$ cat blog/your-turn

Designing Your-Turn Alerts

The interruption is useful only when it is fresh, attributable, and easy to act on.

A background coding run does not need attention while it is producing output. The handoff happens when the run completes and cannot continue without a reply. Agent Island is designed around that moment.

Agent Island alert showing that a Codex thread is waiting for the user
Demonstration capture. The alert identifies the provider, thread, and project before offering a return action.

Freshness before interruption

A completion marker in an old transcript is not a reason to interrupt someone. The app only promotes a session to your turn inside a limited freshness window and after observing relevant activity.

Context before action

An alert should answer three questions without exposing transcript content: which provider finished, which thread is waiting, and which project it belongs to. That is enough context to decide whether to return now.

A return path, not another dashboard

The primary action returns the user toward the relevant coding surface. The alert also provides a clear acknowledgement action. It does not attempt to replace the terminal, session browser, or coding client.

What belongs in a useful report

When reporting an alert problem, include the provider, approximate time, whether the session was still producing output, and whether the alert was missing, late, or attached to the wrong thread. Do not post transcript contents, credentials, or provider tokens in a public issue.

Concrete alert reports can be filed in the public issue tracker after sensitive details are removed.

← All posts