Ask FluxAI Why Your App Restarted

Fluxers! “Why did my app restart?” is one of the most common questions anyone running software asks — and one of the most frustrating, because the honest answer usually means digging through logs, memory graphs and container states to piece it together.

You can now just ask. FluxAI, the assistant in the new FluxCloud preview, can diagnose a restart: it tells you what happened, why, and what it would cost to fix — priced before you change anything.

What an answer looks like

Ask about an app that keeps restarting, and instead of a guess you get something like this:

“It hit its 8 GB memory limit at 14:02 UTC and was stopped. Raising it to 12 GB adds about $0.11 per day.”

That is the whole design in one sentence: the cause, when it happened, the fix, and its price as the difference from what you pay today — so you can decide in seconds whether it is worth it.

Facts first, not a guess

The obvious way to build this — hand a model the log tail and ask what went wrong — does not work well. Logs alone mislead. A process killed for using too much memory often just says “Killed”, and a model reading only that line will happily guess “probably out of disk”.

So FluxAI does not read the log and guess. A dedicated tool, flux_diagnose_app, gathers the raw facts from FluxOS for every component of your app:

• how it last exited, and whether it was killed for running out of memory
• how many times it has restarted
• its peak memory against its limit
• how much of its disk it is using
• the tail of its log

The cause is decided in code

Here is the important part. The cause is not left to the model’s judgement. It is decided in code, from those facts, into one of five outcomes:

• Out of memory — it hit its RAM limit.
• Disk full — it ran out of storage.
• Crashed — the program failed on its own.
• Exited on its own — it finished or shut itself down.
• Moved to another node — the network relocated it, which is not a fault at all.

Where there is a fix, the diagnosis carries it too. For a game server that ran out of memory, for example, the recommended size comes from the marketplace’s own sizing for your player count, not from a number the model invents.

The model’s job is then the part language models are genuinely good at: explaining the diagnosis clearly, in your language, and answering your follow-up questions.

Checked before you see it

Every answer is checked before it reaches you. The cause the assistant states must match the diagnosis, and must not claim a different one. A price change must be stated as what it adds, not as a new total dressed up as the increase. And every per-day figure must trace back to a real monthly price from the network.

If the model names a fix but forgets to price it, the system prices it anyway. You always see what a change will cost.

Tested on every cause

We built five evaluation cases, one for each cause. With the facts gathered and the cause decided in code, FluxAI gets all five right. Across the wider set of thirty-five deploy and management cases it scores 33 of 35, with no invented domains or credentials.

It understands which app you mean

You do not have to remember your app’s exact registered name. Ask about “my minecraft server” and FluxAI matches it to the app on your account you actually meant, then diagnoses that one.

It can explain, never change

As with everything the assistant does in FluxCloud, it cannot sign or spend. It can diagnose, explain and price a fix, and prefill the change for you — but resizing your app only ever happens when you approve it with your own wallet.

Getting better every week

Answers in the assistant now carry a thumbs up and down, so you can tell us when one helped and when it missed. Conversations are kept masked for up to a year to improve FluxAI, as set out in the FluxCloud terms — the same model, trained on Flux and running on Flux, getting sharper from the questions people actually ask.

Try it

Open the FluxCloud preview, sign in, and ask the assistant why your app restarted. If you want the background on the model itself, start with Meet FluxAI.