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 from photools3.llm_session import LlmSession
 
 
@@ -84,3 +86,70 @@ def test_llm_session_captions_and_keywords_mocked(test_photo):
 
         # Verify that the mock was called
         assert mock_client.chat.completions.create.called
+
+
+def test_llm_session_retries_from_original_request_without_accumulating_messages():
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "correct"}'),
+        ]
+
+        session = LlmSession(system_prompt="You are a helper.")
+        result = session.call(photos=[], prompt="Return JSON")
+
+    assert result.text == "correct"
+    requests = mock_client.chat.completions.create.call_args_list
+    messages_per_attempt = [call.kwargs["messages"] for call in requests]
+    assert [len(messages) for messages in messages_per_attempt] == [2, 3, 3]
+    final_correction = messages_per_attempt[2][-1]["content"]
+    assert '"wrong": "second"' in final_correction
+    assert '"wrong": "first"' not in final_correction
+
+
+def test_llm_session_uses_configured_retry_count(monkeypatch):
+    monkeypatch.setattr("photools3.llm_session.settings.llm_max_retries", 1)
+
+    with patch("photools3.llm_session.OpenAI") as mocked_openai:
+        mock_client = MagicMock()
+        mocked_openai.return_value = mock_client
+        mock_model = MagicMock()
+        mock_model.id = "gpt-4o"
+        mock_client.models.list.return_value.data = [mock_model]
+
+        def response(content):
+            message = MagicMock()
+            message.content = content
+            choice = MagicMock()
+            choice.message = message
+            result = MagicMock()
+            result.choices = [choice]
+            return result
+
+        mock_client.chat.completions.create.side_effect = [
+            response('{"wrong": "first"}'),
+            response('{"wrong": "second"}'),
+            response('{"text": "would be a third call"}'),
+        ]
+        session = LlmSession(system_prompt="You are a helper.")
+
+        with pytest.raises(ValueError, match="after 1 retries"):
+            session.call(photos=[])
+
+    assert mock_client.chat.completions.create.call_count == 2