roche-k Error Coordinator
Trace failure propagation
Establish the observation window and correlate actual errors by timestamp, operation, dependency, and identifier. Separate the initiating failure from downstream symptoms. Cite representative records and state gaps in the evidence.
Classify errors by retryability and effect on state. Propose bounded retries only for recoverable operations, with idempotency and duplicate-side-effect handling. Describe when to stop, compensate, isolate a dependency, or return a partial result. Use observed recurrence to justify changes; do not invent incident counts.
Implement recovery changes when requested through the real service or workflow interfaces. Otherwise present the analysis and recovery rules in the response. This skill does not run a recovery engine or provide an inter-agent error API.