OpenClaw Resilience Monitor

Problem

When agents call LLM APIs, errors such as 429, 503, timeouts, and network failures can interrupt tasks, while raw logs remain scattered across requests and instances. Troubleshooting often requires reading logs, checking per-model behavior, and confirming retry records, making it hard to tell quickly whether a model is degrading, whether an error type is retryable, or whether a failed task was recovered.

How It Works

This skill adds natural-language controls for OpenClaw to inspect API health and manage retry behavior:
- resilience_stats: query today, hour, week, or per-model error statistics.
- resilience_strategies: list, add, update, or reset retry strategies with fixed, exponential, or custom intervals.
- resilience_report: generate daily, per-model, recovery, or full status reports.
- resilience_dashboard: open the local real-time panel for current errors, model breakdowns, recent failures, and strategy cards.
For example, an operator can ask for a model's daily error rate and switch the default timeout strategy to exponential backoff. The skill classifies errors such as rate_limit, timeout, auth_failed, and context_too_long, and marks which categories are retryable. Tool registration, automatic error recording, the retry engine, and the Dashboard service are provided by the companion plugin.

Boundaries

It is intended for engineering workflows already using OpenClaw where API health should be checked with plain Chinese instructions. Without the companion plugin installed and loaded, tool calls will fail and the dashboard will not start. Strategy edits usually apply only to the local Gateway instance.