custom model providers (not in Zed, just for the ACP)

[Killusions]

I was very excited when I saw this because I am building agents on top of ACP, with custom models (due to company restrictions).

As Codex CLI supports custom models, I got my setup to work with it, but I cannot integrate it into the agent since this ACP adapter uses the builtin models list and fails if the selected model is not from within.

I have tried just specifying gpt-5 and have a local proxy rewrite the model to the correct one, but then it never sends any message or message update events, even though I can clearly observe both the request and the answer going through my proxy.

And just to confirm, with the upstream standalone agent CLI my setup works (both with and without proxy, even if I set the model to gpt-5), and with an official OpenAI model the ACP also works for me, the agent STDERR is also not helpful.

My suspicion is, apart from the model id lookup issue in the beginning, the issue is that the messages are not streamed (it streams the response from the API but codex only displays it when it's fully generated, for official models it streams, but not for custom providers)

Unrelated, but it would be awesome if you could distribute this ACP binary also via npm (with optional sub-packages for every architecture), as openai also distributes their cli additionally via npm

[benbrandt]

@Killusions thanks for reporting. Does the streaming work for you in the CLI and just not in ACP?
I think we can change it so that the configured model is allowed in the dropdown so it can be selected at least.

And we're working on npm distribution as well

[Killusions]

@benbrandt The streaming doesn't work in the CLI (it does a streaming request, receives the chunks, but only prints them once the full result is available). But since it's not streaming, ACP never sends the text at all (no message event).

[benbrandt]

ah yeah I see the issue. If we show both in ACP then normal models get double messages. But it isn't great that you get neither... Will look into it.

[Killusions]

@benbrandt Any updates on this?

[benbrandt]

there is some in-progress work here #52 that I think could be split out into a few PRs, feel free to tackle any of the list there if you want it sooner, otherwise we'll get it in soon thanks

[cndoit18]

there is some in-progress work here #52 that I think could be split out into a few PRs, feel free to tackle any of the list there if you want it sooner, otherwise we'll get it in soon thanks

I can see that part of it has already been merged in. Are there any updates on the remaining parts?

[benbrandt]

I think both relevant parts have been merged now. I'll do a release soon and then you all can let me know if there is something else missing, but hopefully this solves it 🙏🏻

[Killusions]

@benbrandt Thank you for the work and improvements, excited to try them out. Is there still something blocking the release?

[benbrandt]

apologies, just kicked off the release. Got stuck trying to make sure GPT 5.1 was working on a travel day, and I didn't finish before i had to check out of my hotel 😄 thanks for your patience!

[SteffenDE]

Is it correct that there's currently no way to use custom models when not being signed into OpenAI?

codex-acp/src/codex_agent.rs

Lines 220 to 223 in c3d24ee

	// Check before sending if authentication was successful or not
	if self.auth_manager.auth().is_none() {
	return Err(Error::auth_required());
	}

I did some tests using Codex CLI + OpenRouter, but ACP always aborts with Authentication required.

[Killusions]

@SteffenDE It's possible, you just have to provide the auth method and provide your custom API key or a dummy one via CODEX_API_KEY:

await this.connection.authenticate({
    methodId: 'codex-api-key'
});

[SteffenDE]

Right, that kind of works, thank you. But it feels like it should not ask for authentication in that case at all, no? Also, the model list still shows the OpenAI models (not including the custom one), which are not useable.