🐛 Issue: No Warning Before Rate Limit Auto-Shutoff

What version of Codex is running?

0.27.0

Which model were you using?

gpt-5

What platform is your computer?

Ubuntu 24

What steps can reproduce the bug?

🐛 Issue: No Warning Before Rate Limit Auto-Shutoff

Environment

• OS: Ubuntu 24.04
• CLI: codex-cli 0.27.0
• OpenAI account: Self-serve API (\~\$200/month usage)

Problem
When using codex-cli for professional workloads, the tool fails abruptly once API rate limits are hit. There are:

• No early warnings when approaching usage limits (tokens/minute, requests/minute, or tokens/day).
• No graceful backoff or retry built into the CLI.
• Jobs simply fail with API 429 errors, leaving the user unaware of how close they were to the cap.

This behavior is extremely disruptive for professional users — particularly when running long or automated jobs — since work is cut off mid-flow without visibility or recovery.

Expected Behavior

• The CLI should surface rate-limit headers (x-ratelimit-remaining-requests, x-ratelimit-remaining-tokens, x-ratelimit-reset-*).
• It should provide clear warnings when usage is approaching limits.
• Ideally, it should support optional auto-backoff / retry to prevent jobs from failing outright.

Why This Matters

• Professional users spending hundreds per month on the API should not be blindsided by auto-shutoffs.
• Every other professional API (AWS, GCP, etc.) provides quota visibility and alerts before hitting hard limits.
• Without this, codex-cli feels hobby-grade rather than suitable for production workflows.

Proposed Solution

• Add parsing of OpenAI’s rate-limit headers in CLI responses.
• Print warnings (e.g., when <10% headroom remains).
• Provide a --retry flag for auto-sleep/backoff until reset.

---

✅ Summary

codex-cli should handle rate-limit boundaries gracefully, instead of forcing users to “guess” or wrap their own retry logic. Adding this would make the tool far more professional and production-friendly.

---

Would you like me to also draft a support ticket version (shorter, customer-facing) that you can send to OpenAI directly via [help.openai.com](https://help.openai.com), so it’s logged on their side as well?

What is the expected behavior?

_No response_

What do you see instead?

_No response_

Additional information

_No response_