ca9b510922
- Reset master to upstream/main (16,697 commits) - Overlay 2,271 local-only files (skills, tools, workspace, configs, apps) - Restore IDENTITY.md and USER.md templates - Build verified, gateway running, Discord working Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1.5 KiB
1.5 KiB
OpenClaw Model Usage Policy & Fallback Recommendations
1. Purpose
Ensure reliable model operations and cost efficiency through structured fallback mechanisms, cooldown logic, and token usage thresholds.
2. Fallback Policy Improvements
- Trigger Conditions: Failover when:
- Error rate exceeds 5% over 5 minutes
- Response latency >5s for 3 consecutive requests
- Token quota exhausted (429 errors)
- Cooldown Logic:
- 5-minute cooldown before retrying failed models
- Progressive backoff (5m → 15m → 30m) after repeated failures
- Priority Tiers:
- Tier 1: Primary model (e.g., qwen3-235b)
- Tier 2: Secondary model (e.g., gpt-4o)
- Tier 3: Budget model (e.g., gemma2-27b)
3. Token Thresholds
| Metric | Warning Threshold | Hard Limit |
|---|---|---|
| Tokens/Minute | 8,000 | 10,000 |
| Tokens/Hour | 400,000 | 500,000 |
| Daily Cost | $80 | $100 |
| Concurrent Requests | 25 | 50 |
4. Monitoring
- Real-time dashboard with:
- Token usage heatmaps
- Error rate tracking
- Fallback chain visualization
- Automated alerts via Slack/Email for:
- Threshold breaches
- Model outages
- Cost anomalies
5. Implementation Roadmap
Q3 2026:
- Metrics collection pipeline
- Dynamic model routing
- Alerting framework
Q4 2026:
- Automated cooldown management
- Cost optimization module