Hexagram Circuit Breaker v3

Problem It Addresses

In intelligent quotation workflows, a single request may pass through rule calculation, LLM judgment, external data sources, and downstream services. The risk is not only hard failure, but failure that looks plausible: timeouts, repeated errors, or elevated error rates that feed into LLM reasoning can produce unstable quotations, duplicate calls, or weak fallbacks. hexagram-circuit-v3 addresses this boundary by making the system explicitly degrade, open, and audit itself when the rule layer or model layer becomes unreliable.

How It Works

The skill organizes circuit breaking into three stages: rule detection → LLM judgment → manual takeover. Instead of simply returning an error, it changes system behavior when thresholds are crossed:
- Error rate > 30%: enter a half-open state and limit further traffic.
- More than 3 consecutive failures: enter a fully open state and block subsequent calls while preserving recovery probes.
- Timeout > 5 seconds: trigger automatic degradation to a more conservative handling path.

It also provides complete audit logging and automatic recovery detection, helping operators answer when the circuit opened, what evidence triggered it, and when retries become safe. The documentation states an “add-only, no modification” principle, making it suitable as a protective layer over existing quotation logic rather than a replacement for business rules.

Boundaries And Caveats

This skill is most useful when the upstream chain has observable steps, clear failure signals, and structured validation around LLM output. If upstream fields are missing, error codes are inconsistent, or model responses are unstructured, circuit breaking can contain failure spread but cannot fix the underlying defect. Defaults such as 30%, 3 failures, and 5 seconds should be recalibrated against interface SLAs, traffic peaks, and manual takeover capacity.