Coding agents

Coding agents (OpenHands, Claude Code) work for minutes or hours, often several at once. cue-hook lets them tell you when they finish or need you. Cue decides how: it merges pings from parallel agents, keeps them out of your quiet hours, and sends them over your own channels.

Install

pip install cue-notify-client   # provides the cue-hook command
export CUE_URL=https://cue.example.com CUE_API_KEY=ck_… CUE_RECIPIENT=dev-42

CUE_RECIPIENT is your external id in Cue. Use a key bound to an agent if you want a per-person budget on these pings.

OpenHands

.openhands/hooks.json:

{
  "stop": [{"matcher": "*", "hooks": [{"command": "cue-hook", "async": true}]}],
  "session_end": [{"matcher": "*", "hooks": [{"command": "cue-hook", "async": true}]}]
}

Claude Code

.claude/settings.json. The Notification event fires when Claude is waiting for your input:

{
  "hooks": {
    "Notification": [{"hooks": [{"type": "command", "command": "cue-hook"}]}],
    "Stop": [{"hooks": [{"type": "command", "command": "cue-hook"}]}]
  }
}

Events and a rule

Each hook becomes an event named after it: agent.stop, agent.notification, agent.session_end. Their data holds hook, session_id, project, agent, and, when the agent sends them, message and stop_reason.

POST /v1/rules
{"key": "agents-waiting", "name": "Agent needs me", "event": "agent.notification",
 "template": "agent-waiting", "category": "agents", "channels": ["push", "telegram"],
 "importance": "high", "digest": {"window": "PT2M", "send_after": 5}}

The digest turns five agents stopping within two minutes into one notification. Create the agents category first (POST /v1/categories), with whatever caps suit you.

The hook never blocks the agent. It always exits 0 and writes failures to stderr.