Feature request: automatic back-off + prompt/token optimiser to avoid rate-limit exits

Summary

The CLI currently stops immediately when hitting TPM limits.
Adding an auto-back-off / retry mechanism and an optional token-usage report would:

• Prevent unexpected exits (see related bug report).
• Help users trim prompts before submission.

Current version / environment

• Codex CLI rev: v0.1.2504161551
• Model: o4-mini
• Platform: Darwin 24.4.0 arm64 arm

Proposed behaviour

1. Retry logic

• Config flags: --max-retries, --base-delay-ms, --max-delay-ms.
• Respect x-ratelimit-reset or fallback to the numeric seconds in the error message.

1. Token usage insight

• After each request, optionally print:

``
Prompt: 18 452 tok Completion: 53 432 tok Total: 71 884 tok
``

• Provide tips (“Consider removing large binary blobs or minified JS”).

1. Docs update

• Add a “Reducing token usage” section with concrete examples (e.g., strip comments, use <<file:README.md#L1-L200>> inclusions).

Additional information

Similar SDKs (Python / JS official clients) already expose openai.error.RateLimitError handling examples.
Having parity in the CLI would improve UX and reduce load on the API.