Feature Request: Hooks system for notifications (similar to Claude Code)

[tomerweller]

Summary

Add a hooks system that allows users to run custom shell commands when certain events occur, such as when the CLI is waiting for user input.

Use Case

When running Copilot CLI on remote development machines (via SSH), users often step away while the agent works on tasks. Having a notification system (e.g., Pushover, ntfy, Slack webhook) would alert users when their attention is needed, improving workflow efficiency.

Proposed Solution

A hooks configuration similar to Claude Code's implementation:

// ~/.copilot/hooks.json
{
  "hooks": {
    "prompt": [
      {
        "type": "command",
        "command": "curl -s -X POST https://api.pushover.net/1/messages.json -d 'token=XXX&user=YYY&message=Copilot needs input'"
      }
    ]
  }
}

Suggested hook events:

• prompt - When waiting for user input
• tool_start / tool_end - When tools execute
• session_start / session_end - Session lifecycle

Alternatives Considered

• Terminal beep (beep: true) - Works but not useful for remote/background sessions
• Wrapper scripts monitoring output - Hacky and unreliable

Additional Context

Claude Code has this feature and it significantly improves the remote development experience. Reference: https://docs.anthropic.com/en/docs/claude-code/hooks

[RyanHecht]

Going to let you in on a little secret: we support hooks now! We're finishing up getting docs out this week, and then you'll see more wide communication about it.

You define them in .github/hooks/*.json.

Example schema:

{
  "version": 1,
  "hooks": {
    "preToolUse": [{
      "type": "command",
      "bash": "./scripts/validate.sh"
    }]
  }
}

Here are the events we support

Event	When it fires	Can deny?
sessionStart	Session begins or resumes	No
sessionEnd	Session completes	No
userPromptSubmitted	User submits a prompt	No
preToolUse	Before any tool runs	Yes
postToolUse	After a tool completes	No
errorOccurred	When an error occurs	No

Your script receives JSON on stdin and can return {"permissionDecision": "deny", "permissionDecisionReason": "..."} to deny execution on applicable events.

Scott Hanselman made this nice app to send a Windows toast when the agent needs his input: https://github.com/shanselman/toasty