hermes - 💡(How to fix) Fix Surface rate limit reset time in user-facing status messages [1 pull requests]

Problem

When the LLM provider returns a 429 rate limit, Hermes shows:

⏱️ Rate limited. Waiting 60.0s (attempt 2/3)...

But many providers include when the rate limit resets in the response body. This information is already parsed into error_context["reset_at"] by _extract_api_error_context() but is never surfaced to the user.

For Telegram/messaging users especially, knowing "resets in ~13 minutes" vs just "rate limited" is the difference between waiting patiently and wondering what's happening.

Specific issue: Codex resets_in_seconds not parsed

The Codex/ChatGPT 429 response body includes resets_in_seconds (e.g. 756), but _extract_api_error_context() only checks for resets_at, reset_at, and retry_after keys. The resets_in_seconds field is silently ignored.

Example Codex 429 body:

{
  "error": {
    "type": "usage_limit_reached",
    "message": "The usage limit has been reached",
    "plan_type": "plus",
    "resets_at": 1778839243,
    "resets_in_seconds": 756
  }
}

(resets_at IS there in this case, but in epoch format — the separate resets_in_seconds field is the human-friendly one.)

Suggested fix

1. Parse resets_in_seconds from error payload in _extract_api_error_context() (after resets_at/reset_at but before retry_after)
2. Format error_context["reset_at"] into the _emit_status message at the retry point in the main agent loop (~line 13263 in run_agent.py), e.g.:

⏱️ Rate limited. Resets in ~13m. Waiting 60.0s (attempt 2/3)...

This is a small change — the parsing infrastructure already exists, it just needs the additional key and the format string.

Environment

• Hermes v0.13.0
• Provider: openai-codex (ChatGPT Plus)
• Platform: Telegram gateway