Gemini adapter won't honor the system prompt

[meijieru]

The language configuration is not being respected by the models when using the Gemini adapter. This issue is specific to the adapter, as the configuration functions as expected when using Gemini 2.5 Pro through Copilot. An inspection of the debug window confirms that both the language and system prompt options are correctly applied, suggesting the issue occurs downstream.

[yingmanwumen]

The same happens to me.

1. when no tools are added, it acts normal
2. when any tool is added, gemini lost all system prompts configured before
3. this doesn't happen to copilot

[yingmanwumen]

This could be fixed by this patch:

          gemini_2_5_flash = function()
            return require("codecompanion.adapters.http").extend("gemini", {
              schema = {
                model = {
                  default = "gemini-2.5-flash",
                },
                temperature = {
                  default = 0.5,
                },
              },
              handlers = { -- This is the patch
                form_messages = function(self, messages)
                  local gemini = require("codecompanion.adapters.http.gemini")
                  local raw_messages = gemini.handlers.form_messages(self, messages)

                  -- Collect and merge system messages
                  local system_content = {}
                  local other_messages = {}

                  for _, msg in ipairs(raw_messages.messages) do
                    if msg.role == "system" then
                      table.insert(system_content, msg.content)
                    else
                      table.insert(other_messages, msg)
                    end
                  end

                  -- If we found system messages, merge them and prepend
                  if #system_content > 0 then
                    local merged_system_message = {
                      role = "system",
                      content = table.concat(system_content, "\n\n"),
                    }
                    table.insert(other_messages, 1, merged_system_message)
                  end

                  raw_messages.messages = other_messages
                  return raw_messages
                end,
              },
            })
          end,

This happens due to gemini doesn't support multiple system prompts: it reads the latest system prompt only:
https://discuss.ai.google.dev/t/multiple-system-prompt-messages-not-supported-by-gemini-openai-compatibility/86097

@olimorris shall we apply this patch to https://github.com/olimorris/codecompanion.nvim/blob/main/lua/codecompanion/adapters/http/gemini.lua#L59 ?

[00sapo]

Hey thanks. Gemini was not able to use system prompts until I used this patch. I will open an issue...

[00sapo]

mmm I cannot really reproduce my problem because for simple cases it works, while for complex cases, even your patch is not working. Moreover, for some reason when using the patch the model tries to use insert_edit_into_file even if I haven't provided that tool, and fails using it.

[olimorris]

All,

Can someone actually prove this?

This link points to an unconfirmed post by a user on the Google forums.

I'd be surprised if Google adopt OpenAI for their endpoints and don't account for this.

Also, we have a utility function to merge system prompts together.

[olimorris]

Fixed in #2530