openclaw - 💡(How to fix) Fix Model fallback not triggered on 429 rate_limit_error — retries same provider indefinitely [1 participants]

Bug Description

When the primary model (anthropic/claude-opus-4-6) returns HTTP 429 rate_limit_error, OpenClaw retries the same provider indefinitely instead of falling back to the configured fallback model (google/gemini-2.0-flash).

Config

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-opus-4-6",
        "fallbacks": ["google/gemini-2.0-flash"]
      }
    }
  }
}

Auth setup: Two Anthropic token profiles (anthropic:default + anthropic:manual) with auth order configured. Both are Claude Max accounts (not API keys).

Observed Behavior

• Session receives a user message
• OpenClaw attempts anthropic/claude-opus-4-6 → gets 429 rate_limit_error
• Instead of falling back to google/gemini-2.0-flash, it retries the same provider 20+ times over ~30 minutes
• All retries fail with identical 429 errors
• The session becomes completely unresponsive — user never receives a reply
• The fallback model is never attempted

Expected Behavior

After exhausting retries on the primary provider (including cycling through auth profiles), OpenClaw should fall back to the next model in the fallbacks array.

Evidence from Session Logs

23:24:37 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:24:45 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:24:55 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:29:04 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:29:08 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:29:16 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:42:25 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:42:30 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:42:36 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:42:47 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:43:03 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:43:08 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:43:15 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:43:25 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:59:52 → 429 rate_limit_error (anthropic/claude-opus-4-6)
23:59:57 → 429 rate_limit_error (anthropic/claude-opus-4-6)
00:00:03 → 429 rate_limit_error (anthropic/claude-opus-4-6)
00:00:14 → 429 rate_limit_error (anthropic/claude-opus-4-6)
... (continues, google/gemini-2.0-flash never attempted)

All errors have the Anthropic error message: "This request would exceed your account's rate limit. Please try again later."

Note: The Claude Max account had available usage (only 10% of weekly limit used) — the 429 was a burst rate limit, not a quota exhaustion.

Environment

• OpenClaw version: 2026.3.24
• OS: macOS (Darwin arm64)
• Auth: Claude Max token auth (not API key)
• Two auth profiles configured with auth order

Suggested Fix

1. After N consecutive 429s from the same provider (e.g., 3-5 attempts), trigger model fallback
2. Alternatively, respect Retry-After headers and if the wait exceeds a threshold, fall back immediately
3. Consider cycling through auth profiles first, then falling back to the next model in the fallback chain