aider/tests/basic/test_reasoning.py
2025-03-07 17:16:50 -08:00

323 lines
14 KiB
Python

import unittest
from unittest.mock import MagicMock, patch
from aider.coders.base_coder import Coder
from aider.dump import dump # noqa
from aider.io import InputOutput
from aider.models import Model
class TestReasoning(unittest.TestCase):
def test_send_with_reasoning_content(self):
"""Test that reasoning content is properly formatted and output."""
# Setup IO with no pretty
io = InputOutput(pretty=False)
io.assistant_output = MagicMock()
# Setup model and coder
model = Model("gpt-3.5-turbo")
coder = Coder.create(model, None, io=io, stream=False)
# Test data
reasoning_content = "My step-by-step reasoning process"
main_content = "Final answer after reasoning"
# Mock completion response with reasoning content
class MockCompletion:
def __init__(self, content, reasoning_content):
self.content = content
# Add required attributes expected by show_send_output
self.choices = [MagicMock()]
self.choices[0].message.content = content
self.choices[0].message.reasoning_content = reasoning_content
self.finish_reason = "stop"
mock_completion = MockCompletion(main_content, reasoning_content)
# Create a mock hash object
mock_hash = MagicMock()
mock_hash.hexdigest.return_value = "mock_hash_digest"
# Mock the model's send_completion method to return the expected tuple format
with patch.object(model, "send_completion", return_value=(mock_hash, mock_completion)):
# Call send with a simple message
messages = [{"role": "user", "content": "test prompt"}]
list(coder.send(messages))
# Now verify ai_output was called with the right content
io.assistant_output.assert_called_once()
output = io.assistant_output.call_args[0][0]
dump(output)
# Output should contain formatted reasoning tags
self.assertIn("Thinking ...", output)
self.assertIn("... done thinking", output)
# Output should include both reasoning and main content
self.assertIn(reasoning_content, output)
self.assertIn(main_content, output)
# Ensure proper order: reasoning first, then main content
reasoning_pos = output.find(reasoning_content)
main_pos = output.find(main_content)
self.assertLess(
reasoning_pos, main_pos, "Reasoning content should appear before main content"
)
def test_send_with_reasoning_content_stream(self):
"""Test that streaming reasoning content is properly formatted and output."""
# Setup IO with pretty output for streaming
io = InputOutput(pretty=True)
mock_mdstream = MagicMock()
io.get_assistant_mdstream = MagicMock(return_value=mock_mdstream)
# Setup model and coder
model = Model("gpt-3.5-turbo")
coder = Coder.create(model, None, io=io, stream=True)
# Ensure the coder shows pretty output
coder.show_pretty = MagicMock(return_value=True)
# Mock streaming response chunks
class MockStreamingChunk:
def __init__(self, content=None, reasoning_content=None, finish_reason=None):
self.choices = [MagicMock()]
self.choices[0].delta = MagicMock()
self.choices[0].finish_reason = finish_reason
# Set content if provided
if content is not None:
self.choices[0].delta.content = content
else:
# Need to handle attribute access that would raise AttributeError
delattr(self.choices[0].delta, "content")
# Set reasoning_content if provided
if reasoning_content is not None:
self.choices[0].delta.reasoning_content = reasoning_content
else:
# Need to handle attribute access that would raise AttributeError
delattr(self.choices[0].delta, "reasoning_content")
# Create chunks to simulate streaming
chunks = [
# First chunk with reasoning content starts the tag
MockStreamingChunk(reasoning_content="My step-by-step "),
# Additional reasoning content
MockStreamingChunk(reasoning_content="reasoning process"),
# Switch to main content - this will automatically end the reasoning tag
MockStreamingChunk(content="Final "),
# More main content
MockStreamingChunk(content="answer "),
MockStreamingChunk(content="after reasoning"),
# End the response
MockStreamingChunk(finish_reason="stop"),
]
# Create a mock hash object
mock_hash = MagicMock()
mock_hash.hexdigest.return_value = "mock_hash_digest"
# Mock the model's send_completion to return the hash and completion
with patch.object(model, "send_completion", return_value=(mock_hash, chunks)):
# Set mdstream directly on the coder object
coder.mdstream = mock_mdstream
# Call send with a simple message
messages = [{"role": "user", "content": "test prompt"}]
list(coder.send(messages))
# Verify mdstream.update was called multiple times
mock_mdstream.update.assert_called()
coder.live_incremental_response(True)
# Explicitly get all calls to update
update_calls = mock_mdstream.update.call_args_list
# There should be at least two calls - one for streaming and one final
self.assertGreaterEqual(
len(update_calls), 2, "Should have at least two calls to update (streaming + final)"
)
# Check that at least one call has final=True (should be the last one)
has_final_true = any(call[1].get("final", False) for call in update_calls)
self.assertTrue(has_final_true, "At least one update call should have final=True")
# Get the text from the last update call
final_text = update_calls[-1][0][0]
# The final text should include both reasoning and main content with proper formatting
self.assertIn("> Thinking ...", final_text)
self.assertIn("My step-by-step reasoning process", final_text)
self.assertIn("> ... done thinking", final_text)
self.assertIn("Final answer after reasoning", final_text)
# Ensure proper order: reasoning first, then main content
reasoning_pos = final_text.find("My step-by-step reasoning process")
main_pos = final_text.find("Final answer after reasoning")
self.assertLess(
reasoning_pos, main_pos, "Reasoning content should appear before main content"
)
def test_send_with_think_tags(self):
"""Test that <think> tags are properly processed and formatted."""
# Setup IO with no pretty
io = InputOutput(pretty=False)
io.assistant_output = MagicMock()
# Setup model and coder
model = Model("gpt-3.5-turbo")
model.remove_reasoning = "think" # Set to remove <think> tags
coder = Coder.create(model, None, io=io, stream=False)
# Test data
reasoning_content = "My step-by-step reasoning process"
main_content = "Final answer after reasoning"
# Create content with think tags
combined_content = f"""<think>
{reasoning_content}
</think>
{main_content}"""
# Mock completion response with think tags in content
class MockCompletion:
def __init__(self, content):
self.content = content
# Add required attributes expected by show_send_output
self.choices = [MagicMock()]
self.choices[0].message.content = content
self.choices[0].message.reasoning_content = None # No separate reasoning_content
self.finish_reason = "stop"
mock_completion = MockCompletion(combined_content)
# Create a mock hash object
mock_hash = MagicMock()
mock_hash.hexdigest.return_value = "mock_hash_digest"
# Mock the model's send_completion method to return the expected tuple format
with patch.object(model, "send_completion", return_value=(mock_hash, mock_completion)):
# Call send with a simple message
messages = [{"role": "user", "content": "test prompt"}]
list(coder.send(messages))
# Now verify ai_output was called with the right content
io.assistant_output.assert_called_once()
output = io.assistant_output.call_args[0][0]
dump(output)
# Output should contain formatted reasoning tags
self.assertIn("Thinking ...", output)
self.assertIn("... done thinking", output)
# Output should include both reasoning and main content
self.assertIn(reasoning_content, output)
self.assertIn(main_content, output)
# Ensure proper order: reasoning first, then main content
reasoning_pos = output.find(reasoning_content)
main_pos = output.find(main_content)
self.assertLess(
reasoning_pos, main_pos, "Reasoning content should appear before main content"
)
def test_send_with_think_tags_stream(self):
"""Test that streaming with <think> tags is properly processed and formatted."""
# Setup IO with pretty output for streaming
io = InputOutput(pretty=True)
mock_mdstream = MagicMock()
io.get_assistant_mdstream = MagicMock(return_value=mock_mdstream)
# Setup model and coder
model = Model("gpt-3.5-turbo")
model.remove_reasoning = "think" # Set to remove <think> tags
coder = Coder.create(model, None, io=io, stream=True)
# Ensure the coder shows pretty output
coder.show_pretty = MagicMock(return_value=True)
# Mock streaming response chunks
class MockStreamingChunk:
def __init__(self, content=None, finish_reason=None):
self.choices = [MagicMock()]
self.choices[0].delta = MagicMock()
self.choices[0].finish_reason = finish_reason
# Set content if provided
if content is not None:
self.choices[0].delta.content = content
else:
# Need to handle attribute access that would raise AttributeError
delattr(self.choices[0].delta, "content")
# Create chunks to simulate streaming with think tags
chunks = [
# Start with open think tag
MockStreamingChunk(content="<think>\n"),
# Reasoning content inside think tags
MockStreamingChunk(content="My step-by-step "),
MockStreamingChunk(content="reasoning process\n"),
# Close think tag
MockStreamingChunk(content="</think>\n\n"),
# Main content
MockStreamingChunk(content="Final "),
MockStreamingChunk(content="answer "),
MockStreamingChunk(content="after reasoning"),
# End the response
MockStreamingChunk(finish_reason="stop"),
]
# Create a mock hash object
mock_hash = MagicMock()
mock_hash.hexdigest.return_value = "mock_hash_digest"
# Mock the model's send_completion to return the hash and completion
with patch.object(model, "send_completion", return_value=(mock_hash, chunks)):
# Set mdstream directly on the coder object
coder.mdstream = mock_mdstream
# Call send with a simple message
messages = [{"role": "user", "content": "test prompt"}]
list(coder.send(messages))
# Verify mdstream.update was called multiple times
mock_mdstream.update.assert_called()
coder.live_incremental_response(True)
# Explicitly get all calls to update
update_calls = mock_mdstream.update.call_args_list
# There should be at least two calls - one for streaming and one final
self.assertGreaterEqual(
len(update_calls), 2, "Should have at least two calls to update (streaming + final)"
)
# Check that at least one call has final=True (should be the last one)
has_final_true = any(call[1].get("final", False) for call in update_calls)
self.assertTrue(has_final_true, "At least one update call should have final=True")
# Get the text from the last update call
final_text = update_calls[-1][0][0]
# The final text should include both reasoning and main content with proper formatting
self.assertIn("> Thinking ...", final_text)
self.assertIn("My step-by-step reasoning process", final_text)
self.assertIn("> ... done thinking", final_text)
self.assertIn("Final answer after reasoning", final_text)
# Ensure proper order: reasoning first, then main content
reasoning_pos = final_text.find("My step-by-step reasoning process")
main_pos = final_text.find("Final answer after reasoning")
self.assertLess(
reasoning_pos, main_pos, "Reasoning content should appear before main content"
)
if __name__ == "__main__":
unittest.main()