Bug Description

OpenClaw gateway classifies Gemini 500 (InternalServerError) and 503 (ServiceUnavailable) responses as rate_limit errors, which triggers the exponential cooldown mechanism (1min → 5min → 25min → 60min cap). This effectively takes the agent offline even when API usage is well below rate limits.

Evidence

• Gemini API dashboard shows usage at 5/25 RPM and 20/250 RPD (Paid Tier 1) — nowhere near limits
• The actual errors in Google's dashboard are 500 InternalServerError, NOT 429 TooManyRequests
• Both auth profiles (google-gemini-cli and anthropic fallback) entered cooldown simultaneously, leaving the agent with no working model
• auth-profiles.json showed cooldownUntil set with failureCounts incrementing under the rate_limit category

Expected Behavior

• 429 errors → trigger rate limit cooldown (correct)
• 500/503 errors → retry with backoff but do NOT enter rate_limit cooldown state
• Transient server errors should not disable the agent for extended periods

Actual Behavior

• 500/503 errors → classified as rate_limit → exponential cooldown activated
• Agent goes offline for up to 60 minutes due to server-side errors it has no control over
• Both primary and fallback models can be simultaneously disabled

Impact

• Agent becomes completely unresponsive during Gemini outages
• Even brief Gemini instability (a few 500s) can trigger multi-minute cooldowns
• Fallback model (Claude Sonnet) may also be in cooldown, leaving zero working models

Reproduction

1. Configure gateway with google-gemini-cli as primary model
2. Wait for Gemini to return a 500 or 503 error (happens periodically)
3. Observe auth-profiles.json — failureCounts.rate_limit increments and cooldownUntil is set
4. Agent stops making requests even though rate limits are not exceeded

Workaround

Manually clear cooldowns in auth-profiles.json:

python3 -c "
import json
with open('auth-profiles.json') as f:
    data = json.load(f)
for p in data.get('profiles', []):
    for key in ['cooldownUntil', 'errorCount', 'failureCounts', 'lastFailureAt']:
        if key in p:
            del p[key]
with open('auth-profiles.json', 'w') as f:
    json.dump(data, f, indent=2)
"

Environment

• OpenClaw Gateway v2026.2.19
• Google Gemini Paid Tier 1
• Model: gemini-3-pro-preview