$ define ai-coding-status-companion

What Is an AI Coding Status Companion?

An AI coding status companion is a small, persistent surface that tells you what Claude Code or Codex is doing and when your attention is required. It observes the workflow; it does not silently take control of the agent.

Verified 2026-07-27. The useful category boundary is simple: a status companion reduces monitoring work. It should make session state visible outside the terminal, preserve the identity of the session that changed, and notify you only when the next action belongs to you.

This is different from an agent runner, retry loop, or automation framework. Those systems decide what to execute. A status companion reports observable state so the developer can decide what happens next.

The problem is attention, not execution

Long-running coding-agent tasks create an awkward choice. Keep the terminal visible and repeatedly check it, or switch to other work and risk leaving a question unanswered. The coding agent may still be healthy; the missing layer is a reliable handoff signal.

A companion belongs in a menu bar, system tray, notch, or another stable desktop surface because it remains glanceable while the terminal is covered. Its value comes from shortening the time between an actual state change and the developer noticing it.

A companion needs evidence-backed states

Labels such as working, waiting, finished, and needs you are only useful when each one has an observable rule. A recent tool event can support working. A real assistant turn that asks for input can support your-turn. Silence by itself is not enough to prove completion because a process may be calculating, waiting on a tool, or stalled.

Multi-session workflows add another requirement: the notification must retain provider and session identity. "Codex needs you" is incomplete if three Codex sessions are open and the alert cannot point back to the correct one.

A your-turn alert is a handoff, not a success claim

The alert should mean that the current evidence says the developer can act. It does not mean the requested code is correct, tests passed, or the task is complete. Those are separate verification questions.

A quiet alert policy also matters. If the exact session is already frontmost, another sound may add noise without improving response time. If the developer switches away, the pending handoff must remain visible rather than disappearing with focus.

Local observation needs a clear data boundary

A desktop companion may read local session records to reconstruct state. A responsible design names those sources, limits what is parsed, and keeps the result on the device unless the user explicitly shares it. "Local-first" should describe an inspectable data flow, not a vague security promise.

Agent Island uses no product account and does not upload session data to Agent Island. Sharing a report or image is a separate, user-initiated action.

How to evaluate a status companion

  1. Run two sessions and confirm the UI preserves their identities.
  2. Trigger a real user-input request and measure when the alert appears.
  3. Leave a tool running and verify that silence is not called finished.
  4. Bring the target session frontmost and inspect the notification policy.
  5. Disconnect the network and confirm local state observation still works.
  6. Read the product boundary for telemetry, sharing, and automation.

Current Agent Island scope

Agent Island v1.7.1 is a free, open-source status companion for Claude Code and Codex on macOS and Windows. It provides visible session state and your-turn alerts. This page does not claim that Agent Island automatically resumes, approves, or operates either coding agent.

Continue with the your-turn state definition, the session-state taxonomy, or the Claude Code notification guide.

← All posts