Multiple perfectly innocent codex sessions started complaining about prompts violating usage policy even halfway into a session

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #39742
• #39687
• #38849
• #39539

Powered by Codex Action

Same issue in feedback: 01a02a43-20ec-7d52-a7e9-887d87b68fdb

When I ask codex what I did:

Nothing in your request was inappropriate. You’re debugging ordinary NPC navigation and scheduling in your own game; that’s fully legitimate.
If Codex showed a “usage policy” warning, it was almost certainly a false positive from an automated classifier—not something you should avoid phrasing next time. I don’t have visibility into the exact classifier trigger, but likely candidates in the logs are generic words such as “failed,” “blocked,” or “satisfy need” being interpreted without the surrounding game-simulation context. You did nothing wrong, and I’m continuing the debugging work.

Immediately after this reply from codex it triggered another potential violation flag.

I can reproduce the same issue in Codex Desktop on Windows.

Environment:

• Codex app-server/client: 0.149.0
• Initially reproduced with gpt-5.6-sol, High reasoning
• Changing model/reasoning settings did not resolve it

This happened during benign Unreal Engine C++ development for a colony-simulation game. The task involved ordinary AI scheduling and work-priority logic.

After successful builds and tests, Codex repeatedly failed before starting a local tool call. At least five consecutive attempts to apply a small Markdown documentation update ended with:

“Invalid prompt: your prompt was flagged as potentially violating our usage policy.”

Read-only commands and some patches had worked in the same session, so this was not a general filesystem or Git failure. The rejected attempts did not launch the announced local command.

The local Codex logs show a tool-call output item followed by:
unhandled responses event: "error"

No useful policy category or request ID was exposed.

One potentially relevant — and slightly absurd — lexical detail: the project contains a benign C++ interface/file named INeedSatisfier. In the game architecture, this simply means “an implementation that creates a job for satisfying a simulated character need,” such as eating or resting.

We considered whether an automated classifier might tokenize INeedSatisfier as the natural-language phrase “I need satisfier.” However, this remains only a hypothesis: failures also occurred when that identifier was not present in the immediate command or explanation.

I can provide the affected task ID privately to OpenAI staff, but I would prefer not to publish the task identifier, complete conversation, or proprietary project files in a public GitHub issue.