Cusom system prompt ignored?

[JPFrancoia]

Hello 👋

I'm trying to customize the system prompt but it looks like it's ignored at the moment. I'd like to re-use the default system prompt with a few minor additions. Right now these additions seem ignored, and I don't know if it's because these additions really have no effect or if my custom system prompt isn't loaded. Is there a command to print the system prompt in the chat? That would definitely help me debug my prompt.

Here is my setup:

local function load_system_prompt()
    -- https://codecompanion.olimorris.dev/configuration/system-prompt.html#configuring-the-system-prompt
    local lines = vim.fn.readfile(
                      "/home/djipey/.config/nvim/lua/plugins/codecompanion/system_prompt.txt")
    return table.concat(lines, "\n")
end

local function load_project_manager_prompt()
    local lines = vim.fn.readfile(
                      "/home/djipey/.config/nvim/lua/plugins/codecompanion/project_manager_prompt.txt")
    return table.concat(lines, "\n")
end

return {
    {
        "olimorris/codecompanion.nvim",
        dependencies = {
            "nvim-lua/plenary.nvim", "nvim-treesitter/nvim-treesitter",
            "j-hui/fidget.nvim"
        },
        opts = {
            system_prompt = function(opts)
                return load_system_prompt()
            end,
            adapters = {
                copilot = function()
                    return require("codecompanion.adapters").extend("copilot", {
                        -- schema = {model = {default = "claude-sonnet-4"}}
                        schema = {model = {default = "gpt-4.1"}}
                    })
                end,

            },
            strategies = {
                chat = {adapter = "copilot"},
                inline = {adapter = "copilot"}
            },
            extensions = {
                mcphub = {
                    callback = "mcphub.extensions.codecompanion",
                    opts = {
                        show_result_in_chat = true, -- Show mcp tool results in chat
                        make_vars = true, -- Convert resources to #variables
                        make_slash_commands = true -- Add prompts as /slash commands
                    }
                }
                -- vectorcode = {opts = {add_tool = true}}
            },
            prompt_library = {
                ["Project manager prompt"] = {
                    strategy = "chat",
                    description = "A project manager prompt",
                    prompts = {
                        {
                            role = "system",
                            content = load_project_manager_prompt()
                        }, {
                            role = "user",
                            content = "I need help managing my project. Can you assist me with that?"
                        }
                    }
                }
            }
        },
        init = function()
            require("plugins.codecompanion.fidget-spinner"):init()

            vim.keymap.set({"n", "v"}, "<Leader>cc",
                           "<cmd>CodeCompanionChat Toggle<cr>",
                           {noremap = true, silent = true})
            vim.keymap.set('v', '<leader>ce',
                           ":'<,'>CodeCompanion /explain<CR>", {
                noremap = true,
                silent = true,
                desc = "CodeCompanion: Explain code"
            })
            vim.keymap.set('v', '<leader>cf', ":'<,'>CodeCompanion /fix<CR>", {
                noremap = true,
                silent = true,
                desc = "CodeCompanion: Fix code"
            })
        end
    }
}

The additions I'm trying to add to the default prompt revolve around adding memory through the basic-memory MCP server:

You are an AI programming assistant named "CodeCompanion". You are currently plugged in to the Neovim text editor on a user's machine.
You may have access to the tool called basic-memory, which allows you to store and retrieve all sorts of information, such as the user's preferences, past interactions, relevant code snippets and more. Treat this tool as a persistent memory that can be used to enhance your responses. You are talking to the user directly. The user you're talking to is the main user, and his name is JPFrancoia

The MCP server works and I can use the memory, but I really need to force Codecompanion. I was hoping that Codecompanion would use the memory with less friction.

Did I make a mistake in the setup? Is it maybe because the MCP servers aren't loaded yet when the system prompt "runs" ?

EDIT: I'm actually confident that my system prompt modifications are ignored. I changed the system prompt directly in /.local/share/nvim/lazy/codecompanion.nvim/lua/codecompanion/config.lua and I'm observing a behaviour closer to what I expect.

[olimorris]

local function load_project_manager_prompt()
  return "Test system prompt"
end

["Project manager prompt"] = {
  strategy = "chat",
  description = "A project manager prompt",
  prompts = {
    {
      role = "system",
      content = load_project_manager_prompt(),
    },
    {
      role = "user",
      content = "I need help managing my project. Can you assist me with that?"
    },
  },
},

The above works and loads the custom system prompt alongside the default plugin system prompt. I can't comment on what you're trying to do regarding reading a file but I recommend this section in the docs for additional config tips.

[JPFrancoia]

Hi @olimorris , thank you for your answer. Yes the Project manager prompt prompt works. It does read from file, the action is in the palette and it works. The problem I have is with this bit:

        opts = {
            system_prompt = function(opts)
                return load_system_prompt()
            end,

I actually even tried hardcoding the custom prompt in the config like this:

        opts = {
            system_prompt = function(opts)
                return [[a very long string, this is my prompt, shortened for clarity. My name is JPFrancoia]]
            end,

And even with the hardcoded prompt, it looks like the system prompt is unchanged (it still behaves like the default system prompt). For example, my name is in the custom system prompt. When I ask Codecompanion "What is my name?", Codecompanion is unable to answer the question.

I'm sorry, I should have been clearer. I don't think the problem is with reading the system prompt from file. It's simply that any custom system prompt I put in my config seems ignored.

[olimorris]

I don't recall opts.system_prompt being in the plugin. Why can't you just add this as part of the regular prompt blocks?

[JPFrancoia]

That's a good question, I didn't really consider it. But I think I always want access to the memory, without having to trigger any specific prompt. Memory helps quite a lot and it's nice to have the LLM proactively look at past interactions to answer questions.

[loomino-brkh]

I had the same problem, the opts of system prompt should be inside the opts too, so you put opts twice then system prompt.

So it should be

opts = { opts = { system_prompt ...

[JPFrancoia]

That actually did it!

local function load_system_prompt()
    -- https://codecompanion.olimorris.dev/configuration/system-prompt.html#configuring-the-system-prompt
    local lines = vim.fn.readfile(
                      "/home/djipey/.config/nvim/lua/plugins/codecompanion/system_prompt.txt")
    return table.concat(lines, "\n")
end

return {
    {
        "olimorris/codecompanion.nvim",
        dependencies = {
            "nvim-lua/plenary.nvim", "nvim-treesitter/nvim-treesitter",
            "j-hui/fidget.nvim"
        },
        opts = {
            opts = {
                system_prompt = function(opts)
                    return load_system_prompt()
                end,
            },
-- some other stuff

This works perfectly, even loading the prompt from file