What can we do about the over aggressive content filter at GitHub

[SamSaffron]

I am getting this way too much on Sonnet 3.5, so much that it is becoming unusable:

  choices = { {
      content_filter_offsets = {
        check_offset = 0,
        end_offset = 152,
        start_offset = 0
      },
      content_filter_results = {
        error = {
          code = "",
          message = ""
        },
        hate = {
          filtered = false,
          severity = "0"
        },
        self_harm = {
          filtered = false,
          severity = "0"
        },
        sexual = {
          filtered = false,
          severity = "0"
        },
        violence = {
          filtered = false,
          severity = "0"
        }
      },
      delta = {
        content = "d version:\n\n```javascript",
        copilot_annotations = {
          TextCopyright = { {
              citations = vim.empty_dict(),
              details = vim.empty_dict(),
              id = 0
            } }
        },
        role = "assistant"
      },
      finish_reason = "content_filter",
      index = 0
    } },
  created = 1736400745,
  id = "msg_bdrk_0115ZxJEq23SU48ZmC6mvKB1",
  model = "claude-3.5-sonnet",
  usage = {
    completion_tokens = 17,
    prompt_tokens = 3893,
    total_tokens = 3910
  }

I just discovered this is due to the suggestions matching public code here:

⏫ it was set to block

I guess the 2 possible action items here:

1. Display a nice error message when we hit finish_reason of "content_filter"
2. Add something to the readme about it

[SamSaffron]

another interesting option might be support for BYO open ai compatible LLM?

[deathbeam]

Well this is copilot plugin so we only support copilot stuff. But i agree that we should support better error message. Issue with that is that better error reporting is dependant on bug in plenary.nvim, i made PR for it but its still waiting open: nvim-lua/plenary.nvim#633

[SamSaffron]

I am not sure there is a status issue here, status is 200, it is just once we see a chunk with:

finish_reason = "content_filter", then we know something bad happened and can report an error.