Codex Rate Limits Discussion Thread

Follow-up with numbers from my local Codex logs:

I queried ~/.codex/logs_2.sqlite for post sampling token usage entries and grouped them by codex-auto-review vs everything else.

Result:

  • codex-auto-review: 523 turns, 41.12M tokens, 74.6% of logged usage

  • interactive/other turns: 197 turns, 14.00M tokens, 25.4% of logged usage

So in my case, the majority of local Codex token usage is not coming from my visible interactive agent work, but from codex-auto-review.

This may explain why my Pro weekly limit is draining so quickly even though I am using one agent at a time and not running parallel agents. From the user perspective, I’m doing normal development work. But internally, Codex appears to be spending most of the tokens on automatic review/safety/background checks. This is not transparent!

Can OpenAI clarify:

  1. Does codex-auto-review count against the same Pro weekly Codex limit?

  2. Can users disable or limit auto-review/ambient suggestion runs? Or do we have to use yolo mode.

  3. Can the usage dashboard show per-task/per-source usage, especially interactive agent vs auto-review? Or do we have to check the

  4. Is it expected that auto-review can account for ~75% of local token usage? Or is this maybe a bug, as it just started the last days?