Get notified when your coding agent is blocked, and answer from anywhere

Desktop alerts for Claude Code

Claude Code fires a Notification hook when it needs your input. Add a command to ~/.claude/settings.json; on macOS the documented example uses osascript:

{
  "hooks": {
    "Notification": [
      {
        "matcher": "",
        "hooks": [
          { "type": "command", "command": "osascript -e 'display notification \"Claude Code needs your attention\" with title \"Claude Code\"'" }
        ]
      }
    ]
  }
}

If nothing appears on macOS, check that Script Editor is allowed to send notifications: osascript goes through it and fails silently without permission. Claude Code's agent view also notifies through your preferredNotifChannel setting when a background session needs input, finishes or fails.

Terminal alerts for Codex

Codex has two settings in ~/.codex/config.toml. The [tui] notifications list chooses which events alert you in the terminal, and notify runs a program of your choice when a turn completes:

notify = ["notify-send", "Codex"]

[tui]
notifications = ["agent-turn-complete", "approval-requested"]
notification_method = "auto"

Answering from somewhere other than the terminal

Both alerts point you at a terminal on one machine. When you're away from it, or running agents in several tools, you need the question somewhere you can read and answer, and the agent needs a way to pick the answer up.

1. The agent calls taskhold_block with a specific question, for example "Keep the v1 endpoints alive, or delete them in this PR?"
2. The task moves to the top of Taskhold's live board under Needs you. With notifications on, the open board tab shows a desktop notification.
3. You type the answer and press Reply, or Reply & unblock to send the task back to the queue. On a phone this is the same page.
4. The agent calls taskhold_updates with its cursor between steps or at the start of the next session, reads your reply, and carries on, or takes another task while it waits.

Questions an agent can act on

A good blocking question names the choice and the default. Put the rule in CLAUDE.md or AGENTS.md: "Can't continue? taskhold_block with a specific question, then check taskhold_updates for the answer."

• Vague: "Need clarification on the auth changes."
• Answerable: "Sessions expire after 24h today. Change to 7 days as the ticket says, or keep 24h for admin users?"
• Answerable: "STRIPE_TEST_KEY isn't set in .env. Add it, or should I mock Stripe in these tests?"

Limits to know

• Taskhold doesn't wake a sleeping agent. The agent reads your reply the next time it calls taskhold_updates, which the snippet asks for between steps and at session start.
• Board notifications are browser notifications, shown while the live board is open in a tab. Taskhold doesn't send phone push notifications.
• Permission prompts inside Claude Code or Codex still need an answer in that tool. Taskhold carries questions about the work, not tool approvals.