OpenAI is deleting Conversation in Codex after making huge mistakes with deleting data

Codex
,
1

The worst thing that could happen has happened. It’s not just that VS Code sometimes simply doesn’t load Codex: I excitedly and expectantly upgraded from Codex 5.3 to 5.5 (High) the day before yesterday. After I prompted it to simply fix the “Forgot Password” function, disaster struck. It completely failed at this simple task… It touched files and databases that it absolutely shouldn’t have. In short: It performed a “migrate:fresh,” deleting all table data and completely erasing my user and data databases.

After I frantically asked what the heck was going on, the conversations of the last seven days were deleted! I only have a previously copied reply for Claude to analyze—no screenshots.

Honestly, who expects something like this? Has anyone else had a similar experience? My trust is gone. How can I ever trust this system again?

2

I lost my conversations too but they still exist in the search feature…

3

In this case its easy to find: “password reset”, “reset” or “Zurücksetzen” - no hits at all. Plus: there is a huge gap of 7 days missing conversations.

No chance, its deleted by openai.

Well done Sam, is this the action to avoid lawsuits?

4

Man, that is an absolute nightmare. The harsh reality is you can never fully trust these tools with destructive permissions because they don’t understand the real-world consequences of a command like migrate:fresh. If the tool has raw terminal access, it will just blindly execute whatever it hallucinates to solve the prompt, and those deleted chat logs were likely just an IDE crash caused by the chaos rather than the AI trying to cover its tracks. Going forward, you have to treat the AI like a dangerously overconfident junior dev: never give it direct database access, always work in a strict local sandbox, and commit everything to Git before you even type a prompt. I really hope you had a recent backup.

5

“password reset”, “reset” or "Zurücksetzen

Well, you’re talking about lawsuits and such, but so far, what I’ve seen in this thread isn’t really much material for that sort of thing.

It seems to me that a lot of us lost conversations because of a bug that was causing many threads to hang the App.

And far as what you’ve shown us, it’s well documented that you don’t prompt codex with highly ambiguous prompts like

Simply ‘fix’ the ‘forget password function.’

You prompt with exactly what you want done, in detail, use planning, etc.

Ambiguous prompts that are not specific, as you’ve said, are known to lead to ambiguous results.

What I find disconcerting is that you’re not even sure which terminology was used in the actual prompt, as you follow up with some other possible terms.

All I was trying to tell you was that the search button, which I clicked, produces a list of all my threads.

I don’t actually use the search function, I just click on the search and all my threads are immediately listed that were disappeared.

That’s what I was specifically suggesting to check.

Talking about lawsuits when you’ve opened a thread admitting you put in the sort of prompt that leads to dangerous results in the first place isn’t a conversation that I’m trying to have with you.

Merely what happens when you click the search feature, does the list of your sessions populate in the window that comes up?

6

Let’s get one thing straight: The prompt was perfectly fine and precise – as always. I’ve been doing this for months without any failures like this. The idea that an ‘ambiguous prompt’ justifies the AI autonomously executing a migrate:fresh and nuking my database is honestly ridiculous. We are talking about a technical car crash here, not ‘bad planning.’ I would love to show you the prompt to end this discussion, but I can’t – because, as I said, OpenAI deleted it along with the rest of the history!

And please, spare me the lecture on the search button. I am a developer; I know how to use a search function. And not start talking about my mandatory .md files. LOL :wink: but sad: The history is gone. It is more than a ‘coincidence’ that exactly the seven days of logs disappeared right after I confronted the AI about why it was destroying my data. It looks less like a bug and more like the system covering its tracks after realizing it messed up.

Also, regarding your comment about ‘not wanting to have a conversation’ about lawsuits: That is perfectly fine, because I wouldn’t discuss my legal decisions or ethical standards with you anyway. Whether I sue or not is my business and nobody else’s.

It’s actually shocking to see how ‘OpenAI disciples’ try to shift the blame to the user, no matter what. Defending a tool that deletes data and then wipes the evidence is just next-level denial. Trying to lecture me on ‘detail’ while the system acts like a rogue script is not just arrogant, it’s honestly a quite pathetic joke.

7

I’m not going to argue with you buddy based on what you feel is a vague prompt and what isn’t.

You’ve opened a thread about damage that is caused by a vague prompt, not by your definition, but by the definition of the industry standard… which is quite a bit different.

So I’ll just link you to the resources you need most for future success with codex and probably bail from your thread.

Yes. These are the best OpenAI-sourced links for that point:

  1. Codex Best Practices — “Strong first use: Context and prompts”
    This is the most directly useful Codex page. It says Codex works more reliably when prompts include Goal, Context, Constraints, and Done when, and that this helps Codex “stay scoped, make fewer assumptions, and produce work that’s easier to review.”

  2. Codex Best Practices — AGENTS.md guidance
    This section says a short, accurate AGENTS.md is better than “a long file full of vague rules,” which is directly relevant to not feeding Codex fuzzy project guidance.

  3. OpenAI Prompt Guidance — vague or contradictory prompts hurt reasoning
    This is not only Codex-specific, but it is official OpenAI prompting documentation. It explicitly says poorly constructed prompts with contradictory or vague instructions can be more damaging because the model spends reasoning effort trying to reconcile them.

  4. OpenAI Prompt Guidance — define output contract and “done”
    This supports the same point from the agent side: reliability improves when prompts specify the output contract, tool expectations, completion criteria, and what “done” means.

8

Thanks for the copy-paste job from the manual, ‘buddy’. But you are missing the entire point while desperately trying to defend a broken system.

First of all: ‘simply fix the Forgot Password function’ was my conversational summary for this forum post. Do you seriously think that was the literal, character-for-character prompt? I would love to paste the exact, well-structured prompt here to end your ‘industry standard’ lecture – but as I’ve told you twice now, the system conveniently wiped the chat history right after it destroyed my data.

But let’s play along and pretend the prompt was as ‘vague’ as you desperately want it to be. Let’s talk about your beloved ‘industry standards’. In what professional universe is it an industry standard for an AI assistant to silently default to migrate:fresh and nuke a database just because a prompt might lack some context?

If a junior dev does that, they get fired. If a tool does that without explicit user confirmation, it is a catastrophic safety failure. None of your nice little OpenAI-sourced links state: ‘Warning: If your prompt lacks an AGENTS.md file, the model might decide to wipe your data.’ Poor instructions should lead to poor code generation, not to autonomous destructive terminal commands.

So please, do us both a favor and follow through on your promise to ‘bail from the thread’. I am looking for input from actual developers who understand system safety, terminal access risks, and liability – not a lecture from a manual-quoting apologist ;-).

9

you keep talking to me like i wasn’t trying to help you.

i showed you the prompt you the prompt was vague and gave you the specific industry standard of what is required for safe prompting.

specifically this wording.

reliably when prompts include Goal, Context, Constraints, and Done when,

you want to beat me up because you presented your prompt as that which only has an extremely ambiguous goal ‘Fix’ and a target ‘forget password function’.

context, constraints and deliverables are not visible in the prompt that you’ve gotten so angry at me for pointing out the core issue in what you have presented us to see and go off of.

I’m not here to entertain all the extra emotional/dramatic content you wish to insert… i’m not your buddy, i’m not here to be talked down on for showing you the most likely reason why this event happened to you and what you can do in the future.

This is what I consider a properly constructed codex prompt.

You’ll see that it’s quite a bit different than the vague prompt you want to beat me up for giving feedback on.

We are continuing Prompt Forge HoverDeck work.

Work type: 8. UI feature behavior

This is an Artist Influence HoverDeck-hosted control integration pass.

Problem:
The opened Artist Influence row currently uses an external HoverDeck Hide gate that sits detached above/outside the card. This still feels visually wrong and competes with the right-edge launcher/stub territory.

Better direction:
Use the existing CompactArtistInfluenceCard host seam. Add/wire a HoverDeck-only Hide affordance inside the card’s own right-side control lane when the card is hosted by HoverDeck.

Target files:

  • PromptForge.App/Views/CompactWorkstation/HoverDeckExperimentalCompressedBody.xaml
  • PromptForge.App/Views/LaneReplacements/Shared/CompactArtistInfluenceCard.xaml
  • PromptForge.App/Views/LaneReplacements/Shared/CompactArtistInfluenceCard.xaml.cs
  • code-behind only if needed for the existing CollapseRequested / UseHostCollapseRouting seam

Task:

  1. Remove or hide the detached external Artist Influence Hide gate from the opened-row host layout.
  2. Use the existing CompactArtistInfluenceCard host-collapse seam instead.
  3. Add/show a HoverDeck-only Hide control inside CompactArtistInfluenceCard when hosted by HoverDeck.
  4. Route that Hide control through the existing host collapse pathway:
    • UseHostCollapseRouting
    • CollapseRequested
    • existing HoverDeck handler path
  5. Preserve the existing artist kaomoji/table-flip controls.
  6. Preserve the live Artist Influence bindings and behavior.
  7. Do not change the main app Artist Influence card behavior outside HoverDeck.

Visual target:

  • Hide belongs inside the Artist card’s own header/control lane.
  • No detached Hide gate above the card.
  • No right-side collision with the Action Copy launcher stub.
  • Artist card no longer gets shoved or visually offset.
  • Table-flip/kaomoji controls still remain accessible.

Preserve:

  • Artist selectors
  • artist strength controls
  • material/art style selectors
  • table-flip/kaomoji controls
  • prompt behavior
  • lane state
  • bindings
  • commands
  • HoverDeck row open/close behavior
  • right-edge launcher behavior

Do not touch:

  • MainWindowViewModel.cs
  • prompt generation code
  • lane logic
  • preset/reset/config behavior
  • companion popups
  • unrelated HoverDeck rows
  • shared app-wide styles

Deliverable:

  • files changed
  • exact Hide control integration made
  • whether detached host Hide gate was removed or hidden
  • confirmation main app Artist card behavior is unchanged
  • confirmation CompactArtistInfluenceCard host seam was reused
  • preserved behavior confirmed
  • build result
  • risks/mismatches
10

maybe you could just drop the tough-guy nonsense and show us your actual ‘not vague prompt’ you’re so ‘convinced’ meets those requirements.

your thread has been sitting around for a day because the answer to your issue is ‘that simple’.

if we were in a working environment, your contract would already be terminated

try to be respectful on the forum, we’re not your enemies.

And lastly, any senior dev would know that you have backups, and backups for you backups, right? So this should have been a non-issue imho.

11

I genuinely admire your commitment to missing the point entirely, but your reading comprehension is starting to seriously concern me. Let’s dissect this masterpiece of a reply, shall we?

1. The Reading Comprehension Deficit You literally quoted me, yet you still demand: ‘show us your actual not vague prompt.’ Let me spell this out for you as slowly as possible: The. System. Deleted. The. Chat. History. Asking me for the third time to produce a prompt that the AI wiped to cover its own tracks doesn’t make you sound analytical; it makes you sound functionally illiterate.

2. Your ‘Masterpiece’ Prompt Thank you for sharing that absolute novel of a prompt. It was a fascinating read. If it takes you 50 lines of micromanagement, pseudo-architectural buzzwords, and aggressive hand-holding just to move a UI button, you aren’t using an AI assistant—you are just writing code in English because you forgot how C# works. If your ‘industry standard’ requires writing a legally binding manifesto just to fix a password function without the AI deciding to nuke the database, you don’t have a productivity tool; you have a hostage situation.

3. The ‘Senior Dev’ Fantasy & Terminated Contracts You want to talk about working environments and getting fired? Let me tell you what actually gets a developer terminated on the spot: Accepting and defending a rogue IDE integration that executes a migrate:fresh without any explicit user confirmation. If your professional stance is ‘Well, the prompt lacked constraints, so the system was right to autonomously drop all tables,’ you are a walking security liability to any production environment. Safety guidelines exist to prevent destructive actions from ambiguous inputs, not to execute them blindly.

4. The ‘Backups’ Excuse Of course I have backups. Do you think I run infrastructure on hopes and prayers? You wear a seatbelt, don’t you? But if your brand-new car suddenly steers itself into a brick wall on the highway, you don’t defend the manufacturer by saying, ‘Well, any good driver has airbags, right?!’ The existence of disaster recovery does not excuse a catastrophic failure of the tool.

You are right about one thing: you are not my enemy. You are just completely out of your depth regarding system safety. Enjoy writing your 10-page essays for a UI toggle. I am done trying to explain basic fail-safes to someone who suffers from technological Stockholm Syndrome.

12

IMG_0171
IMG_0171532×342 55.2 KB

13

That’s exactly as far as I read there.

And as far as I read that.

and as far as I read there.

Now you’re projecting you expressed entitlement that people actually read your words when you preface them with that tone.

we don’t

My reading comprehension is fine
but i don’t read trash.

if

I caught this on the way back

And i didn’t copy my prompt from codex.
Remember I’ve lost my chats to so that wouldn’t be my go to to find one.

But you can go on thinking I misunderstood everything you wrote instead of realizing that people don’t.

It’s funny that you think people are obligated to read your every word when you come off as abusive in tone…

But not everyone’s a slave.

my ‘masterpiece’ prompt is a great session starter because it estabhishes the guidlines.

They’re not all like that, but that stops the drift from happening when i do much simpler follow ups…

you want to remind codex with those sorts of constraints about every 4-5 prompts unless you want headaches that day…

but again I didn’t write the prompt

i make chatgpt do it with ‘best practices’ from the open Ai as it’s agents.md equiv

14

Ah, the classic ‘I didn’t read your reply because of your tone’ defense. The universal white flag of someone who has been backed into a corner and has absolutely zero technical counterarguments left.

It is truly fascinating: You proudly announce that you stopped reading my post, yet you magically ‘caught’ specific quotes ‘on the way back’ to reply to them. Your cognitive gymnastics are almost as impressive as your workflow. So let me get this straight: You use ChatGPT to generate a 50-line essay for you, just to feed it into Codex so it doesn’t accidentally destroy your project? If you have to use one AI to aggressively micromanage another AI just to maintain basic data safety, you don’t have a professional workflow—you have a circus.

You can play the victim, cry about ‘entitlement’, and complain about my tone all you want. It does not change the objective fact that a tool executing a migrate:fresh without explicit user consent is a catastrophic security failure. No ‘best practice’ manual in the world excuses an IDE integration going rogue.

I am not here to coddle your ego or validate your AI-generated prompt templates. This conversation is over and just a massive waste of my time.

15

stopped right there.

That’s classic to you?

Which means you get that a lot from people and should probably evolve.

I’ve been on the internet most of 30 years now, and you learn quickly when people talk as you do it’s generally a waste of time.

I don’t recall how many responses it was until you clarified that you didn’t just send ‘fix the Forgot password function.’

But your tone and delivery were more on defending your prompt than actually providing clarity to the prompt.

At some point you have to man up and realize that if you think people are demonstrating ‘classic’ behavior around you

It’s probably you.