bug: deepseek + ai mode generates broken mjml, thinking is not conveyed in the chat UI

[linuxd3v]

I use deepseek v4 pro openai api.
I use following query: "Update general theme for this email - make it more earthy looking".
Same query works great when I use it from opencode/openchamber.

Issue 1:
sometimes broken mjml is generated I think - but there is no indication in the notifuse UI about that.
So then I click on "save" template in notifuse - but then cannot preview anything because mjml is broken.
Also - because mjml sometimes is broken - preview does not update, so people will be confused what is even happening if query finished, but template kind of looks the same.
Typically that would require a simple follow up prompt to fix, but because UI does not tell me that mjml is broken, or what exactly is broken - non-technical notifuse users just would have not idea what is even broken.
Experience for non-developers will be that - "it would seem that notifuse broke my template".
Proposed solution: add mjml validation when user query completes - automatically submit a follow up query "Fix these issues" if mjml validation does not pass.

Issue 2:
Thinking models typically output a lot more information while generating response - is it possible to enable streaming of all that information? Perhaps could it be made configurable when configuring openai api?
Otherwise people are left to wonder if things are even working.
See for example openchamber deepseek streaming output:

Issue 3:
models typically have an effort switch. Could effort switch be made available in the openai configuration screen? it could just be default by default, but if needed would be great to select medium or high.

[linuxd3v]

When I attempt to resolve mjml issues via notifuse chat window - it reports all issues as fixed, but preview still does not update and mjml still shows those same validation issues.
I'm not sure if there is some notifuse mjml code state issues? - where up to date context is not passed to the openai api? Or maybe chat updates dont update the mjml correctly?

Perhaps I have to click on "save" in order to preview and try to fix in another chat iteration?

But the problem is - non-technical notifuse users are not going to be clicking on "save" before they can ascertain template is looking good. And they can't do that without working preview functionality first.

[linuxd3v]

Ultimately I think there should be an SSE or websockets used here to broadcast the state.

[pierre-b]

Thinking streaming has been added in https://github.com/Notifuse/notifuse/releases/tag/v35.0 with other improvements

[linuxd3v]

It seems to be better.
But I would question couple things:

1. it asks user to fix mjml. Non technical users wont know what to do with this.
   Do I have to ask chat to fix broken mjml? why? this seems like should be mechanical on backend - backend validates mjml and should automatically dispatch fix request to ai.
   non-technical people dont know what mjml is - they will think its notifuse broke.

2. collaborative users don't see updated template.
   see old template on the left, new on the right. (different browsers)

3. There is no template locking - how come 2 users can trigger redesign queries at the same time?
   Which template would win - user 1 or user 2?

4. when both users trigger ai design build - user B design succeeds, user A design just fails entirely and user A email view never updates to the user B design.

**SSE or websockets state sharing likely needs a lot more thinking through. **
For starters SSE subscription would start on page load (not on ai request) - for both (all) browsers, - so both browser get streaming results.
Also it likely should not be allowed to start ai design in one browser, when another browser is running ai design at the same time. Locking is needed.

• many other details like shared streaming vs individual events. Some events should be public - apply to all users looking, some private - just for that specific user.

[pierre-b]

Hi @linuxd3v ,

• if the AI model is not smart enough it will NEVER be able to generate proper MJML. the issue seem to come from the model you are using
• websockets/live-collaboration won't be on the roadmap, it's way too complex to implement

[linuxd3v]

I don't think that's a realistic expectation of any LLM. None of the current models - whether it's Claude, GPT, or others- are going to generate valid MJML 100% of the time. They all make mistakes occasionally.

From an end-user perspective, I think the application should be designed to account for that rather than assuming perfect model output. Things like validation, automatic fixes, or clear error reporting make the experience much more robust regardless of which model is used.

Of course, it's your call - I just wanted to share some feedback from the perspective of someone using the product.

[linuxd3v]

if the AI model is not smart enough it will NEVER be able to generate proper MJML. the issue seem to come from the model you are using

deepseek fixed the problem. it took 1 additional request. when I tested in vs code it sometimes one-shots it, sometimes it takes a follow up. gpt-5.5 is same way in my tests

[linuxd3v]

websockets/live-collaboration won't be on the roadmap, it's way too complex to implement

perhaps simple locking the template when someone opens it for edit would fill the gap meanwhile?
sort of like wordpress does?