Title generation tends to be very generic if copilot-instructions.md is part of the conversation

[bsolar17]

I use a .github/copilot-instructions.md file which CodeCompanion can include automatically in the context through the memory feature. With the feature enabled the .github/copilot-instructions.md file becomes part of the conversation so that the AI can be instructed on generic guidelines to be used in the project.

The codecompanion-history plugin attempts to generate a relevant title for the conversation, but it tends to interpret the copilot-instructions.md files as relevant to determine the specific conversation as opposed to generic instructions that are not really meaningful for the conversation's title.

As example, prompting the text "Test prompt":

• Without memory it generates the title "Test Prompt Discussion".
• With copilot-instructions in memory it generate the title "GitHub Copilot Instructions for Neovim".

The first title is much more relevant than the second one.

I wonder if there is a way to generate the title more accurately by somehow excluding generic instructions from the context taken into account by the title generator.

[grddavies]

Yeah I had a poke around the code and it looks like rules messages aren't filtered from the chat when sending to llm to generate the title: see here

-- title-generator.lua L101
    local relevant_messages = vim.tbl_filter(function(msg)
        -- Include user and assistant messages with actual content
        local has_content = msg.content and vim.trim(msg.content) ~= ""
        local is_relevant_role = msg.role == config.constants.USER_ROLE or msg.role == config.constants.LLM_ROLE
        -- local not_tagged = not (msg.opts and (msg.opts.tag or msg.opts.reference or msg.opts.context_id))
        -- FIX: we are looking in the wrong place for a tag when creating a title
        local not_tagged = not (msg._meta and msg._meta.tag)
        return has_content and is_relevant_role and not_tagged
    end, chat.messages)

@ravitemer with a bit of guidance I could try get a PR together

[chunleng]

I am wondering if we could have an option to use the title generated/set from codecompanion for this? Maybe it's much more accurate in that sense because code companion actually have it's own feature to generate title that does not get affected by additional context provided.

Also, for me personally, I rename the chat sometimes to more easily navigate through my conversation. When I open up the history, the name changes and it becomes hard to differentiate 😅