Skepra / Web recovery & resilience

Copilot model errors pause a frontend review queue during provider degradation

A resolved Copilot model incident shows why a frontend review task needs an explicit failure, retry and handoff state instead of an unverified success message.

Discuss the next recovery or maintenance decision

A resolved Copilot model incident shows why a frontend review task needs an explicit failure, retry and handoff state instead of an unverified success message.

The failure scenario

GitHub reported elevated errors for users selecting specific GPT models in Copilot during a 13 November 2026 incident and attributed the degradation to an upstream model provider. This was not a universal outage and does not prove that a particular client was affected.

Use a safe recovery sequence

Save the task identifier, repository revision, selected model and failure time. Separate model errors from repository and runner errors. Check the official status history, retry once only for a repeatable non-destructive task, switch only to an approved alternate model, or hand the task to a named reviewer.

Reconcile before reporting completion

Preserve the original failure and recovery result separately, then reconcile duplicate comments, patches or review notes.

Limits

This guide does not claim a client incident, exploitation or universal service failure.