What version of Codex CLI is running?
codex-cli 0.157.1 (Homebrew cask)
What subscription do you have?
ChatGPT sign-in
Which model were you using?
gpt-6-luna, reasoning effort low
What platform is your computer?
macOS 27.0 (26A428), arm64 (Mac Studio)
What issue are you seeing?
Twice within ~8 hours, a long-running agentic turn on reasoning effort low ended with:
Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: https://platform.openai.com/docs/guides/reasoning#advice-on-prompting
(error code invalid_prompt)
Both turns were ordinary software-development work in a private Python repository (working through failing unit tests / running a test-suite plan). Neither flag came on a new user message: Codex had started the turn itself to continue the thread's active goal, and the flag came mid-turn after 70 and 260 ordinary tool calls respectively. The last tool call before each flag was a read-only sed -n of a Python source file. So the flagged "prompt" appears to be the accumulated conversation context.
- Incident 1: 2026-09-27T23:08:28Z, turn
01a0e512-b13c-7363-8e88-21cc4226e528
- Incident 2: 2026-09-28T06:39:53Z, turn
01a0e6a5-1ef4-7c72-b430-3053757e67fb
The same goal later continued without flags (once on low, and on medium after I switched), so I can't say low is proven to be the cause, but both flags happened on low. It's costly: to avoid losing long turns I've moved to medium.
What steps can reproduce the bug?
Uploaded thread: 01a0e3ce-15c3-7770-ad4d-3afb25f17eaa
Long autonomous goal-continuation turns on low in the same thread. Not deterministically reproducible.
What is the expected behavior?
Routine coding/test work isn't flagged as a usage-policy violation, or at least the flag is attributable (request ID shown) and appealable from the CLI.
Additional information
A local rollout doesn't record the request ID of the flagged request, which makes reporting harder. Also reported via OpenAI Help Center chat.
What version of Codex CLI is running?
codex-cli 0.157.1 (Homebrew cask)
What subscription do you have?
ChatGPT sign-in
Which model were you using?
gpt-6-luna, reasoning effort
lowWhat platform is your computer?
macOS 27.0 (26A428), arm64 (Mac Studio)
What issue are you seeing?
Twice within ~8 hours, a long-running agentic turn on reasoning effort low ended with:
(error code
invalid_prompt)Both turns were ordinary software-development work in a private Python repository (working through failing unit tests / running a test-suite plan). Neither flag came on a new user message: Codex had started the turn itself to continue the thread's active goal, and the flag came mid-turn after 70 and 260 ordinary tool calls respectively. The last tool call before each flag was a read-only
sed -nof a Python source file. So the flagged "prompt" appears to be the accumulated conversation context.01a0e512-b13c-7363-8e88-21cc4226e52801a0e6a5-1ef4-7c72-b430-3053757e67fbThe same goal later continued without flags (once on low, and on medium after I switched), so I can't say low is proven to be the cause, but both flags happened on low. It's costly: to avoid losing long turns I've moved to medium.
What steps can reproduce the bug?
Uploaded thread: 01a0e3ce-15c3-7770-ad4d-3afb25f17eaa
Long autonomous goal-continuation turns on
lowin the same thread. Not deterministically reproducible.What is the expected behavior?
Routine coding/test work isn't flagged as a usage-policy violation, or at least the flag is attributable (request ID shown) and appealable from the CLI.
Additional information
A local rollout doesn't record the request ID of the flagged request, which makes reporting harder. Also reported via OpenAI Help Center chat.