Debugger won’t stop and reports completion incorrectly

[harryworkbiz]

I’ve run into this a few times now. The debugger starts running and just keeps going. This time, I pressed the cancel button repeatedly and nothing happened, it just kept debugging. In the screenshot, it had already been running for almost eight minutes. It’s let me cancel in the past, but this one was completely stuck with no way to stop it.

---

I’m also seeing the “Debugging Complete” card show up even when the issue isn’t actually fixed. This happens most often when the preview fails to load. It reports everything as resolved, but the preview is still broken. Screenshot for reference:

---

I've started seeing this message (screenshot below) after every initial chat message, right after it finishes the first phase, even when nothing looks wrong. Is this just a default message? It kind of throws you off, because it makes it seem like there are bugs even if we can’t see any.

[AshishKumar4]

I've started seeing this message (screenshot below) after every initial chat message, right after it finishes the first phase, even when nothing looks wrong. Is this just a default message? It kind of throws you off, because it makes it seem like there are bugs even if we can’t see any.

So this comes up when there are linting or type errors detected by the static analyzer. Most often, these dont lead to immediate failure but can lead to crashes if a certain code flow is hit, hence the agent flags it.

I’m also seeing the “Debugging Complete” card show up even when the issue isn’t actually fixed. This happens most often when the preview fails to load. It reports everything as resolved, but the preview is still broken. Screenshot for reference:

The agent thinks its fixed when it's clearly not. In this case, is the preview not working at all even after multiple refreshes? Can you share the appId?

I’ve run into this a few times now. The debugger starts running and just keeps going. This time, I pressed the cancel button repeatedly and nothing happened, it just kept debugging. In the screenshot, it had already been running for almost eight minutes. It’s let me cancel in the past, but this one was completely stuck with no way to stop it.

When it does that, does it keep generating the exact same text over and over again i.e, enters an infinite loop, or does it just keep going around in circles?

[harryworkbiz]

So this comes up when there are linting or type errors detected by the static analyzer. Most often, these dont lead to immediate failure but can lead to crashes if a certain code flow is hit, hence the agent flags it.

Totally get that, but it’d be way more helpful if it showed the actual errors instead of a generic message. That way we can tell whether something is harmless or actually critical. Even better, whenever any errors show up, small or major, there could be a “Fix issue” button so it can jump in right away instead of us having to copy errors back manually. Example below.

The agent thinks its fixed when it's clearly not. In this case, is the preview not working at all even after multiple refreshes? Can you share the appId?

The preview didm't work at all, even after several refreshes. I’ve hit the "preview not available" issue many times. Usually I can fix it after a few tries with the deep debugger, but it’s a consistent problem. I deleted the app already, but if I run into it again I’ll send the appid here.

When it does that, does it keep generating the exact same text over and over again i.e, enters an infinite loop, or does it just keep going around in circles?

It didn’t generate the same text over and over. It looked like it was doing something, but it just kept running, and I couldn’t abort or cancel the deep debugger. But that was the first time I wasn’t able to stop it.

[AshishKumar4]

Btw I saw in the logs that for your app, it seems the system picked up a 'presentation' type template which it wasn't supposed to. This is a new feature that is still in development. This lead to your app unable to be sandboxed properly.
Are such things happening for you for other apps too though?

[harryworkbiz]

You’re right about the presentation template, I specifically asked for that. But the “preview not available” issue happens when building normal landing pages and web apps. If it comes up again, I won’t ask the debugger to fix it, I’ll just send the appID here instead.

[AshishKumar4]

Yes! That would be right. Also, there was an incident yesterday that did cause deployment issues. That has been resolved though.
Also, we have been recently experimenting with different default model configs to move out of gemini-only system.
Do give feedback on the performance of various aspects - UI Quality, depth, error frequency, deep debugging experience, conversation etc. We are in the process of upgrading the model config to more powerful and cheaper/efficient models

[harryworkbiz]

Absolutely, I noticed the new AI models. Definitely going to test them out and see how they perform on various tasks.

Quick question on cost optimization: If I used something like Sonnet 4.5 for the main coding phases (Phase Implementation, Deep Debugger) and Gemini 3 Pro for planning/screenshot analysis, with lighter models for the other agents - are the costs still manageable? Does the vibesdk architecture help optimize token usage even with more premium models?

Trying to figure out the right balance between quality and sustainability for my own production deployment.

Regardless, I'll get back to you with feedback on the new models!

[AshishKumar4]

We haven't started much work towards further token usage optimization, but the architecture of the phasic agent in vibesdk allows a lot of flexibility in this already - different model for different tasks. I experimented a lot with non-claude models though - claude is just too expensive for what it's worth. I found gpt 5 mini (medium) to be comparable to gemini 3 pro for blueprints, and for now have stuck to gemini 2.5 pro (low, 0.2 temp) for phase impl task, but switched to grok-4 fast for phase gen (gpt 5 mini wasnt great here), and conversation is also handled by grok 4 fast.
gpt 5 mini now powers the deep debugger.
These seem to have reduced the costs for us by ~30% already, and hopefully switching phase impl model with kimi k2 or gpt oss should save us more.
On the official build.cloudflare.dev, i won't recommend running any claude models because it would exhaust the free tier quota pretty quickly.

Trying to figure out the right balance between quality and sustainability for my own production deployment.

we are doing the same! Will keep you posted on this

[AshishKumar4]

Another thing - the Phasic system works great for app development but it's phasic nature and extensive context engineering makes it suboptimal for caching by providers (implicit caching atleast). Currently working on a true 'agentic' coding agent in feat/general-agents-ui, that should be much better for caching purposes although It would require a lot of effort getting it to anywhere close to phasic agent's quality and speed

[harryworkbiz]

Thank you for the insight! Makes total sense for build.cloudflare.dev to optimize for cost given it's free.

Looking forward to the updates.