Regarding showing #Me and #CodeCompanion (user / LLM text hard to see)

[alexghergh]

Sometimes it's a bit hard to see the separation between the user text and the LLM text. Even with show_header_separator and a custom separator, the LLM (in this case GPT-5.2) still outputs things like H2 headers and --- markdown separators on its own, making it confusing (outputting H2 headers also trips up {{ bindings, which are supposed to (?) juggle between the user and LLM turns).

I didn't see any way around this (e.g. implementing custom foreground / background for the separator or title) in the UI, e.g. this piece of code:

---Get the formatted header for the chat buffer
---@param role string The role of the user
---@return string
function UI:format_header(role)
  local header = "## " .. role
  if config.display.chat.show_header_separator then
    header = string.format("%s %s", header, config.display.chat.separator)
  end

  return header
end

Is there something else that people do for better separation of chats between the user/LLM? Or is there any possibility to update this function to use something like a custom background / foreground for rendering the header + separator?

Thank you!

[olimorris]

An accompanying screenshot would have been helpful to see what you're seeing.

I'd recommend using a markdown rendering plugin. They generally improve the look of all markdown files:

Even with show_header_separator and a custom separator, the LLM (in this case GPT-5.2) still outputs things like H2 headers and --- markdown separators on its own

Yep this is a frustration of mine too. A current limitation of LLMs unfortunately. I observe this with Sonnet and Opus 4.5 too.

[alexghergh]

Thanks for the answer!

Well, here's a screenshot (blurred, but alas). The 4), 5) and 6) are part of the LLM output. I'm using render-markdown.nvim.

I think it's worth pointing out that [[ (not {{ as I initially mentioned by mistake) is broken (so it cycles between H2 headers, regardless of whether it's the LLM output or the headers for LLM/user turns).

Is going to H1 headers not an option? It seems that the LLM never outputs those.

Also, would it not be desirable to augment the header separators to have a different foreground/backgroud color? I understand if that would be low on the priority list. I've already somewhat fixed it by using / separators, which are at least somewhat more visible.

[olimorris]

I think it's worth pointing out that [[ (not {{ as I initially mentioned by mistake) is broken
Is going to H1 headers not an option? It seems that the LLM never outputs those.

I wrote it specifically to cycle between H2 headers. Two years ago when I added it, LLMs seemed to be rather good at listening to the basic system prompt of "don't use H1 or H2 headers". I think it's a valid ask to update the TS parsing to detect the user and LLM headers though. I'll put that on my todo list.

Someone else raised H1 headers a couple of months back. My argument against that is H1 is normally reserved for a title rather than sub-headings.

Also, would it not be desirable to augment the header separators to have a different foreground/backgroud color? I

Why can't you do that with CodeCompanionChatSeparator ?