auto_generate_title results in messages no longer being streamed & unexpected behavior with Qwen3 thinking

[atkr]

Hey ravitemer,

I noticed that if auto_generate_title is enabled, all messages received from the LLM after the very first one are no longer stream into the codecompanion chat, all subsequent messages appear only once the message has ended.

Is this expected behavior?

I have also noticed that when using a reasoning model such as Qwen3, the <think>[...]</think> content is included as part of the title, making the title too long and not as useful.

As I am using codecompanion exclusively with open weight local models, I went ahead and disabled this feature, but figured the feedback could be useful.

Thanks for the very useful extension!

[ravitemer]

Yes, we are currently using the model the user is chatting with to generate the title. I see how this can be an issue with thinking or more expensive models.

Will add a field like "title_adapter" which takes the adapter name and model so that can use cheaper and faster models dor this.

[ravitemer]

@atkr Added with 09d40d7

You can now specify which adapter and model to use for title generation:

			history = {
			---@type HistoryOpts
			opts = {
				auto_generate_title = true,
				title_generation_opts = {
					adapter = "copilot",
					model = "gpt-4o",
				},
			},
		}

[olimorris]

I went to raise an issue about this because I too have just discovered the same thing. Alas, I spent ages debugging the Copilot adapter trying to find the issue 😆.

@ravitemer - I do plan on adding a background strategy very soon that will allow these types of LLM requests to be handled by CodeCompanion, separately. I think that will solve this by providing a discreet communication channel for plugins to speak to LLMs.

[ravitemer]

I spent ages debugging the Copilot adapter trying to find the issue 😆.

Oh no, It was my mistake I forgot to address the streaming part of the original post. Saw that just now. It was due to the following, I didn't copied/cloned the adapter and settings properly. It works fine with the changes below. Sorry 😅

codecompanion-history.nvim/lua/codecompanion/_extensions/history/title_generator.lua

Lines 222 to 235 in ceb02d2

	function TitleGenerator:_make_adapter_request(chat, prompt, callback)
	log:trace("Making adapter request for title generation")
	local opts = self.opts.title_generation_opts or {}
	local adapter = vim.deepcopy(chat.adapter)
	local settings = vim.deepcopy(chat.settings)
	if opts.adapter then
	adapter = require("codecompanion.adapters").resolve(opts.adapter)
	end
	if opts.model then
	settings = schema.get_default(adapter, { model = opts.model })
	end
	settings = vim.deepcopy(adapter:map_schema_to_params(settings))
	settings.opts.stream = false
	local payload = {

I do plan on adding a background strategy very soon that will allow these types of LLM requests to be handled by CodeCompanion, separately.

Awesome!