Crash after large refactoring

[efunneko]

Issue

I did asked for a reasonably large refactoring of some markdown files in which it removed two files, created a sub-directory and then created 4 files in that sub-directory. After the LLM response was complete, I got this stack dump and aider quit:

File "/home/github/chat/.venv/bin/aider", line 8, in <module>
    sys.exit(main())
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/main.py", line 661, in main
    coder.run()
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/coders/base_coder.py", line 725, in run
    self.run_one(user_message, preproc)
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/coders/base_coder.py", line 768, in run_one
    list(self.send_message(message))
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/coders/base_coder.py", line 1129, in send_message
    self.show_exhausted_error()
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/coders/base_coder.py", line 1206, in show_exhausted_error
    input_tokens = self.main_model.token_count(self.format_messages())
  File "/home/github/chat/.venv/lib/python3.10/site-packages/aider/models.py", line 590, in token_count
    msgs = json.dumps(messages)
  File "/usr/lib/python3.10/json/__init__.py", line 231, in dumps
    return _default_encoder.encode(obj)
  File "/usr/lib/python3.10/json/encoder.py", line 199, in encode
    chunks = self.iterencode(o, _one_shot=True)
  File "/usr/lib/python3.10/json/encoder.py", line 257, in iterencode
    return _iterencode(o, 0)
  File "/usr/lib/python3.10/json/encoder.py", line 179, in default
    raise TypeError(f'Object of type {o.__class__.__name__} '
TypeError: Object of type ChatChunks is not JSON serializable

I restarted and tried the same refactoring again and I got the same crash.

Version and model info

Aider v0.54.4
Model: Sonnet 3.5 accessed through an OpenAI API on a LiteLLM proxy

[efunneko]

I debugged this a bit more and I think it is hitting Sonnet's response token limit and then it crashes when trying to count tokens in the show_exhausted_error() function

I will try to reduce my request to stay within the response limit. I had thought I had read in aider's blog that for Sonnet it would automatically send 'continue' if this happened, but perhaps not...

[paul-gauthier]

Thanks for reporting this! Should be fixed now in the main branch, with test coverage.

To get the latest build:

aider --install-main-branch

[paul-gauthier]

I'm going to close this issue for now, but feel free to add a comment here and I will re-open. Or feel free to file a new issue any time.